Research and Development in the Computer and Information Sciences

DOCUMENT RESUME

ED 039 891 LI 001 944

AUTHOR Stevens, Mary Elizabeth TITLE Research and Development in the Computer and Information Sciences. Volume 1 r Information Acquisition, Sensing, and Input: A Selective Literature Review. INSTITUTION National Bureau of Standards (DOC), Washington, D.C. Center for Computer Sciences and Technology. REPORT NO NBS-Monogr-113-Vol-1 PUB DATE Mar 70 NOTE 169p. AVAILABLE FROM Superintendent of Documents, U.S. Government Printing Office, Washington, D.C. 20402 (C 13.44:113/Vol.1, $1.50)

EDRS PRICE EDRS Price MF-$0.75 HC Not Available from EDRS. DESCRIPTORS Automation, *Computer Science, Information Networks, *Information Processing, *Information Science, *Literature Reviews, Research, *Telecommunication

ABSTRACT The series, of which this is the initial report, is intended to give a selective overview of research and development efforts and requirements in the computer and informationsciences. The operations of information acquisition, sensing, andinput to information processing systems ate considered in generalized terms. Specific topics include but are not limited to:(1) source data automation and remote sensing techniques,(2) communication systems and data transmission links, (3) audio and graphic inputs, (4) preprocessing operations upon input items,(5) character recognition, (6) speech recognition and (7) various other aspects of automatic pattern recognition. Supplemental notes and abibliography of over 640 cited references are included. (Author/NH) UNITED STATES DEPARTMENT OF COMMERCE Maurice H. Stans, Secretary NATIONAL BUREAU OF STANDARDS Lewis M. Branscomb, Director

CO tr1 cD Research and Development in the um' Computer and Information Sciences

1. Information Acquisition, Sensing, and Input: A Selective Literature Review

Mary Elizabeth Stevens Center for Computer Sciences and Technology National Bureau of Standards Washington, D.C. 20234

"PERMISSION TO REPRODUCE THIS COPY- RIGHTED MATERIAL BY MICROFICHE ONLY HAS BEEN GRANTED BY

U.S. DEPARTMENT OF HEALTH.EDUCATION & WELFARE TO ERIC AND ORGANIZATIO OP TING OFFICE OF EDUCATION UNDER AGREEMENTS WITH E U.S. OFFICE THIS DOCUMENT HAS BEEN REPRODUCED OF EDUCATION. FURTHER REPRODUCTION EXACTLY AS RECEIVED FROM THE PERSONOR OUTSIDE THE ERIC SYSTEM REQUIRES PER- ORGAN:ZATION ORIGINATING IT. POINTSOF MISSION OF THE COPYRIGHT OWNER" VIEW OR OPINIONS STATED DO NOT NECES- SARILY REPRESENT OFFICIAL OFFICE OFEDU- CATION POSITION OR POLICY.

National Bureau of Standards Monograph 113, Vol. 1 Nat. Bur. Stand. (U.S.), Monogr. 113-1, 170 pages (Mar. 1970) CODEN: NBSMA Issued March 1970

For sale by the Syperintendent of Documents, U.S. Government Printing Office Washington, D.C. 20402 (Order by SD Catalog No. C 13.44:113/Vol. 1), Price $1.50 o .4!luAki .94

Ira Foreword The Center for Computer Sciences and Technology of the National Bureau of Standards has responsibility under the authority of Public Law 89-306 (the Brooks Bill) for automatic data processing standards development, for consultation and technical assistance to Federal agencies, and for supporting research in matters relating to the use of computers in the Federal Government. This selective literature review is one of a series intended to improve interchange of information among those engaged in research and development in the fields of the computer andinformation sciences. Considered in this volume are the specific areas of information acquisition, sensing, and input, including the problems of character and pattern recognition. Names and descriptions of specific proprietary devices and equipment have been included for the convenience of the reader, but completeness in this respect is recognized to be impossible. Certain important developments have remained proprietary or have not been reported in the open literature; thus major contributors to key developments in the field may have been omitted. The omission of any method or device does not necessarily imply that it is considered unsuitable or unsatisfactory, nor does inclusion of descriptive material on commercially available instruments, products, programs, or processes constitute endorsement.

LEWIS M. BRANSCOMB, Director

HI Contents Page Foreword III 1. Introduction 1 1.1. General introductory considerations 1 1.2. Basic background considerations 2 1.3. A generalized information processing system 3 1.3.1. Automatic production control 3 1.3.2. Character recognition systems 3 1.3.3. An experimental ISSR system in the U.S.S.R 5 1.3.4. Implications of the generalized scheme 7 1.4. Specific background considerations 8 1.4.1. Plan of attack 8 1.4.2. Some general caveats 9 1.5. Scope and coverage of this report 10 2. Information acquisition, source data automation, and remote- data collection 10' 2.1. Basic operations and requirements 10 2.2. Source data automation 12 2.3. Measurements automation and feedback control 13 2.4. Remote sensing and data collection techniques.. 14 3. Communication systems and data transmission links...... 16 3.1. Communication links, services, and systems 16 3.1.1. Voice and digital data transmission 16 3.1.2. Graphic and facsimile transmission 17 3.2. Computer systems and communication networks 18 3.3. Store and forward techniques and delayed response.... 19 3.4. Technical problem areas 19

4. Audio and graphic inputs 21 4.1. Audio signal inputs 21 4.2. Graphic inputs 22 4.2.1. Alphanumeric inputs 22 4.2.2. Special symbol and diagramatic inputs. 23 4.2.3. Graphic inputs and on-line drafting 24 4.2.4. Pictorial and three-dimensional data input. 26

5. Preprocessing operations and pattern recognition 27. 5.1. Preprocessing operations: Image and information enhancement 27 5.2. Property filtering and feature extraction 29 5.3. Optical character recognition 31 5.3.1. Multifont character reading 31 5.3.2. Performance requirements and quality control 33 5.4. Recognition of handprinted characters, handwriting, geometric shapes, and line drawings 33 5.4.1. Constrained handprinting 34 5.4.2. Handwritten characters and handwriting. 35 5.4.3. Shapes and drawings 38 5.5. Speech and speaker recognition 38 5.6. Other pattern identification and recognition areas 42 5.7. Some theoretical approaches to pattern recognition and categorization 43 6. Conclusions 50 Appendix A. Background notes on research and development requirements in information acquisition, sensing, and input 53 Appendix B. Bibliography 151

/U/V List of Figures Page Figure1. A generalized information processing system... 4 Figure2. A generalized character recognition system 6 Figure3. Areas of continuing concern, information acquisition 11 Figure4. Areas of concern; information sensing, input, and preprocessing 21 Figure5. Model typescript page at expanded scale 25 Figure6. Two illustrative display formulas 25 Figure7. Image enhancement for noise reduction 28 Figure8. FOSCRIPT decoding criteria 34 Figure9. Patterns for different speakers 42 Figure 10. Fingerprint pattern showing bifurcation_.. an_d_ridge ending 44 Figure 11. Material read by Soviet print reader 56 Figure 12. Automatic preparation of smooth sheet 64 Figure 13. Data processing shared by two computers 80 Figure 14. Dickinson hand-print embellishments 113 Figure 15. FOSDIC alpha-numeric marking document 114 Figure 16. Handwritten numerals, vector crossing technique 116 Figure 17. OpScan document 119 Figure 18. PAPA machine assembly 144

. VI. Information Acquisition, Sensing, and Input: A Selective Literature Review Mary Elizabeth Stevens

Center for Computer Sciences and Technology National Bureau of Standards, Washington, D.C. 20234 This report, the first of a projected series on research and development efforts and requirements in the computer and information sciences, is concerned with a selective literature review involving the operations of information acquisition, sensing, and input to information processing systems considered in generalized terms. Specific topics include but are not limited to:source data automation and remote sensing techniques, communication systems and data transmission links, audio and graphic inputs, preprocessing operations upon input items such as image enhancement and property filtering, character recognition, speech recognition, and various other aspects of automatic pattern recognition. Supple- mental notes and a bibliography of over 640 cited references are included. Key words: Audio inputs; automatic pattern recognition; character recognition; communication systems; data transmission; graphic inputs; image enhancement; remote sensing; source data automation; speech recognition.

1. Introduction This report is the first of a series intendedto the Budget Circular A-71 and Public Law 89-306 providea selectiveoverview of research and (the Brooks Bill of 1965), it is not unreasonable to development efforts and requirements in thesome- try to provide a framework for program review and what overlapping fields of the computer and infor- to pinpoint, so far as is practical, specific needs mation sciences and technologies. The projected and difficulties to be noted in selective reviews of series of reports will attempt to outline the probable the literature and other appraisals of the state of range of R & D activities in the computer and in- the art. Such clues may often indicate the desira- formation sciences and technologies through selec- bility of further research and development efforts tive reviews of the literature and to developa whether Federally conducted, Federally sponsored, reasonable consensus with respect to the opinions or independently initiated. of workers in these and potentially related fields There is also the desirability of bringing together, as to areas of continuing R & D concern for research in one place, a single perspective with regard to program planning or review in these areas. In both current efforts and foreseeable future needs general, each individual report will be self-con- for research and development in the computer and tained, and will provide notes inan appendix geared informationsciences, however biased that one to the order of sections in the text. perspective may be. This series of reports is thus 1.1. General Introductory Considerations intended to provide background material for the identification of research needs as a contribution Itis,of course, presumptuous, as Licklider* to improved cooperation in the field. It is hoped and others have emphasized, to attempt to provide also that a "framework" for considering R & D directives for the course of research and develop- requirementsinthe computer and information ment planning in so broad an area as that of the sciences and technologies such as is proposed here computer and informationsciencesandtech- may provide a continuing basis for annual reviews nologies. Within the scope of Federal Government of continuing progress in the field. responsibilities,however,andwithparticular In this introductory section to this first report, respect to the specific responsibilities assigned to we will consider a few of the major background the National Bureau of Standards under Bureau of factors; outline the general plan of attack for the projected series of reports; provide some general A bibliography of all cited literature references, as well as other pertinent refer- caveats; discuss, with a few examples, a diagram- ences of interest, is inciuded as Appendix B. In addition, notes are provided in Appen- dix A for each of the discussion sections of this report. They represent selective matic schema of a generalized information process- samplings of corroborative opinions to be found in the literature, together with additional specific examples and citations to specific research and development ing system which will form theorganizational efforts to which we have referred. See note 1.1, Appendix A. framework for at least the first few of the reports in

1 sir =0, ,'ti'

the planned series; and summarize the scope and cuts into the available lead-time in which coverage of the present individual report. laissez-faireapplicationsorsolutionsare practical. 1.2. Basic Background Considerations (3)Exponentially, the recording of potentially applicableinformation,fromtelemetered Machines, especially radical new developments in data captured in outer space through lab- computer, communication, and information process- oratory measurement recordings to published ing systems, have made possible revolutionary literatureand computer-controlled graphic gains in data acquisition and data manipulation displayof theprobable consequences of over time, over space, over human limitations of decision choices, continues to increase in reactionandresponsiveness,overmanually- volume and in bulk. For every accretion to the inaccessible conditions, and more especially over sum of recorded, potentially usable infor- human limitations of span-of-attention, consistency, mation, there are corresponding accretions and objective judgment. to the problems of location, selection, con- When, in 1946. ENIAC was instructed to perform densation, and validation. ballistic computations (at high speed and great (4) Critically, in terms of resource utilization, accuracy), it was, in effect, substituting for many we need today an increased use ofpotentially human fingers on many desk-calculator devices. available information to avoid replication of In the early 1950's, computers could be used for effort, to avert false starts and blind alleys, or payroll calculation, inventory control, accounting to assist us by the suggestion of practical and bookkeeping, and similar processes, where tools, techniques, and theoretical methods. the substitution was that of the machine for, roughly, We will also need increased release of human high-school level actions and judgments of clerical intellectual capabilities from machine-do-able personnel. Today, the challenge is of at least an tasks in order to free these capabilities for the order of magnitude greater: how can machines truly creative and/or prudent action-choices substitute for, or at least provide effective aids to, that willstill require human judgment and the air traffic controller, the professional meteor- valuation. ologist and the laboratory scientist, the architect For these reasons, we have seen over the past or city planner, the referencelibrarian, the student two decades the phenomenal growth of whole new or the professor? industries involving computers, communication and- Pressures of time-gap with respect to human control systems, and information processing tech- response-time (whether measured in fractions of nologies. For these reasons, there have been sub- seconds or in terms of mastering the U.S. Patent stantialresearchanddevelopmentprograms Office backlog) dictate increasing use of machines. established in industry, the universities, non-profit Pressures of concern in the accuracy of strategic organizations and private foundations, and Federal forecasting, medical diagnosis, and predictions of Government agencies. Despite substantial progress actions necessary to avert foreseeable disasters in many of these programs, however, there continue (political, economic, social, ecological, and the like) to be many areas of research and development again indicate the importance of machine aids and requirements that are of major concern to the U.S. correctionstofallible and inconsistent human Government and inparticular to the Federal judgments. Thirdly, there are the related pressures Council of Science and Technology and its Com- to avoid duplication of effort, to shepherd the utiliza- mittee on Scientific and Technical Information tion of available scarce-manpower resources, and (COSATI). to challenge the most effective output of creative Of particular concern in the United States as of minds directed toward the Most urgent problems 1967-1968 are the information processing system of our times. requirementsintheanticipatedphenomenal The rationale for research and development in growth of multiple access networks and informa- the computer and information sciences and tech- tion processing utility systems, with concomitant nologies is thus manifold. Among the currently problems of shared data banks and the measures most critical considerations are the following: necessary to protect privacy on the one hand, and (1) Increasingly,theproblems and decisions withhigh promise foron-line problem-solving, facing man impose conditions of organized machine-aided design, and machine-aided decision- complexity, multiple interdependence making applications, on the other hand. of many uncontrolled and largely uncontrol- There are R & D requirements that stem from the lablevariables,astaggering number of desirability of foresighted consideration being given possible outcomes with their comparative to probable future standardization or compatibility payoffs largely unknown, and limitations of efforts; there are R Sz-D requirements that arise bothavailable-response-timeandhuman because of current difficulties and bottlenecks, and span-of-attention. there are those that relate to fact-finding, hypothesis (2) Steadily, the rate of technological advance, formation, and experimental investigation in areas and concommitantly that of growth in the where we know so little of the underlying basic magnitude of problem factors to be resolved, phenomena that it is not clear whether automatic 2 7.3r7.71FrV,Vri, -,1"' .-477 yst /1,

information processing techniques could be feasibly is a case of a relatively novelsource data collection, applied at all. processing.recording,andresponsesystem. In addition, special R & D requirementscan be Another exampleisarepresentativesystem foreseen in specific potential areas of application, designed for automatic character recognition. The such as library automation or question-answering third case involves both cGnv:ntional andnon- systems for the personal use of the intelligence conventional techniques for the handling of recorded analyst. For example, we shall turn to a more de- scientific and technical information theareas of tailed consideration of the problems of information traditional library science, mechanized documenta- storage, selection and retrieval (ISSR) systems in tion,andinformationselection,storage.and other reports in this series. In the first reports, retrieval systems. however, let us discuss some R & D implications in It may also be appropriately suggested that, in systemdesignrequirementsgenerally,with human analogy, information acquisition, sensing emphasis on certain recent developments in infor- and input operations are comparable to perception mation acquisition; input-output subsystems and and ingestion systems in the human; pre-processing links to communication systems;processor sub- operations to digestive actions; processing to both systems and multiple access systems in particular; metabolicandcommunicationfunctions;and storage subsystems; advanced technological selective recall and output to effective psycho- developments, and programming considerations physical reaction to the real world. including the special problems of on-line instrumentation and the mutual protection of clients and 1.3.1. Automatic Production Control systems. Any consideration of generalized research and In a relatively recent example of automatic development requirements in the computer and production and control systems, the Bevercotes information sciences should obviously be based Company in England has designed techniques to on one or more aspects of the fundamental infor- provide for coal mining operations under remote mationprocessingcycle theacquisitionof control. Information acquisition, input and sensing information; the processing of information to find, operations (such as those indicated in Boxes 1 and 2 fix, and focus significant features present; the of Fig. 1) will involve the use of a nuclear sensor that recording and storage of processed information emits gamma rays. The backscatter rates can be for subsequent use, and the selective recallor measured (preprocessing and processing opera- retrieval of processed and stored information for tions in Boxes 3 and 4) to determine the thickness use. of the mine walls around coal veins and the thick- Certain themes will be necessarily recurrent: (1) ness of coal remaining in a given vein. Direct output System design must becomemore and more inte- (Box 5 of Fig. 1) is to the automatic guidance control grated and more andmore dynamic and flexible; of the cutters, with provision for stopping them just (2) Effective man-machine interaction at hardware, short of the rock protecting the mine itself. Other software, and behavioral factor levels isan in- inputs are processed to provide automatic monitor- creasingly significant problem; (3) Programming ing of methane gas levels and to adjust ventilation languages will require bothmore formal theoretic equipment. Still other outputs (Box 14) go to human and also more user-oriented development, especially operators who govern, by appropriate processing in the areas of man-machine dialog, parallelprocess- specifications, the movements of hydraulic chocks ing, multiprocessor systems, simulation, and self- that follow the cutters and support the mine roof. diagnostic capabilities; (4) The essential interplay (Electronics 38, No. 20, 213 (1965).) of man-machine reaction requires that muchmore This experimental productioncontrol system must be realistically learnr d of man-in-the-system designthusemphasises automatic information needs, of human advantages, frailties and prejudices acquisition (Box 1 .of Fig. 1), sensing (Box 2), pre- in perception and apperception, and basicprocesses processing (Box3),post-processing by human in man or machine procedures for pattern detection, operators (Box 13), and outputs (Box 14) fed directly concept formation and amalgamation, the tracing of back to the system itself as, in effect, processing associative retrieval trails, and the application of service requests (Box 7) which may be used to both deductive and inductive inference. develop processing specifications, suchasthe required amount of adjustments (Box 9) whichare 1.3. A Generalized Information Processing then "matched" (Box 10) with the actual processing System operationsneededtoeffecttheadjustments. 1.3.2. Character Recognition Systems For preliminary purposes of indicating trends and defining certain obvious research and development In the generalized character recognition case, requirements we provide in Figure 1 a generalized information acquisition (Box 1) involves the receipt function chart of information processing systems. of a carrier medium (such as paper or film) in oron The intended generality of Figure1 can be which the character or characters to be'read have illustrated by tracing through several examples of previouslybeeninscribed,typed,orprinted. information processing system applications. First Box 2 of Figure 1, input and sensing, involves the 3 1. Information Acquisition r

2. Information 7. Processing I Sensing an r Client Service I Input Requests I

I Preprocessin 8. Preprocessing I

I 10. 4. Processing Matching i9. Processing I Operations Specifications

I I a. b. I On-Line Stor- Storage age I I

11. I earch a Selec - I tio I

I 12. Retrieval 13. Postprocessin i //14.Outputs I I

15. Use and I

Evaluation I

111,11200 OM/0 =NM 110 IP 111.10 =IMO MN= =MM.11= =11=1 .I11111. INIMMIO AMMO MEMO MMON 11,

Legend: process flow

=11 ONO 0/10 feedback flow

FIGURE 1. A generalized information processing system.

4 positioning and illumination of the carrier on which scale acceptance level so that more "gray" is the character is recorded and the physical sensing sensed either as "black" or as "white") and pre- of a source patternfor example, by optical pro- processing operations, especially those which affect jection to a bank of photocells. By preprocessing localneighborhoods(orsubsets of the source operations (Box 3), such as integration o.-,er photocell pattern image areas) are of particular importance response to particular subareas of the total source and of significant concern for R & D efforts in this pattern image area, the sensed source pattern is area; to be discussed in later sections of this report. transformed into an input pattern suitable for the subsequent processing requirements of the system. The processing requirements are usually in the 1.3.3. An Experimental ISSR System in the U.S.S.R. form of matching and "best-match" selection by Finally, although we shall consider information comparison with previously stored reference patterns storage, selection, and retrieval (ISSR) systems in (exemplars of the salient characteristics of the detail in one or more separate reports of this series, various members ofthecharactersetto be letus note a system combining optical char- recognized by the system) or templates (which, in acter recognition, (OCR) thesaurus look-up, semi- their simplest form, might be the photographic - automatic indexing, and machine expansion and negative images of the set of characters in a partic- reformulation of search query inputs, that is in ular font and type size). experimental operation at VINITI, the All-Union Such reference patterns (whether merely simple Institute of Scientific and Technical Information of templates or sophisticated pattern-property require- the Academy of Sciences, Moscow, U.S.S.R.4a ment listings involving syntax and context) serve Information acquisition at VINITI relies on the the function of processing service requests (Box 7) submittedabstractsof scientific and technical in our generalized representation. For example, papers that are supplied by a corps of some 10,000 processing specifications (Box 9)inthe photo- Jr more scientists and engineers throughout the graphic-negative mask-matching system may require Soviet Union. Information input and sensing opera- check for matching against "F" before "E" in a tions involve, first, a preprocessing operation in the sequentially processed match-decision procedure. form of manual re-transcription of the original input Other possibilities involve feedback to threshold citations and abstract texts, using a somewhat settings and the directed re-scanning and repeti- stylized font (Cyrillic) typewriter. The typed mation of pr:Trocessing operations to enhance signal- terial is then read by optical character recognition to-background contrasts 6: to eliminate noise in equipment and is fed to a computer for the experi- the case of "0" and "Q" in either sequential or mental preparation of author and title indexesb parallel matching operations. The input text has been limited, at least up until In a simple mask-matching system, the matching late 1966, to bibliographic citations and short ab- operations of Box 10, Figure 1, require that the stracts for approximately 25,000 documents in the suitably preprocessed input pattern will be optically field of electrotechnology. The processing opera- superimposed against the various reference patterns tions for this input material, whether manually and, when this input pattern coincides with its prepared in machine-usable form or automatically proper negative, light will be extinguished to a par- read by the OCR equipment, involve operations of ticular photocell. The resultant "blackout" (extinc- analysis and indexing by computer using thesaurus tion of light to the photocell), together with suitable look-up to identify related terms, synonyms, and means for identifying the particular pattern for the like, and also to provide word-by-word transla- which the coincidence occurred, can then be used tion of document input to a numerically encoded toeffectappropriate post-processing operations documentary language, which is that of a coordinate (Box 13 of Fig. 1), such as to punch the appropriate descriptorsearchsystem, involving 3,500de- machine-language for the character so "identified". scriptors at the present time. These operations may alsoprovide an output, A machine-based, 4,000-item,Russian word- (Box 14), which may include the information that the stem dictionary is used. For each text word, the system was unable to decide as between two or appropriate stem is selected. Words not found in more ambiguous input patterns.(It may be of the dictionary are printed out, so that misspellings interest to note that such a character recognition and possible new words can be caught on manual scheme was at least implicity disclosed by Goldberg inspection. Print-outs may also be made of possible as early as 1931.) beginnings of complex expressions (e.g., where For later and more sophisticated automatic char- "installation" or "resistance" may be beginnings of acter recognition systems, Figure 2 illustrates how phrases rather than single words). If a word found the basic schema of Figure 1 may be expanded, in the dictionary is not marked as being a possible constrained, and made more specific in order to "complex" or "homograph", it is translated directly provide a generalized picture of automatic character into a 4-digit concept code. and pattern recognition processes. In many actual The documentary language used involves the or proposed systems for character Jr pattern recog- application of concept-codes or descriptors as man- nition purposes, the importance of feedback with ually established.Its code structure directly in- respect to threshold settings (e.g., to reset the gray- corporates certain semanticrelations, including 5 WEIGHTING FUNCTIONS I i

I SCANNING THRESHOLD de --, I

* 1 I s'6:) ERM ENR MEM 20FrRIOT PATTERN - - ---4 TO INPUT PATTERN 1 i 1 3. TRANSFORMATION 1 OF INPUT PATTERN TO 1 FORMAT OF REFERENCE 1

PATTERN 1 t i 4 1 SET REFERENCE 411 FATTERN SELECTION 7 1

I 5. INPUT PATTERN IELEMENTS: REFERENCE 1 Z REFERENCE PATTERN PATTERN STORAGE ELEMENTS I olIF 1FEEDBACK 1MATCHING THRESHOLD FUNCTIONS * I NO 6. ALL TRANSFORMATION I REFERENCE OF ELEMENT BY I PATTERNS ? ELEMENT MATCH TO eMINED 1111,-- - -41 IDENTIFICATION YES FORMULA i I if SET MASTER 1 IDENTIFICATION I SELECTION I I I 9. OBSERVED MASTER I IDENTIFICATION IDENTIFICATION MASTER I IDENTIFICATIONS L STORAGE I NO I RECOGNITION THRESHOLD f I

YES 12. SELECT TARGET PATTERN ---, LANGUAGE

15' REJECTS 13. TRANSFORMATION TARGET OF IDENTIFICATION I PATTERN STORAGE 1 DECISION TO TARGET PAT TERNS I1 TO FEEDBACK t FUNCTIONS

FIGURE 2. A generalized character recognition system.

6 some hierarchical relations, between descriptors. It may often be desirable to record relayed !- In addition, manually determined links or connec- formation in several different input sensing modal- tions between descriptors have been set up for ities in parallel. In particular, temporal interval retriei-al pnrposes. They have been tested byex- lapse-time data may be required for content and perimental searches for the total list of descriptors diagnosticanalysesofverbalprotocols1.5 but conducted independently but in parallel by machine also for "precise pin-pointing of time ofoccurrence and by subject specialists, and they have been of events of an oscillographic record ofan experi- subsequently refined for machine use on the basis ment regardless of the speed of the chart during the of analyses of noise, omissions, and desirable recording."1.6 Formechanizeddocumentation improvements. applications in particular "Verbal. graphic, alpha- The retrieval logic depends upon these associa- numeric inputs and outputs are intermixed and tional connections in the sense that, if a query term intertwined." (Licklider, 1965, p. 186.) does not find a direct match, then search is made The concept of effective interface involvesnot for match on related terms. If any descriptor in the only the prospects and the problems of man-machine question finds no direct or related term match with interactive collaboration but also those ofa variety descriptors assigned to the stored item, the item and intermixture of information inputs (by voice, is not selected. drawing, keyboard, andso forth)."a As in the Output of citation references for selected items case of a multiplicity of forms and types of inputs, is in the form of two listingsthose which the a generalized information system must also accom- machine considers positive responses, and those modate a variety of output modes andoutput media, which because of the tracings of possible associa- whether as direct outputs (Box 5 of Figure 1)or tional connections to presumably related descriptors as outputs derived from subsequent search, selec- are only probable responses. The machine also tion, and retrieval processing (Box 14). prints out listings of the connections that were It will be noted, next, that suggested techniques used in the search specification modifications so for a specific system may apply tomore than one that statistical analyses may be made as to the operational box or function shown in Figure 1. extent to which proper use is being made of these Speech synthesis developments, for example, using associatienal connections. Very preliminary results vocoder 1.7and other techniques,are primarily indicate that for 100 relevant documents in the directed to the output of artificially spokenmes- collection, about 55 are selected by direct match sages, but, at the same time, R & D efforts in speech and 25 by virtue of the pre-established connections. synthesis may have a direct impactupon improve- It is hoped that the latter procedure can be improved ments in speech recognition appropriate to informa- to 30 or 35 percent, to give up to 90 percent overall tion input and sensing. This is an approach currently recall. being explored at Gunnar Fant's laboratory, Royal Institute of Technology, Sweden, and by King and 1.3.4. Implications of the Generalized Scheme Tunis(1966)1.8among others. Certain features of the information flow and Similarly,ina givensystemconfiguration, process schema of Figure 1 are to be noted in these specific operations or processes mayoccur in dif- examples. It is assumed, first of all, that informa- ferent sequences and several differentones may be tion processin systems of the future should provide combined in various ways. Thus, for example, for automatic access from and tousers at many questions of remote console desideratamay affect points. This implies multiple inputs in parallel, original item input, queryor processing request interruptability,andinterlacingsofcomputer input, output, or user feedback. Further, the specific ,programs. Thus an obvious continuing research solutions adopted may be implemented in each of need is for the development of sophisticatedexecu- these operational areas,or combined into one, e.g., tive and monitor systems, to control tie-ins from by requiring all inputs and outputs to flow through communication networks, to manage scheduling the same hardware. Because of the difficulties in and priorities, and to provide for efficient utiliza- organizing the discussion of these reports, however, tion of system resources. the choice of the most appropriate placeor places to discuss a particular technique or topic has been Other major assumptions include the following: somewhat arbitrary. It is to be noted, in particular, (1) The overall system involves a hierarchy of that the separation shown as between "Information systems, devices and procedures. Input and Sensing" (Box 2) and "Input of Process- (2) Processing involvesmulti-stepoperations, ing Service Requests" (Box 7) is often, ina specific each step of which may be bypassed, cur- system, an artificial distinction. tailed, or expanded on the basis of man- Again, many of theprocesses can be iterative and machineevaluationoftentativeresults recursive at many levels. Thus, in the spectrum achieved at a present or immediately prior analyzer developed by King and Tunis (1966) fbr step. their speech recognition experiments, the results (3) Multi-mode operation is possible, depending of preprocessing operations(theinstantaneous on job requirements, prior or tentative results, outputs of a contiguously tuned bank of band pass accessibility, costs, and the like. filters) become, ineffect, inputs equivalent to 7 "Processing Service Requests". These are con- which is relatively unstructured. The computer, of tinuously compared (i.e., Box 10, Fig. 1, Matching course, must finally accept and file the information Operations) in order to locate the instantaneous in a highly structured form so that it may be re- peaks in the envelope of the speech spectrum." trieved efficiently. The computer can direct a pro- The iterative, recursive nature of the process cedure which allows the human to input the data in flow in particular exemplars of a generalized system the order and in the form which the machine can (Fig. 1) is especially to be noted in the Soviet ISSR accept." (1965, p. 19). example. Thus, preprocessing operations (Box 3) It should be noted, finally, that, in this and subse- occur first at the level of re-typing of input informa- quentreports, when we discuss the systems tion, and then, within the optical character recog- implications of Figure 1, we will in some instances nition part of the system, at the level of quantizing stress currently available examples, sometimes we the input pattern to a 32 x 32 matrix (of which about will be pointing toward foreseeable trends of future 300 black-or-white cells are actually used for recog- applicationofadvancedtheories,techniques, nition purposes). Additional preprocessing opera- media, and devices, and sometimes we will empha- tions involve the detection of significant "informa- size current difficulties. .tion areas" and may also involve the re-initiation and modification of both preprocessing and rec- 1.4. Specific Background Considerations ognition processing operations. For example, there may be re-scanning of special areas of the quantized The specific background considerations in the input patterns to decide between patterns that are preparation of the projected series of reports involve near to each other in the sense of pairwise confu- both questions of the plan of attack and of necessary sions. Such preprocessing operations may also be caveats which the reader should note. In particular, applied in holding a given character recognition- problems of organization of a wide range of mate- identificationdecisioninstoreuntilthe next rials, on a wide variety of topics, and with con- character has been read, so that context-predictive siderable logical overlapping between topics, do rules, based upon digram and trigram probability not appear to be soluble without many omissions, statistics, may be applied. without some necessary repetitions, and without Perhaps the single most important feature of the arbitrary allocations of specific topics and examples generalized schema (Fig.1) is the provision for of current R & D efforts to particular discussion threshold requirements under feedback control. Let areas. For example, questions of boundary and con- us consider, for example, some of the varied trast enhancement and property filtering in graphic threshold factors that might be applied to input- data processing applicationswill be considered scanning operations for various types of items to in the present report, but closely parallel operations be stored and subsequently retrieved. These might in mechanized documentation applications (such include the following: as synonym reduction or homograph resolution) (1) The "purge" (or "stop"), lists used in key- will be deferred for reports dealing more specifically word-in-context (KWIC) indexing to eliminate with such areas of application. Similarly, many "common" (and presumably "nonsignificant") of the R & D requirements affecting more than one words from further processing.l.w Modifi- or two of the Boxes of Figure 1, such as program- cations (changes in threshold) may be applied ming languages, character sets, or advanced hard- on inspection of results either to restore words ware developments, will be considered in a separate previously eliminated, but later considered report in this series, as valuable access points, or to add words for general elimination because of the unmanage- 1.4.1. Plan of Attack ably large blocks of index entries that they There are certain obvious difficulties with respect generate. to the organization of the material for a series of (2) Time-interval sampling of experimental sig- reports on research and development requirements nals to reduce total volume of data to be inthe computer and information sciences and processed, to eliminate redundancy, and to technologies. These problems stem, principally, "smooth" minor data variations. from the overlaps between functional areas in (3) Minimum gray-scalelevelquantizationof which man-machine interactionsof both com- graphic pattern input, such as printed or munication and control are sought; the techniques, handwritten characters. Here, a particularly tools and instrumentation available to achieve such sophisticated approach to feedback thres- interactions, and the wide variety of application hold control is suggested by the work of areas involved. For example, foreseeableapplica- Rankin et al. (1965) on the syntactic analysis tions range from astrophysical considerations of of Chinese ideographs where "wellformed- events that occurred many thousands of light years ness" criteria are applied to reduce ambiguities ago to investigations of on-going synaptic transfers of character-component recognition or errors in studies of human neurophysiology. Beyond this of either transcription or encoding. are many diversified areas of applied technology Then there is the further man-machine reactivity involving adaptive sensing, processing, and control discussed by Bauer: "Consider the input of data operations in many fields. 8 How can we order, much less appraise, the preparation of a series of reports such as is planned, massive body of experimental data and resultant certain disclaimers and observations as to the pur- literature that is already pertinent in these fields? pose and scope of this report, its necessary selec- This is a major problem in itself. The material that tivity, and the problems of organization and em- has been collected and reviewed to date is so phasis are to be noted. Obviously, the reviewer's multifaceted and so extensive as to require organi- interests and limitations will emerge at least in- zationinto reasonably tractable (butarbitrary) directly in terms of the selectivity that has been subdivisions. These are being grouped, first, in applied. accordance with the diagrammatic schema of a There is, nevertheless, heavy reliance placed on generalized information processing system given the experience and opinions of others, whether in- previously as Figure 1. side or outside of the Federal Government, and The first few reports planned for this series can whether their projects are or are not supported by therefore be quite clearly identified. In this first Government contracts and grants. In general, con- report we consider, selectively, current problems troversial opinions exprebsed or implied in any of and developments in information acquisition, sens- the reports in this series are the sole responsibility ing and input, and preprocessing operations with of the author(s) of that report and are not intended respect to computerized information processing in any way to represent the official policies of the systems (Boxes 1, 2, and 3 of Fig. 1). Next, we will Center for Computer Sciences and Technology, the be concerned with processor system management, National Bureau of Standards, or the Department of with the organization and storage of data (at pos- Commerce. However, every effort has been made il sibly many levels of storage and access), and with to buttress potentially controversial statements or processing operations upon this stored data as such. implications either with direct quotations or with Discussion of outputs from and use of generalized illustrative examples from the pertinent literature information processing systems will involve some in the field. Thus corroborative evidence, largely necessary overlapping of research anddevelop- from the ope! literature, is adduced, generally, in ment implications in the input-output and proc- terms of direct quotations in the notes of Appen- essing management areas. A separate report will dix A. be directed to some of the overall system design The author of the present report must apologize operations with respect to Figure 1. for a necessary selectivity and for many inadvertent Next. in view of widespread interest and concern, omissions in coverage. It will be appreciated if we plan to examine the domain of potentialappli- specific omissions are called to our attention. It is cability of automated information processing tech- to be noted that neither inclusion nor citation is niques to the improved utilization of scientific and intended in any way to represent an endorsement technical information and recorded knowledge. of any specific commercially available device or Then, a more detailed functional flowchart or system, of any particular investigator'sresults schema of information processing operations as with respect to those of others, or of named project applied specifically to information storage, selec- objectives. Conversely, omissions are in no sense tion, and retrieval purposes will be discussed in intended to imply adverse evaluations of products, one or more reports. media, equipment, systems, project goals and proj- Other specific topics under consideration for ectresults,or of bibliographic references not inclusion in this series of reports include the current included. statusof library automation developments, the There will be quite obvious objections to our problems of evaluationof information 'storage, selectivity from readers who are also R & D workers selection, and retrieval systems and subsystems, in the fields involved as to the representativeness and the questions of maintaining the integrity of of cited contributions from their own work or that privileged files (i.e., some of the background con- of others. Such criticisms are almost inevitable. siderations with respect to the issues of privacy, Nevertheless, these reports are not intended to be confidentiality,and/or securityinthe case of state-of-the-art critiques as such, but rather they multiply-accessed, machine-based files, data banks, are intended to provide provocative suggestions for and computer-communication networks). In general, the plan of attack in each individual further R & D efforts. Selectivity must also relate report of the series will be to give in relatively to a necessarily arbitrary cut-off date. This factor short and largely discursive text the topics of con- is in itself an illustration of the limiting factors cern, supplemented by an Appendix of notes and that affect the problems of preparing state-of-the-art quotations, and another Appendix giving the biblio- reports and critical reviews. graphic citations of quoted references. It is planned, The present series of reports, subject to the fore- however, that there will be a comprehensive sum- going limitations, is offered as a possible contribu- mary, bibliography, and index for the series as a tion to the understanding of the general state of the whole. art, especially with respect to long-range research possibilities in a variety of disciplines, potentially 1.4.2. Some General Caveats applicabletoinformation processing problems. Since problems of organization, terminology, and The reports are therefore directed to a varied coverage have all been difficult with respect to the audience among whom are those who plan, conduct, 9 and support research in these varied disciplines. a few earlier references have also been included They are also addressed to applications specialists where especially pertinent. who may hope eventually to profit from the results We will consider first selected problems of infor- of current research efforts. Inevitably, there must mation acquisition in the areas of source data be some repetitions of the obvious and over- automation, measurements automation and feed- simplifications of certain topics for some readers, back control, and remote sensing techniques. and there must also be some too-brief or inade- Next, there are a number of considerations involving quately explained discussions of other topics for communication systems and data transmission these and other readers. What is at best tutorial links as means of input to information processing for one may be difficult for another to follow. It systems, including questions of services, networks, is hoped, however, that notes and bibliographic and facsimile transmission and consideration of citations will provide sufficient clues for further some technical problein areas. follow-up as desired. Direct audio and graphic inputs, including the special cases of pictorial and three-dimensional data processing, are next considered. Preprocessing 1.5. Scope and Coverage of This Report operations such as filtering and image enhancement lead to topics of pattern recognition generally, In the area of information acquisition, sensing, includingopticalcharacterrecognition,voice and input (Boxes 1, 2, and 3 of Fig. 1), which we recognition and speaker identification, and other attempt to cover in this particular report, a selective areas of potential application. Finally, some theo- literature review has been conducted with emphasis reticalapproachestopatternrecognitionare on the 5-year period from 1962 to 1967.Some addi- consideredasportendingfutureadvancesin tional coverage is provided through mid-1968 and information acquisition, sensing, and input.

2. Information Acquisition, Source Data Automation and Remote Data Collection In the general area of R & D efforts affecting the world. Physical and environmental considerations operations of information acquisition, sensing, and are reflected, for example, in the physiological and input, we consider first, briefly, basic operations cultural factors that aggravate the problems of and input requirements and then discuss selectively automatic speech recognition. some of the problems of source data automation, Then, in the medical, biological, and physiological: measurements automation and feedback control, fields, there are systems for the sensing and trans- and remote sensing and data collection techniques. mitting of electrocardiographic and other data. On-line or real-time monitoring and reporting of 2.1. Basic Operations and Requirements patient condition, not only for the critically but also for interchange of information between Under the basic; operations of information acquisi- cooperating institutions," isan important area tion, input, and sensing we include such input items for further development and possible networking as: Data collected at source, including that auto- .applications.2:4 matically collected at remote locations; laboratory Figure 3 exemplifies one of several possible meth- measurements, on- and off-line; control and feed- ods of identifying and predicting R & D require- back information; photographic, acoustic, magnetic, ments in the computer and information sciences and light-pen, and other recordings of events and data; technologies. Thus, with respect to the problems auxiliary control signals; machine-usable data rec- of information acquisition raised with respect to ords; written records and reports; communication physically and environmentally restrictive condi- system messages; documents, document surrogates, tions, we may look to continuing efforts in science and feedback commentaries; processing service and technology generally for improvements in the requests; search, selection, and recall requests:, scope, accuracy, and effectiveness of physical instructions and data for revision, correction, and measurements. We may also count on develop- updating of information; processing control speci- ments and improvements by mathematicians and fications, and system operation and usage statistics. statisticians in both the techniques and the evalua- Areas of continuing R & D concern with respect tion of statistical sampling.* Problems and pros- to the first process shown in Figure 1, that of Infor- pects for R & D efforts in the computer and information Acquisition, are exemplified in Figure 3. mation sciences, however, are more directly con- First are the questions of the physical, environ-cerned with the questions of effective sensing mental factors that affect observations, experimental manipulations, and recordings of both physical It should be noted that, because of necessarily selective coverage in these reports, phenomena and various transactions in the human not all of the possible considerations shown in various figures and tables will be ex- social, cultural, political, economic, and ideological. plicitly discussed in the text.

10 Areas of Continuing Concern

1. Information Acquisition

Physical measurements Physical and Simulation Environmental Remote access instrumentation Conditions Statistical sampling

Resource analysis Market analysis Intelligence evaluation i Availability Improved communication techniques ! Considerations Improved dissemination techniques Requirements analysis

Improved sensor capabilities Sensor selectivity and filtering Data capture 11-- Detection requirements analysis I and collection Quality control of source data organization Feedback-controlled re-scanning

Improved transmission facilities Data Error detection and correction transmission Improved data filtering and compression

FIGURE 3. Areas of continuing concern, informationacquisition.

366-107 0 - 70 - 2 11 techniques and with the development of remote processing systems design begin with the source access instrumentation. data. Both "source data automation" and "remote Problems relating to the availability of informa- data collection" have been significant systems plan- tion dictate first the analysis of present and potential ning concepts in business and management infor- resources for its capture, recording. and use. Auto- mationprocessingapplicationsoverthepast matic information processing design requirements decade.2-7 with respect to information acquisition, sensing, and The concept includes hardware possibilities that input are of increasing concern because of the range from mark-sensing principles for Census-data rapidly growing trend to attach analog sensors and gathering and microfilming principles (as in FOSDIC transducers to computers. This enables the com- equipment for Census data-gathering and records puter to be used for space vehicle control, process management 29 or the taking of inventories of control, medical research, and the like. For example, stocks-on-hand, through both magnetic ink and telemetry monitoring of human reactions in unusual optical character reading devices, to the principle environmentsisprovocatively described in the of automating and integrating the entire publishing report of a 1966 panel discussion, covering, in partic- cycle from typing of the author's original manu- ular, the monitoring of high-altitude climbers, the script to computer-controlled typographic compo- study of head injuries to football players by radio- sition.27b The latter possibility includes not only telemetry techniques and of electrocardiographic the preparation-publication cycle for journal articles reactions of spectators, as well as aerospace ap- and books but also of machine-derived byproducts plications (Nevison, 1967). such as catalog cards, printed indexes, and ab- In general, the areas of automatic production stracts.2-8 control,numericallycontrolledmachinetools, In a theoretical sense, "remote data collection" and the like, will not be covered in this series and "source data automation" might well be re-. ,of reports. Several examples, however, are illustra- garded as synonymous expressions. In practice, tiveof three-dimensional data processing and however, source data recordings that are collected machine-aided design applications. In automotive automatically at remote or generally inaccessible design, for example, proximity probes have been locations fall within the area denoted by the first used experimentally to measure the length, width, of these expressions and, conversely, "source data .and height of clay mockups of new models and the automation" is typically taken to apply to manual measurements were then converted by computer to recordings of observations of data values whether provide punched tape for the control of a drafting by physical marking, embossing, hand printing, plotter2.6, and electronically. sensed photographs keystroking, or alphanumeric reactive display opera- of automotive parts have similarly been converted tions. to numerical control tapes for machining.2-6 In a special sense, source data automation implies We can single out certain of these areas of con- the mechanization of information-carrying records tinuing R & D efforts of interest, but we will hope at the point of origin, as in automatic employee to stress relatively new or novel examples (e.g., identification and time-clock recording systems.2.8a automatic or partially automatic remote sensing Another common method is that of by-product or and control systems) rather than the more conven- dual language recording as in the generation of a tionalareassuchasanalog-digitalconversion punched paper tape with the necessary information techniques. We may note, however, such recent about a transaction produced automatically in the contributions as the use of hybrid computers for proper machine processing code by the device that the analysis of feedback control systems, with records the transaction.2-6 special reference to the handling of non-linearities Automaticcharacter and symbol recognition in equations of neutron kinetics and heat flow as devices are obvious tools for source data automa- reported by Sanathanan et al, (1966). tion where the information to be processed has been Let us consider next, a few of the current poten- pre-recorded in printed, typewritten, handwritten, tialities and implied research requirements for bar-coded, color-coded and other visible forms. For source data automation, remote data collecting, example, the Honeywell data station accepts trans- sensing techniques, and tie-ins to communication fer of information from optical bar code readers 2.10 networks feeding input to automatic processing and IBM's System/360 provides for input from two networks. different optical character readers, as well as mag- neticinkcharacterrecognitiondevices(Data Processing Mag. 7, 290-302 (1965).) Some character 2.2. Source Data Automation and pattern recognition processes may be applied directly on input, as in the sensing of optical There is first of all the question of the physical marks. A special feature of machine-controlled relationships of the information processing equip- pattern recognition capabilities is represented by ment to the outside world, to people, and to other the Myocoder system developed by Philco's Bio- machines via a very wide variety of direct and in- Cybernetics Engineering Laboratories.2.12 direct communication and display devices. Ob- Othersuggesteddevelopmentsof automatic viously, such relationships with respect to data recognition techniques include checking of labels 12 during packaging processes.2-13 An interesting sim- areaidedbyremote-sensinginstrumentation. ilar suggestion for computer systems management Particular emphasis is placedupon the reporting is as follows: of observed measurements in machine-usable form "What is needed is for the manufacturers to as close to the information or data source as developsome system whereby anexternal possible. This involves the reduction of observed graphic label can be prepared and attached to the values,togetherwithappropriateinformation reel and subsequently read by the tape unit. An about the conditions of observation, toa message- adhesive tape with reflective dots and dashes language that can be fed into appropriate data- encoded on it would be one possibility." (Lynch, transfer or communication links, transmitted, and 1966, p. 122). subsequently processed by automatic dataprocess- ing equipment. Recording and input techniques System-performance monitoring applications of for these purposes will be considered in Section source data automation techniques are exemplified 2.4.2. by a Sylvania system, developed in collaboration Improved techniques ofsource data automation with the Association of American Railroads for the and interrupt processing increase the potentialities automatic identification and inventory control of ofdatacollectionunder machinecontro1.2.15a freight cars.213a including those problems of collecting data from Source data collection for direct transmission to relatively inaccessible locations where conventional a central processor or processor network is also instrumentation generally fails either because of achieved by touchtone dialing and voice telephone physical dimensions, interference with the phenom- inputs.2.14 More typically, automatic source data ena being observed, or the fact that the phenomena collection systems involve time clock and other are occurring in the remote reaches of space.2.15b measuring and identification device inputs in a Similarly, in the electric utility industry increasing variety of business and industrial applications. One attention is being given to combined computer/ example involves collection and scheduled forward- telemetry systems 216 ing of work hour distribution data from physically General-purpose dataacquisition and control separate production facilities to a processing center systems, such as Digitronics Diai-O-Verter 2.17 or for computer report generation.2.15 IBM's 1800, can "monitor an assembly line, control For another example, the source data acquisition a steelmaking process or analyze the precise status system under development for the U.S. Post Office of a missile during test firing, [and] can gather Department will link more than 5,000 data gathering, information at the rate of eight billion bits a sec- transmission and measuring devices in 75 major ond," 2.18 post offices to two collection and processing centers. Next, centrally located processing systems may (Commun. ACM 10, No. 2, 132 (Feb. 1967).) accept digitally encoded input messages via a In 1961, eight major members of theaerospace variety of information-carrier recording and trans- industry established an informal study groupon mission channels. Such techniques typically in- data transmission. The needs for data transmission volve both the recording and encoding of messages were surprisingly similar for different members of at the source of transmission and decoding and re- the group. It has been reported that these mutually recording at the destination. In many cases, several recognized needscouldbe describedinfour levels of recording may be involved in order to step categories, as follows: up to progressively more stringent performance "a. Loadsharingamongmajorcomputer requirements, e.g., as in increasing the speeds of centers ... information transfer, through various stages, from "b. Data pick-up from remote test sites (or from the very slow (several characters a second only) airbornetests).Insome cases real-time rates of manual keyboarding to the many thousands processing and retransmission of results to of characters per second rates acceptable to modern the test site would be desirable. computers. "c. Providing access for Plant A to a computer center at Location B. Plant A might have a 2.3. Measurements Automation and medium-scale, small-scale, or no computer Feedback Control of its own. "d. Data pick-up from dispersed plants and 4, second type of example of source data automa- offices for processing and incorporation in tion and/or remote data collection is found in auto- overall reports. The dispersed points might matic measurement techniques that provide for be in the same locality as the processing dynamic capture of data and for the feedback con- center,orpossiblyas much asseveral trol of on-going experiments or tests.2-i" In the area. thousand miles away." (Perlman, 1961, p. 209.) of laboratory measurements automation, McGee and Petersen (1965), for example, discuss a micro- Other' thanfor telemetry, data capture and program control system for processing, display, and collection technologies, as of 1967, still predomi- output printing of information derived from input nantly require human observation, monitoring, and data from CRT scanners, counters, pulse height recording processes, whether or not these processes analyzers, telemetry converters, and the like.2.19 13 Other projects for automatic measurement of filmed Again, at the Argonne National Laboratory, a data of nuclear particle events include an on-line variety of laboratory measurement instrumentation man-computer system at the University of Illinois,2-20 techniques have been developed. These include on- and a system at Rutgers University.'' -' -2' line interrupt facilities for users of the accelerators Other recent examples include the on-line collec- and film scanning and track counting devices for tion and analysis by computer of data derived from analysis of o:Ita such as bubble chamber records.2.29 nuclear physics experiments using the University of Neilsen reports on radioactivity measuring instru- Wisconsin'standemelectrostaticaccelerator mentationtechniquesusingaudioinputand (Commun. ACM 10, 385 (1967), work at the Brook- output techniques for transmission to and from the haven Laboratories of the U.S. Atomic Energy laboratory and a computer system.2.3° Commission (Rabinowitz, 1968),2.2" and atthe Earlier developments in telemetry; remote data Lawrence Radiation Laboratory (White, 1968). In collection, data logging and transmission; auto- addition, the CHLOE film digitizing system de- mationof laboratory measurement procedures, veloped at Argonne has been applied to spark cham- and on-line processing of remotely sensed data ber data (Hodges, 1968) as has the PIP device.2-2'b are being extended in many areas of experimental Another approach isthat of PEPR (Precision application.230a One such area is that of the auto- Encoding and Pattern Recognition). 2.21c matic control of experiments and tests, based upon Another laboratory measurements automation rapid transmission, processing, analysis, and feed- example is provided by an IBM system for crystal- back with respect to the input data. In particular, lographicanalysis.2.22 The Universityof Man- "the value of rapid transmission and processing of chester computing center also uses x-ray diffractom- test data to allow decisions to be made while tests etertechniquesforon-linemeasurements.2-23 are in process has already been proven in the Cooper (1964) points out that while measurements development of jet engines and rocket motors." automation might be expected in x-ray diffraction (Perlman, 1961, p. 210.) analysis where thousands of repetitivemeasure- It is also to be noted that "machine-collected data ments of reflections must be made, automated are not only more accurate (being less subject to techniquesarealso being applied in complex errors in reading, writing, or transcribing) and more experimental situations with man-machine inter- conveniently obtained, but make possible experi- action.2-24 Cooper also cites the PEXRAD (Pro- ments that produce data too fast, too infrequently, grammed Electronic X-Ray Automatic Diffractom- too irregularly, or over too long a period of time to eter) designed by Abrahams at the Bell Laboratories be taken by a laboratory staff." ("Specialized Com- and work on neutron diffraction measurement and putational Equipment from General-Purpose Mod- analysis at Brookhaven. ules", 1967, p. 175). Other examples of measurements automation 2.4. Remote Sensing and Data Collection and feedback control include the following: Techniques Current experimental evidence indicates many (1) Data Analysis by computer interfaced with potentially fruitful approaches to an extremely wide gas chromatograph equipment.2-25 diversity of source data automation applications, (2) Use of infrared radiometry to detect flaws in with or without human monitoring and interaction. multilayer printed circuits.2-26 (3) A Public Health Service grant has been made In addition, relatively direct sensor-processor link- to design and construct a scanner-computer ages are being explored, especially for photographic, for automatic observation and analysis of auditory signal, radar,2.3" and other electromagnetic medical and biological specimens to identify sensing, recording and transmitting potentialities micro-organisms and, further, for the com- that are being investigated. A data logging example puter to"direct the automatic petri dish is provided in a 1963 NBS project to develop and machine to spray penicillin and a variety of test an automatic data acquisition system for hydro- other drugs of the growing microbial colonies graphic survey data, a high precision plotter to locate position fixes and to print out information on in order to determine the drug sensitivities the soundings taken, and a computer program to of possible disease-causing organisms convert the position-fix data into earth coordinates which are found." 2.27 and then into x-y coordinates fed to the plotter.2.31 At the National Bureau of Standards computer An example of commercially available modular scientists and engineers have worked on a variety data logging equipment is provided by Berkeley o'l measurements automation projects designed to ScientificLaboratories.2.32 Mobidac, for another assist other agencies of Government and to serve example, is a mobile data acquisition system de- other R & D programs at NBS. In particular, they signed for use at remote test sites?33 have stressed a modular systems design program More recently, large-scale data collection arrays in which various previously developed components and networks have begun to emerge. These include and sub-assemblies can be assembled in configura- seismic arrays for the detection of earthquake and tions dictated by different requirements in different nuclear tremors 2.34 and a communication system areas of laboratory experimentation.2.28 used forthe monitoring of water pollution.2.35

14 Meteorological observations are an important area of systematic investigations, at the earliest possible of application.236 Then it is noted that "the Na- date, of overall source data acquisition require- tionalMeteorological Service System and the ments in a variety of subject fields for the foresee- World Weather Watch are planning the real-time able future and of both available and potentially data-acquisition system that will ensure continued available instrumentation, devices, techniques, and improvements in the future." (Aron, 1967, p. 65.) systems suitable for application. In particular, it Here again, system requirements for rapid process- is increasingly necessary to consider new research ing of data received from multiple sensors and for and development efforts in areas where "space real-timeanalysisof sensed data are empha- flight and hostile environments now imp )se require- sized.2.36a ments on logical elements of adaptation and self- Some of the more ingenious recent developments repair." (Swanson, 1966, p. 83.) in source data automation, processing, and control Space vehicle instrumentation and telemetering includetransmissionofvideophotographsof as such we have perhaps somewhat arbitrarily Mars 2.36b and the use of a 7094 computer not only excluded from the general scope of this report. to improve the quality of lunar photographsbut Nevertheless, we note that "our ability to acquire also to prepare three-dimensional contour maps of data is so far ahead of our ability to interpret and the moon's surface (see The Computer Bulletin manage it that there is some question as to just how 9, 62.) 2.37 Nathan, (1968) discusses picture enhance- far we can go toward realizing the promise of much ment procedures for television imagestransmitted of this remote sensing. Probably 90 percent of the from the moon and from Mars. data gathered to date have not been utilized, and, Hunt (1967) describes a laser system under de- with large multisensor programs in the offing, we velopment for the Air Force Cambridge Research face the danger of ending up prostrate beneath a Laboratories in cooperation with NASA, involving mountain of utterly useless films, tapes, and charts." the reflection of laser beams from the lunar surface (Parker and Wolff, 1965, p. 31.) Moreover, a con- for purposes of detecting small variations in the tinuing R & D requirement for adaptive sampling motion of the moon and of measuring earth-moon techniques is noted.2.42a distances.2.38 Remote sensing techniques also under advanced Thus R & D requirements for the future include development include those using gamma rays, firstthe severe problems of sifting or filtering seismic waves and measurements of magnetic enormous masses of remotely collecteddata.2.43 and gravitational fields (Parker and Wolff, 1965). Another indicated area for research in remote data Among examplescitedaremultihandrecon- collection and source data automationisthat naissance,infraredmappingsofforestfire "computer designers should search for a particularly boundaries, and multi-frequency radar detection of adaptable form of sense input,....Say a sense soil conditions down to. 18 inches or more below organ which could perceive spatial outlines of solid surface. 2.39 An earlier paper by Cantrell (1964) objects from a distance." (Strom, 1965, p. 111.) A emphasizes that as a new remote sensing tool an special-purpose development in this direction is to infrared scanning system can provide the geologist be noted, however. Thus, Kroger reports on a with unique information about the surface features helmet-mounted observationcontrol system, in of the earth. Similarly, remote infrared scanning which "...polarized light sources and sensors techniques may be used from spacecraft or satellite are used to determine the head position inthree todetect changing patternsof crops,coastal rotational axes, locating the target in relation to hydrologic features, and urban populations.2.39a the pilot and his aircraft." (Kroger, 1965, p. 269-270.) We note further that novel remote information In addition Parker and Wolff point out that "there acquisition procedures such as detection of cre- are great gaps in our knowledge of spectral emis- vassesinsnow-coveredsurfacesbyinfrared sivity and reflectivity for natural materials, espe- measurements (Parker and Wolff,1965)2-40;a cially at wavelengths outside the visible portion radio-wave water prospecting system in the Soviet of the spectrum...." Further, "right now, not Union 2.41, and measurements of the patterns of enough is known about natural materials to be cer- light-reflecting propertiesof water and carbon tain of the role radar can play in identification...." dioxide frost to be used with space probes of (Parker and Wolff, 1965, p. 27.) Venus 2.42) are merely scratching the surface of a Finally we note that in the space-data collecting potential wealth of future applications involving area, for example, current new development con- appropriatecombinationsofsensordevices; siderations (to be discussed in another report in communication links; pre-processing operations to this series) range from small physical size, com- select, extract and enhance significant data, features pactness, integrated circuit, batch fabrication tech- orproperties;computerprocessing;efficient niques and components withultrahigh-speed storage, and subsequent search and selective switching, storage, and read-out capabilities to fit recall, retrieval or reproduction. the requirements of spacevehicle dimensional and Nevertheless, the very variety and ingenuity of weightconsiderations,through'fail-softly'or these isolated examples suggests the importance `graceful degradation' provisions and multiprocessor

15 design2'44to the experiments in optical interspace sensing, and input.Itis noted further that "a message transmission tried out byGemini VII with variety of adapters are being provided to permit laser beams. the connection of computers to [communication] As was indicated in Figure 3 and in several of lines of differing qualities. This permits the com- these examples, improved data transmission fa- puter to be used as a message switching center cilities and both error detection and correction and it also permits separated computing installa- capabilities are among the areas of continuing R & D tions to be linked together into a giant multi- concern with respect to informationacquisition, processor." (Clippinger, 1965, p. 209.)

3. Communication Systems and Data Transmission Links The information sensing and input area broadly Map "Compilation System 3'4or the Automatic- defined to hiclude the typical physical relationships Picture-Transmissionrecorderwhichtransmits between the machine and its environment obviously standard weather maps in one mode of operation and involves the interconnecting links by which the records direct transmissions from meteorological machine system receives both information and satellites in another mode.-4a processing instructions and then communicates the results of the processing operations thatit 3.1.1. Voice and Digital Data Transmission performs. Of particular interest is the development of hardware and communication linkssuitable Considering first the case of digital data and voice for timed-shared and remotely accessed networks transmissions, currently available services include and systems. For example, the July 1%5 issue of the teletype, with both upper and lower case capabilities IBM Journal of Research. and Development provides promised for forthcoming models; voice grade a sampling of research in the communicationslink telephone lines for both data and voice transmis- aspects of such systems, with an overview by sion;3'41)broadband systems, and links via com- Andrews (1965). The technical papers cover such munication satellites. 3-5 In some cases, duplicate aspects of the problem as pseudonoise modulation transmission may be made over more than one type techniquesthatenablebetter performance in of communication link, such as via microwave and random-access communication systems where many also by coaxial cable.3'8In addition, there are a users communicate simultaneouslywithout going number of special arrangements such as private througha centralcontrol 3-1and questions of telephoneexchanges, 3-7Wide Area Telephone codes for optimum trade-offs between error rates Service,3.8and Telpak.3."One example of the use and speeds of transmission.3.2 of the latter service is for data convertibility be- Under overall system design considerations for tween computers.3'10 input to generalized information processing systems, Then there is the Dataport, a terminal device that we must obviously consider the R & D requirementsis claimed to be capable of being 'plugged in' to for improved and more versatile data transmission any computer in the country.(Commun. ACM 9, and data communication systems.3.2a Increasingly, 705, 1966). Increasing use is being made of Data- the computer and communication technologies are Phone, whether or not including automatic calling becoming interdependent. 3.3 In general, we will or polling mechanisms, 3."and of Touch-Tone reserve consideration of communication-computer telephones.3'12Davidson (1966), has described an system complexes, especially for multiple-access economical way to add alphanumeric input capa- applications, for a separate report in this series. bility to the Touch-Tone keyboard. 3.13 A similar However, in the sense of some of the areas of R & D approach is taken in the Tele-CUPL telephone concern with respect to information acquisition, as input time-sharing system at Cornell University. shown in Figure 3, we will outline briefly below some (Conway and Morgan, 1967). When coupled to Auto- of the problems and prospects in communication matic Calling Units,. Data-Phone service allows systems and data transmission links. computer-to-computer transmissions without the need for human intervention.3.14Higher data rates 3.1. Communication Links, Services, and for Data-Phone services are now being offered.3.14a Systems In particular, high speed burst transmission of data isprovided for multiple users with Data-Phone Commercially available communication links of 50.3.14b varioustypesprovidetransmission-inputand The use of communication satellites for both data output capabilitiesfordigital, voice, and even and video transmission is becoming increasingly graphic information items (e.g., facsimile trans- commonplace. 314e McManis (1966) predicts sig- mission,closed-circuitTV).Forexample,in nificant gains in communication channel capacity addition to the relatively well-established techniques from further developments of synchronous satellites for the facsimile transmission of weather maps, we and from the use of lasers.3.15Goettel, however, is note much more complex systems as: an Automatic somewhat pessimistic as to the time scale. "A fur- 16' ther step into the future,the laser, holds great with emphasis upon use by the newsservices and promise for both voice anddata transmission. The for weather map transmission.326b laser uses a beam of light to carryinformation. Crooks, reporting at the 1954annual meeting of already been done the American DocumentationInstitute, cited the Although a great deal of work has U.S. Air in this area, years ofcontinued research remain use ofTimes Facsimile equipment by the laser-transmission system Weather Network, the use by newsservices of before an economical Western Union becomes a reality." (Goettel,1966, p. 193). Mufax and Hogan facsimile systems, developments capabilities for transmitting onelettersize page Other advanced communication in 45 seconds, aimed at higher capacity,higher speeds, or lower between New York and Washington costs include pulsecode modulation systemsfor a Japaneseexample involving 3,000 ideographs, the Federal Reserve Bank system,and a publication voicetransmission, 3.16millimeterwaveguide in microwave application at McCall's involving the transmissionof techniques, 3.17 and improvements publisher to the printer and the technology. 3-17a Higher speedsof data transmission raw copy from the via voice and datanetworks, sometimes accom- return of proofs.3.27 reductions in error rate, are Ten years later, in 1964, Mayopredicted that the panied by significant transmission of also under development.".18 The speed mark-up economics of "rapid error-free already demonstrable diagrams, signatures, and tests will,I feel, sub- for textual data transmission (1964, p. 78). In is to be noted in contrasting a1953 report of the stantiate the apparent high cost." of the text of Gone With the following year, 1965, Gentlereported on a hand- long- distance transmission and the Wind in a few minutes3" with new technology writing transmission system for both messages transmit the contents hand-drawnsketches.328 By 1967,handwriting that would make it possible to voice com- of Webster's New IntercollegiateDictionary in less transmission combined with two-way munication could be demonstrated.329An interest- than a second. 3.20 networks, ing example of the use of regulartelephone lines In terms of communication system for transmission of graphic informationis provided Swanson(1967)reportsthat the U.S. Army's character recognition workload of approxi- in a system for remote optical STRATCOM with an annual proposed by Cogit-:c.ts Corporation. The system mately 60 million messages tiestogether various locations and countries and provides involves desk-top st,...ners at remote agencies in more than 30 a central recognitionunit capable of handling multi- connections to other Governmentnetworks includ- correction.329a In ad- digital network) 3.2' and font characters and on-line ing Autodin (the automatic colorfacsimiletransmissionisnow Autovon (for voicecommunications), with plans dition, communication satel- available3-29b under way for greater use of Currently available techniquesfor graphic and lites.3.22 Increasing cooperationbetween Federal Long Distance exemplified in the case of facsimiletransmissioninclude and local governments is Xerography(LDX)3.3°,wideband services 3.31, the National Crime InformationCenter system.3.23 Datafax Government interest is Magnafax3.32, Stewart-Warner Dial A further example of U.S. equipment 3.32a, Muirheaddevelopments 3.32b and the Federal GovernmentARS (Advanced Records Network with tie-ins to Picturephone 3.33,as well asTV transmission. System) Data Transmission Facsimile transmission links maybe combined with teleprinter, facsimile, and otherfacilities as well as voice message capabilities.3.34It is noted further public carriers, Telex, and theAutodin system. This be digitized and en- of social that "facsimile signals can network is used extensively in support crypted for secure communications overtelephone security and medicare programs3.24 lines." (Schatz, 1967, p. 3). As inthe case of digital In the planning for the proposedEDUNET system higher speeds of fac- universi- data transmission as such, (interconnecting a number of cooperating smile transmission, e.g., ofTV signals, are being ties and other ediicationalinstitutions), the discuss- sought by means ofdigitalization.3.35 Facsimile ants considered avariety of communication media transmission of information orimages stored in from teletype to video, with theconclusion that each Three examples are difficulties of cost, speed,technical microform is also available.3.36 type presented theAldenElectronicand Impulse Recording or administrativeproblems.3.25 Equipment 3.36a,theGeneralPrecisionMicro- In general, communicationlinks for digital data system as alphanumeric data televiser 3.37 and the Alden/Miracode transmission are required for used, for instance, by thehelicopter repair shops of and for digitized forms of speechand television of there are increas- the Army Materiel Command.3.38 facsimile signals.3.26 In addition, Special purpose facsimiletransmission networks ing requirements forhigh-speed, but economical, video information capabilities. include those providing medical graphic and facsimile transmission to physicians; 3.38athose linking libraries andother information centers, and thoseinvolving cooperative Transmission . 3.1.2. Graphic and Facsimile use ofeducational TV circuits.3.39 Thepossible use of closedcircuit TV mobile networksfor a Although facsimile transmissiontechniques have is also to be been available for more than a century,it has on4 variety of law enforcement purposes that facsimile com- noted.3.39a On at least anexperimental basis, link- been in the past dozen years cooperating libraries usingclosed- munication systems havebeen widely used,3.26a ages between 17 circuit TV and other facsimile transmission tech- tion complexes as used in university-oriented and niques have been under investigation forsome years. other multiple-access system, especially forre- Inparticular,thedevelopment of cooperative search purposes, will be discussed in anotherreport "telereference" services has beena continuing in this series, a few additional examplesmay be concern of the Council on Library Resources.3-4° Oven here. The CLR annual report for 1966comments, how- Examples of U.S. agency networks indicate the ever, as follows: "The demonstration of closed- potential significance ina variety of applications. circuit TV between libraries at t:-: University. of First is the interagency rapid communicationstest Virginia in 1957-8 identifiedsome of the problems. link that transmits data via teletype and dictaphone A demonstration in 1966 involving Reno and Las from survey ships operating atsea to the Netienal Vegas, Nevada, and Davis, California, showed Oceanographic Data Center.3-48 On another front, progress in some particulars since the earlier data ARPA's Nuclear Test"Detection Officeis spori,,oring with further improvementstillto be sought." the development ofa seismometer array and central (Council on Library Resources 10th Annual Report, data processing facility for the detection of under- 1966, pp. 24-25). ground nuclear tests. (Maguire (1965)4 Athird In the Houston Research Institute studies (1964- example is a Department of State proposalto link 1965),3.41 it was found that facsimile transmission of together officials stationed in the U.S., Canada, technical informationwas feasibl e via pipeline and Mexico .3 49 Then it is noted that "in law enforce- microwave network links, especially for off-shift ment a full scale nationwide system time shared periods, but it was concluded that "the establish- between Federal Government andstate agencies is ment of a facsimile transmittal system is not justified currently in operation. This system isa typical until major improvements have been made inpres- example of a computer utilitytype application of ent library procedures for information storage and dataprocessingwithintheFederal-State-local retrieval." (Schatz, 1967, p. 10). government complex." (Johnson, 1967, p. 14). Other recent library experiments in theuse of Applications in business and industry include facsimile transmission techniques include those first the air and space industries with, in particular, conducted by the New York State public library airline reservation services 3.50 andcomputer load- system 3.42 and by the John Crerar Library.3-43 For sharing capabilities.3-5' Banks 3.52 insurancecom- library applications, however,a requirement noted panies,3.53 and hospitals 3'54are among the customers in 1953 that for direct copying and transmission for the newly emerging computer-communication from bound book pages is still largely unsatisfied system complexes.354a The communication industry today.3-44 is itself exploring increased provision ofcomputer services to its customers.3-55 3.2. Computer Systems and Communication At Case Institute of Technology, Univac 1004 Networks reader/printers are "connected to the 1107 viaa Direct input to computer from remote locations half-duplex telephone line with 201A MODEM- units at each end.... via communication linkswas first demonstrated by Full error detection and Stibitz of Bell Laboratories in 1940, usinga console correction is provided. The telephone line can be at Dartmouth to communicate over telegraph lines severed and reattached hours later without loss of with a relay-computer in New York.3-45 A second data....Error detection is accomplished bya example is to be found in several demonstrations horizontal and vertical parity bit scheme similar of SEAC results as transmitted toremote teletype to that employed on magnetic tape. Error correction printers in Washington during early 1950. Another is accomplished by retransmittinga message until example of NBS computer design and application it is correctly received and a verification is correctly dates back to 1954. At that time,a brief experi- returned to the sender." (Lynch, 1966,p. 119). mental demonstration of botha multiprocessor Certain current difficultiesare to be noted, how- network (through the interconnection of SEAC and ever. Special problems may arise in multipath DYSEAC) and also ofprogram time-sharing was communication systems andsystems where many run.3-46 Overseas connections tocomputers via users communicate simultaneously witha shared radio teletype links became feasible withinthe processor or data bank. In the lattercase, for next 10 years.3-47 example, "high-speed digital datatransmission is Of increasing interest in thisarea, today, are a vital requirement in many data processingsys- multiple access systems involving on-linetransmis- tems. The reasons vary widely and include (1) the sions to a central data processingsystem from need for rapid processing of remotely gathereddata, remote locations. For example, the GE Datanet-70 as in a satellite tracking and control system, (2) the system involves a multiline controller "whichcan desirability of load sharingamong remote processors handle simultaneously theaccess of up to 248 or files in large management information systems." remote stations into the central computer via Tele- (Andrews, 1965,p. 226). This example loops back type, voice grade and TELPACA communications to the need for improved communication techniques facilities." (Commun. ACM 8, No. 5, 343 (May generally. For example,a question of continuing 1965).) Although, in general, computer-communica- R & D concern is: whatcan be done to increase the 18 rate of graphic transmission on voice communication transmitted speech, detect breaksor pauses and fill lines by several times its present rate? 3.56 gaps in the telephone circuit usage with coded data, Questions less likely to lead to breakthroughs in IDAST (Interpolated Data And Speech Trans- the near future are exemplified by Baker's complaint mission). (Data and Control 3, No. 8, 11. 1965). that "even though the slowest and noisiest loop in Delayed transmission is thusone possible solution the system is the man-to-computer communications to some of the problems of speed and economy with link, tittle research is being done to alleviate the respect to communication links for input and output problem" (Baker, 1965, p. 430), and by considera- transmission. tion of the fact that "the problem of high-speed Other advantages of using store and forward communication of data between remote centers in techniques are that they can be used in off hours,3-59a an information systems network is largely unex- they enable batching until there is enough data plored...." (Scientific Information Notes 6, 4, 11 accumulated to meet minimum transmission time (1964).)Nevertheless, increasing importance of period requirements,3-6° and they allow botha computer-communication complexes, and remote permanent record for audit purposes and retrans- multiprocessor networks can be foreseen.3-57 mission in the case of garbled transmissions 3.6' Mills (1965) points to some of the difficulties of Gains from delayed transmission. usinga "burst" the use of presently available communication links technique, are exemplified in Esso's transoceanic withmultiple-accesssystems,especiallythose message system where up to 2,500 words per minute, involving interchanges of graphic data displays. He in contrast to 100 wpm in leased teleprinter circuit suggests that: "The transmission requirements for systems, are transmitted 3.62 multiple-access systems probably cannot be met Then there is: "A trick devised by remote users by straightforward extensions of the services avail- [which] is the method of linking to a dummy station able through the present common-carrier systems", located in the system. Soon after being admitted but that: "The message-store-and-forward concept, into the system, they begin the production type run together with techniques of multiplexing (i.e., time- of the dummy station, and then quit the remote sharing communication channels, offer the most station. When the remote user feels that his produc- promise." (pp. 237, 241). tion jobs are completed, he can be readmitted into the time-sharing system, link to the dummy station, 3.3. Store and Forward Techniques and retrieve his output and quit. Many hours of line Delayed Response charges have been saved with this method. There are undoubtedly many other methods that can be Delayed-recording outputs may thus hold the devised, and these should be investigated and promise of greater efficiency in priority scheduling encouraged when savings in time and money will and system utilization for multiple-access informa- result." (Fiala, 1966, p. 164). tion processing facilities as well as at least the possibility of lower cost data transmission and facsimile 3.4. Technical Problem Areas communication links.3.5" Relatively earlyrefer- ences in the literature include the discussion by The science, technology, and theory of communi- Genetta et al. (1960), of the RCA Autodata sys- cation systems represent areas as broad and as tem.3-59 Helman et a3., 1963, describe the ITT 525 diversifiedas those that we attempt to cover equipment, a store and forward switching center generally in this report. Examples of some of the capable of interfacing with both synchronous data technicalproblemareasrelatedtocomputer- and start-stop teletype lines. "Dalcode" (Data Line communication networks may be mentioned briefly, Combiner and Demultiplexer) is a Western Union however. development.3.58a Advantages of store and forward Directly related to input/output transmission via techniques include protection against message loss various communication links are questions of data or interruptions of service.3.58b and message compression. Problems of high-cost In a 1962 paper, Maass describes the possibility bandwidth requirements in communication links of using a special type of TV set than can turn itself for both pictorial and voice transmission processes on at a predetermined time and record information thus continue under R & D attack.3.62a Vocoding received onto tape that can be replayed at the techniques may be applied, for example, to achieve client's convenience. This possibility suggests that efficient compression of speech while maintaining transmissions can be made over the regular net- adequate quality at the receiver station.3-63 works but in off -hour time and hence perhaps at Relatively recent R & D work in this area is lower cost. This investigator adds: "When you also exemplified by compression of words or seg- consider the advantages of product or process ments of speech into short time-slots,3-64 by square- demonstrations in color, the possible applications rooting of analytic signals of speech3.64a and by for this specialized communication form [off-hour studies of the losses of intelligibility of digitized TV recorded for replay] develop rapidly." (Maass, speech transmissions.3.65 It is claimed by some that 1962, p. 47).3.59 redundancy reduction, as for example by the use of As another example, investigators at Standard channelvocodertechniques,maypotentially TelecommunicationsLaboratories,Ltd.,have reduce the bandwidth requirements for digitized developed a magnetic tape storage system to delay speech from 48,000 or more bits per second to 2,400

19 bps, or less.3-65a It is also to be noted that speech pictures, using a conditional replenishment technique. That is, after a first full frame has been sent .compression techniques resulting in speeded-up speech playback are of value to certain special (with eight bits for amplitude at each sampled types of information usersnotably tostudents and picture point), thereafter only significant changes others seeking a rapid review of material and to need to be transmitted. Results of subsequent scans blind persons using recorded books.3-66 are routed through a buffer oftwo-frame capacity For textual message data compression, R & D and replenishment occurs in accordance with req...irements may be exemplified by an ingenious significant-change thresholds that vary with the scheme reported by White (1967), in which com- relative amount of motion in successive frames. 3-wg monly used words or phrases are economically Also related to the communications link aspects of encoded by dictionary lookup and short-code sub- generalized information processing systems are stitution, with a resultant reduction of transmitted questions of error detection and error correction in English-language message length of 50 percent or messages transmitted viasuch links. Advanced better.3.66a technological investigations may also be directed, In the picture-data transmission area, Huang and for example, to minimization of noise in TV trans- Tretiak (1965) report experiments of their own and missions, by introduction of "noise only" and of Schreiber and others on the reduction ofredun- "picture plus noise" redundancy such as has been dancy inpictorial data transmissions.3.67They considered by Poppelbaum and Faiman (1968). 3-7" have used computer-simulation techniques, coupled Redundancy to improve correct message recep- to CRT displays, to evaluate thedegradation of tion can be achieved by various coding schemes picture quality when various redundancy-reduction which can be subsequently processed not only to schemes are tried. These 'investigators have also detect the occurrence of errors but in some cases to considered problems of coding of digital color restore garbled messages. To asignificant extent, pictures and of motion pictures.3.68 however, such error-detecting and error-correcting Earlier investigations in this area by Cherry et al. codes increase the complexity of receiver-decoding (1963) and by Kubba (1963) stressed run-length equipment and decrease the efficiency of use of the 3.71 coding and variable velocity scanning.3.69Julesz communication channels themselves, or both. (1959) proposed the use of edge detection techniques Franco et al. (1965) provide a review of relative ad- as a means for codingtelevision signals and Gabor vantages and disadvantages and currentdevelop- and Hill (1961) investigated television band comments with respect to this problem.3.72 pression by means of contour interpolations. Inose The question of the extent of error control coding and Yasada (1963) developed an experimental to be applied will involve clientele requirements3.72a encoder for compressed digitalization of video including tolerances for delays, the desired data signals. 3.7° Another approach to compressed storage rates and expected error rates, factors ofchannel of pictorial data is that developed by Pfaltz and efficiency,characteristicsofthetransmission Rosenfeld(1967),consisting of describing an medium,3-73 and special factors such as the protec- arbitrary region as the union of "maximal neighbor- tion of privileged communications 3.74 In the latter hoods" in terms of the centers and radii of these case, special R & D requirements arenoted by neighborhoods. Geddes et al. (1963) with respect to facsimile trans- A pseudo-random scanning techniqUe developed mission.3-75 by Roberts (1962) preserves relatively good repro- Finally, we note in this, as in many other areas ductibility at the receiver, with a reduced number of of information processing system design R & D gray-levels to be transmitted.""In addition, such requirements, the incidence and effects of human transmissions lend themselvesto some noise- errors. Franco et al., point out that "in settingthe elimination techniques""and to encrypting.3-70e1 error rate requirement for teleprinter transmissions, Predictive coding, where redundancy is reduced by it is necessary to keep in mind that errors can be subtracting from a message those parts that can be introduced by the operators and terminal equip- predicted from previously transmitted and received ment. It is not profitable to specify undetected bit signals (as discussed by Graham in 1958,3-"c2 has error rates which are more than 10 times greater been investigated, by Wholey (1961) for weather than the error rate of the information source, maps,3.70d by Atal and Schroeder (1968) for speech simply because the improvement will not be dis- waveforms3"70eand by Lucky (1968) with respect.to cernible in the output error rate. Experience has error control.3.70f shown that operator-caused character error rates Developments at Bell Laboratories involve an 8:1 may be greater than 1 X 10-3." (Franco et al., 1965, .reduction in data transmission requirements for TV p. 130).

20 4. Audio and Graphic Inputs If the process of item input and sensing is to be Balkovic (1967), Conway and Morgan (1967), Dale speeded, improved, or made less costly by increased (1966), and Davenport (1965).4-' Examples involving use of machine techniques, then expanded and more audio inputs as developed in remote sensing and intensive research and development efforts in audio data collection techniques have been noted previ- signal input and graphic sensing, including pictorial ously (Sec. 2). and three-dimensional data input, are required. Hogan (1966) providesa survey of voice input/out- In Figure 4, we see the operations of information put developments, concluding that the telephone sensing and input and of preprocessing of inputs will become an increasingly important computer together with some of the areas of specific research I/O device. A review conducted by Business Forms and development concern. First, inputs in the form International, Inc., includes the prediction that of auditory signals of various types will be required. "when the voice input is fully developed it will inaugurate the beginning of living in an area of 4.1. Audio Signal Inputs complete convenience. Not only will the voice method of input be used by business firms for busi- Auditory signal inputs, to date, have largely been ness transactions, it will be used for many operations limited to touch-tone telephone techniques and to not directly linked with business." (Business Forms the use of telephone lines to carry meta-information International, Inc., 1967, p. 4.) about other input items. Examples of references in Mills (1967), Van Dam and Michener (1967), and the literature to touch-tone and Data-Phone systems Lee (1968) are somewhat more pessimistic, however, used fordirectinput and processing include particularly with respect to the time-scale in which

AREAS OF CONCERN:

(Audioand graphic input Pattern recognition 2. Information Sensing Speech recognition and Input i Text processing (Feedback Executive and monitor programs control of thre shold 4+ Remote inquiry stations settings)

3. Pre-processing J Operations

C. Criterial feature extraction Image enhancement Information enhancement Error detection Error correction Graphical manipulation languages \--- FIGURE 4.Areas of cm.cern; information sensing, input, and preprocessing.

21 speech and speaker recognition experiments might information derived from special symbols suchas be reduced to practice.42 A simple special-purpose those used in mathematical equations and in chem- exception is noted, that of the experimental Cogi- ical and other types of structural diagr-ms. A tronics system which can ask a certain set ofques- number of relativelynew devices permit on-line tions and can understand the answers "yes"or drafting and direct entry of two-dimensional infor- "no".4.3 Beyond this, moreover, "when the tech- mation into the data processing systems. Insome nique reaches a stage where a spoken wordor cases, the graphic input may consist of handprinted letter can be translated to an impulse that uniquely or cursively handwritten characters, words, or types out the required work, what will be the signatures. Automatic recognition of such input, efficiency, the freedom from errors, the stress im- typically for a small, known "family" of users,' -'a2 posed on the 'dictator' and what training will be has been successfully demonstrated andsome ex- necessary? Research is needed on all theseques- amples are given in Section 5.4 of thisreport tions." (Van Geffen, 1967, p. e). Finally, inthis area, we consider some of the An interesting experiment involving acoustic prospects for pictorial and three-dimensional data signal inputs isdescribed by Neilsen (1965).4-4 processing. The area of application is that of radioactivitymeas- urements instrumentation where thehardware 4.2.1. Alphanumeric Inputs techniques combine use of a scintillation detector and a scaler-counter at a remote laboratory to reduce First, graphic input requirements for generalized the number of detector pulses to be fed to computer information processing systems may involve the via a telephone data link and with simple codes and dire& entry of hand-printed or hand-written infor- timing indicia being transmitted via the telephone mation and the automatic extraction of words from dial. text. Ent, of hand-drawn material may be by stylus In the area of sonic wave measurements,we note with x and y coordinates fed to the computer,or first that "in 1963, Bogert, Healy, and Tukey dis- by light pen movements on the face of a console cussed an approach to the detection and timing of screen. For example, light pen inputs have been echoes. Later Noll successfully applied these tech- used since the early 1950's, notably with the M.I.T. niques [nonlinear filtering of convoluted signals]to Whirlwind computer and in certain militarysys- the detection of pitch in speech waveforms." (Re- tems.4.4b Special equipment suchas an elastic search Laboratory for Electronics, M.I.T., Quarterly diaphram switch device 4'may be used. Another Progress Report No. 80, 168, 1966).-Another example handwritten character input example is the Data is provided by Douglas Aircraft studies of target Trend "MIMO" Hand-Print Model input-output identification in terms of three-dimensional echo terminal for use in real-time on-line operation with reflectionsof acousticenergy.(Dreher,1966). a data processor, where the user writes his message Similarly, in Sweden, Dr. Wennerbarig of the Insti- on a platen with a special stylus. Teletype-to- tuteforAcousticsisconcerned with general computer transmission is to be used." problems of marine acoustics, especially considering Automatic extraction of words from textwas problems of methods of detection, decision-making, apparently first proposedas a physical process and the elimination or reduction of noise emanat- guided by manual selection in terms ofa "light ing from sources other than those of primary pencil" proposed by RCA and discussedat a con- interest -4a4 ference on Mechanized* documentation held under The long-range possibilities for input-output in the auspices of the Welch Medical Libraiy of Johns terms of natural or artificial spoken language will Hopkins University in 1953. If the proposed "pencil' involve close relationships between research in were to be moved across a page of text, words or speech compression, speech synthesis and auto-. phrases of interest could be pickedup and recorded matic speech recognition (see Sec. 5.5). In general, so that, in effect, an underlining of words or phrases the more sophisticated types of auditory signal could result automatically in machine-useful input input, like many graphic input developments, will recordings. (A somewhat OrweTlian prediction has involve to a greater or lesser degree advanced been made that in the future automatic wordextrac- preprocessing and pattern recognition capabilities, tions from voice signals could result in further to be considered in Section 5.7 of this report. Thus, invasions of privacy.) 4.7 "the problem in speech recognitionis one of More modern versions of the word-from-text identifying, within the total acoustic output of the extraction concept are represented by the IBM human voice _or its mechanical reproduction, those "Scanistor",4.8 the Institute for Scientific Informa- elements which are significant for communication." tion "Copywriter",4.8a and by the Eccetron device (Garvin, 1963, p. 111). reported by the National Science Foundationas being in use at the Bureau of Geological and Mining 4.2. Graphic Inputs Research in Paris. (Current Research and Develop- ment, No. 13, 1964, pp. 68-69). A future possibility In the area of graphic inputs to generalized infor- has been suggested by recent advances in the tech- mation processing systems, the questions of entry nology of coherent optical systems.4.8b In general, of alphanumeric informationare first to be noted. however, automatic word extractionoperations Next, there are problems of machine processing of require given texts to be in machine-usable form. 22 1

Among the methods available for conversion of 4.2.2. Special Symbol and Diagramatic Inputs such textual material to machine formare the following: Wigington (1966) defines "graphical input-output as those forms of I/O in which are employed graphs, (1) Initialrecording of thematerial in dual pictures, charts and other forms of information language form: that is, ina form legible to where the two dimensional positional relationships the human eye and simultaneously ina form of information elements have importance beyond readable by machine, such as the hardcopy that of the ordinary formats of printedtexts and typescript and the punched paper records lists." (p. 86). It is noted further that "the acquisition produced by a tape typewriter.4-8c of data directly from graphicalsources such as (2) Manual re-transcription by various keyboard- maps (for terrain, land-use data, etc.) is a developing ing operations such as a keypunch operation capability which is proving its worth to the planning to prepare punched cards, the use of punched and design of transport facilities." (Breuning and paper tape typewriters, or the use of steno- Harvey, 1967, p. 262). typing techniques.4.9 (3) Scanning of .the printed, typed, or handwritten Graphic information of a highly stylizedor sche- record by optical character reading devices matic nature, such as electrical circuitor chemical and automatic re-transcription into the de- structure diagrams, should eventually be amenable sired computer-usable form. topatternrecognition techniques that can be (4) Various combinations of any of thes methods, demonstrated, and at least the beginnings have been as in the re-typing of original documents in a made on problems of pictorial data recognition. standardized font readable by available char- Diagrammatic representation of chemicalstruc- acter recognition equipment 4.9x1 tures, circuit diagrams, word-association "maps" and other two-dimensional complexes of information The state -of -the -art in optical character recogni- are of obvious importance in man-machine inter- tion, both practical and experimental, is promising, action operations. Burger points out, for example, but many challenges still remain (see Sec. 5.3 of that the early work "of Op ler and Baird in drawing this report). Current success in terms of practical chemical structure diagrams on a computercon- applications is largely limited to thosecases where trolled CRT gave much promise for the future there is a high degree of controlover character applicability of such devices." (1964,p. 2.) As of input quality, where the character sets to berecog- 1967, however, the input of special symbols and nized are. limited (and often consist of specially diagrams depend almost exclusively on the encoding designed character fonts),49-a2 and where the alter- and digitalization of the two-dimensional informa- native of key-stroking the input material isexces- tion, either by keyboard, light pen, or scanner sively costly in terms of availablemanpower and techniques. time.49b The latter factor is ironically emphasized Keyboard devices for such purposes have been in those situations where it has been determined explored as in the Army Chemical Typewriter sys- to be desirable to keystroke the source data informa- tem,4-13 studies at Monsanto 4'14 and at Redstone tion using a typewriter with a type font amenable Arsena1,4-15 and in a system described by Cossum to character reading equipment requirements and et al. (1964). In the latter case, an optical scanner then to "read" such typed re-transcriptions by looks for black areas, applies a feature detection machine, logic for symbol and character recognition, and constructs and checks the connection table. Both This wouldappear, in principle, to be a backward step in the light of modern technology. Yet it is in this system and in the earlier Monsanto experi- faster than most alternative keystroking methods ments, a degree of error detection is provided, as currently available and the combined economics especially to check whether or not there is a well- may therefore be more favorable at the present formed structure with every element connected to time.4io at least one other and with all connections recorded Hopefully, such expedients will be of in both directions.4-16 temporary duration, pending the advent of more versatile character-readers and/or improvedmeans Amann and Klerer (1966) reporton a modified for initial generation of source data in machine- Flexowriter device for two-dimensional input of useful form. mathematical equations where the machine system will subsequently recognizeeven badly formed Where the documentary input contains inter- symbols4-17 and Klerer and May (1965) describe spersed graphic material as well as text, additional the two-dimensional programming and other details problems arise:i." For example, itmay be required associated with this device.41" Another system for that manual pre-editing operations mark offareas on-line input and display of mathematical symbols of the page which the character reading device is and equations is AMTRAN (Automatic Mathemati- to ignore.4.12 On the other hand, in the future we cal Translation), described by Reinfelds et al. (1966) may find input scanning systems capable both of and Clem (1966).4.17b A special terminal for key- recognizing printed or typed characters and symbols board entry of mathematical symbols has been and alsoof processing complex pictorial ma- developedbytheOceanographicLaboratory, teria1.4.12a Columbia University.4.170 23 Hoffman (1965) discusses the use of the Army without the use of computer facilities. A Japanese Chemical Typewriter in the IDEEA (Information example is provided by Okajima, et al., 1963.4.24 and Data Exchange Experimental Activities) net- Another illustration of progress in this application work and points out that, by interfacing the ACT area is provided by Whipple etal. (1965) who dis- with a computer, error checks and preprocessingcusses an electrocardiographanalysis program such as data classification can be performed on relating also to special communication networks, input of both alphanumeric data and symbol repre- that is, the use of frequency-modulated transmission sentations:" Tate (1967) includes in a review ofof the data from the Massachusetts Memorial developments in the handling of chemical com- Hospitals to a GE225 computer and teletype display pound information referencestomodified key of the resulting diagnosis.425 Stark et al. (1965) also punchdevicestorecordstructuraldiagrams report on remote access computer diagnosis of directly (Waldo and de Backer, 1959, and Horowitz EKG data. and Crane, 1961), modifications to tape generating Man-machine interactiveinput of two-dimen- typewriters, and the INVACTT-200 structure sional data typically involves the use of such devices typewriter that recordsdirectly onto magnetic as the RAND Tablet4.26 and commercially available tape° Bozman (1967) describes information con- equivalent devices including Data Equipment Com- trol codes for encoding and decoding the use of pany's Grafacon 1010 4.27 Itek's EDM (Electronic "Typits" in the special typewriter designed by Drafting Machine) 4.28Sylvania's Data Tablet 4'2° Duncan and Weissberg at NBS, with applications or an IBM handprinting input device.4.29a At the for chemical diagram input 4.2° Later NBS develop- IIT Research Institute, Chicago, the "Electrosketch" ments are incorported in the "Document Image is similar in function to the RAND Tablet, but it Code System",4'26a as shown in Figures 5 and 6. consists of "a small plastic knob or stylus which the Scanner input examples include investigations user grasps and is free to move about the working by Meyer (1965), Led ley (1963), and again Cossum surface." (Cameron et al., 1967, p. 350). Simek and et al. (1964).4-21 Many of these developments will Tunis (1967) describe a development also similar to be discussed in more detail in another report in the RAND Tablet, but using conventional paper and this series, in connection with character set require- pencil:1.29° ments for both input and output. There are ques- It is claimed that the III (Information Interna- tions, however, as to the development of effective tional, Inc.) Computer Eye is a sensor and processor means for- the regeneration in two-dimensional capable of selectively measuring, inputting, and form for display of chemical structure information processing a variety of graphic images. Suggested represented in linear ciphers, connection tables, applications of the III Computer Eye include micro- and other special notations. There is also the ques- scope measurement and classificationof biological tion of the further operational development of tech- samples,digitizationof drawings,maps, and niques whereby the chemist enters the two-dimen- photographs, and measurement of the texture granu- sionalstructuralinformationdirectlyintothe larity and color of various materials.430 The IBM machine.4.22 Cartographic Scanner is used by Canada's Geo- graphic Information System to develop maps of 4.2.3. Graphic Inputs and On-Line Drafting land use .4.3°a Still other examples of techniques available- for With respect to two-dimensional graphic data on-line drafting and graphic input obviously include inputs we note generally, of course, that analog-to- the SKETCHPAD System at M.I.T.430a2 and the digital conversion and hybrid analog-digital tech- "Light Handle" at the Aiken Computation Labora- niques are required in many applications such as tory, Harvard University,4.30b as well as systems the use of the sound spectrograph 4.22a and the developed by SDC 4:30cControl Data Corpora- automatic processing of electrocardiographic and tion,4.31 RCA,4.32 General Motors Research Lab- electroencephalic recordings. In this connection oratories,4.32aand Computer Sciences Corpora- we note also that computer analysis of encoded tion.4.33 At the National Bureau of Standards, the electrocardiographic and electroencephalographic MAGIC System (Machine for Automatic Graphics data has been the subject of a number of investiga- Interface to a Computer) is used for a variety of tions over the past several years. By 1965, Bobrow man-machine communication and display experi- et al. (pp. 125-126) could report 93 and 94 percent ments. (Rippy et al., 1965). accuracies of electrocardiographic amplitude values Then we note that the "Eyeball" device developed as measured by computer and could conclude that: at the Lawrence Radiation Laboratory provides for "The computer actually provides a more precise the digitizing of pictorial data for regeneration on standard than does visual measurement. An auto- computer output interspersed with text. (Conn, mated system can accurately identify and measure 1967), and that, in a Bolt, Beranek and Newman the waveforms of an electrocardiogram." 4'23 system, "the visual pattern to be analyzed by the Klingemann and Pipberger (1967) emphasize that system is drawn on the face of the computer CRT automatic techniques of EKG analysis are needed display scope by means of a light pen." (Marill because a large number of measurements and more et al., 1963, p. 27). Pictorial data input is described efficient classification methods are, not practical by Ledley as follows: "...The computer must

24 a

1 r r-1 1. SIMPLECAPS 3 4 5 6 111.: 1-.311: 1

7 I I 8 IrL011 9 I a 10 4 11 12. x 13 14 17a 16 17 18 19 20 21 ..H...1 El 22 23 V\,\/k 24...... (... 27 \ 7/. 26 271234,67891011121314122 FIGURE 5.Model typescript mg.,: zt expanded cc,,ln.

V2 3 4 C H 141 6 2 2 . . 7 2 a 9 10 11 00 12 13 14 15 16 1/ 18 .110 19 I 20 . . 21 22 23 Litx?+a 24 23 26 27 U 28 2 3 4 7 6 7 8 9 10 11 1213 14 1516 17 1819202122 X FIGURE 6.Two illustrative display formulas.

25 then interpret the symbols and structural relation- He suggests (p. 185) further that: "After a list ships of the drawing, in preparation for the final of three-dimensional objects has been obtained in coding task. The picture scanning is accomplished some manner, it should be possible todisplay them by a device called FIDAC (Film Input to Digital from any point of view. The sections of objects Automatic Computer). Pattern recognition tech- behind other objects should not be seen, nor should niques are used for the structure and relationship the back lines and construction lines of individual analysis." (Led ley, 1963, p. 201). objects. The three-dimensional display program will do all this and more.".4-39 Roberts also briefly describes earlier work by 4.2.4. Pictorial and Three-Dimensional Data Input Julesz4:39 on machine depth perception applied to Other areas of graphic input of current R & D binocular images. This is a procedure which shifts concern are for the pickup andprocessing of dia- the two images to find the areas at different depths grams, drawings, andphotographs 4.34 Rushforth using texture, rather than detected edges, to develop (1965) provides one example, among many, of this information. pattern analysis, identification andclassification by An approach to two-dimensional scanning, data computer as applied to aerial photographs. Ber- analysis, and three-dimensional reconstruction is tram provides an example of mixed inputrequire- being investigated by Kirsch at the National Bureau milts for such applications, that is, toprovide such of Standards (Lipkin et al., 1966)4-39a Here there is pertinent camera data as the position and altitude a multi-level gray -scale scan applied to a seriesof for each stereoscopic pair of photographs as taken, microphotographs of successive cell tissue slices the focal length, and the distortion characteristics which by their transparent nature give defocusei of theleni.4.35 A scanner developed at NBS, images that are out of their focal plane. A computer described by Moore et al. (1964) digitizes 9-inch differentiation program then enables the determina- aerial photographs for computer analysis of stereo- tion of proper plane for the image with respect to graphic differences 4.36 At Aeronutronics, a PDP-7 its original locus in the tissue. system will be used for film-scanning experiments In 1966, Roberts described the Lincoln WAND, designed to test techniques for automatic detection extending tablet principles to three-dimensional and analysis of data contained in aerial photo- input by the use of a hand-held receiver accepting graphs.4.36a1 signals from four ultrasonic transmitters,4-4° and Stereoscopic initial data recording and three- Denil reported on an IBM development of a language dimensionaldepthreconstructionby machine and three-dimensional light pen input technique processing and computational techniques isof for computer-aided design applications. In 1964, obvious significance in the automatic or semi- McDonough suggested placing an object between automatic analysis in such areas of application as two grids and using light beam reflectance measure- aerial photograph interpretation4.36a2target identi- ments to determine the thickness of the object at fication,4.36b map or contour matchin0.36C and various points.4.41 A similar technique had been analysis of bubble-chamber data with respect to proposed even earlier by Russian scientists. (Gar- thedetectionofhigh-energy-particlecollision mash, 1961). events.4-36d Then we note that three-dimensional x-ray photo- As noted previously, McGee and Petersen report graphs may be used to measure automatically the on a computer-controlledCRT scanner used to distances between components in a potted module measure automatically bubblechamber tracks as and in medical applications to measure the depth recorded on film but applicable also to pulse height of a bone infection 4.42 There are also important analyzers, telemetry converters, and other data three-dimensional aspects in machine-aided design collectionsystems.AlsoattheUniversityof processes, especially in the aircraftand automotive Illinois,arelativelyearlycomputer program, industries, for which far more progress is promised FILTER, was designed for recognition, separation in the near future. Proximity probes and a light- and measurement of star-type nuclear events.4-37 scanning measuring machine (Ex-Cello-0) have A The use of ILLIAC III as a processor of visual been used to measure length, width and height of information with application to problems of cloud clay models. A proposed laser device would also pattern detection was described at an IFIPCongress operate as a fixed scanner above a movabletable 65 symposium. (McCormick et al., 1966.)4.3" on which the model is mounted.However, Gomolak Research and development interests in the area reports: "Because of its coherent beam and resolu- of three-dimensional data input andmanipulation tion, the laser probe closely approaches measuring are exemplified first bySketchpad III, which is a with a true point....Reported accuracy of this computer program to facilitate the entryinto the new laser machine is 0.0001 inch inall three coordi- M.I.T.processing systemof three-dimensional nates (1965, p. 66)." 4'43 drawings. (Johnson, 1963).4-3" Roberts (1965, p.211) An idealized graphic input-output system, poten- reports that: "Sketchpad IIIextends the graphical tiallycapable of processing "three-dimensional constructionprocesstothreedimensionsby curve and surface perspective projectionsinto two enabling one to draw in any of three views onthe dimensions" is reported by Parker (1965, p. 100). scope, thus constructing a computerrepresentation Certain developments in new technologies, such as of a three-dimensional drawing." those of holography (to be discussed in a later report 26 inthis ,series)offer additionalpossibilitiesfor the Cyclops I experiments at Bolt, Beranek and pictorial and three-dimensional data processing, Newman (Marill et al., 1963),1.45 work by Rosen- storage, and retrieval.4.44 For another example, feld 4.45a and by Kanal and Randall 4.45b an experi- Okaya (1966) considers an automated method for mental program described by Evans (1964),4.46 and three-dimensional data-gathering, in thecase of current research under Minsky at M.I.T.4-47 single-crystal crystallographic diffraction studies. As of today, it is feasible to scan graphic and However, techniques of automatic pattern detec- photographic input and convert the initial two- tion and recognition of two- and three-dimensional dimensional and grayness level information into data are as yet highly experimental and much digital form which in turn may be reproduced as a further research and development must be pursued reasonable facsimile of the original, either for trans- before they can be advanced far enough for practical mission purposes (most spectacularly in the case applications. Most of the techniques thatare being of Surveyor's color photographs of the moon's explored for automatic pattern recognitionare surface), for purposes of digital storage, and for highlydependentuponefficientpreprocessing purposes of further computer processing as in the operations. These operations typically include the case of qualitative pattern recognition, that is, situa- extraction of significant or criterial features; image tions where certain pattern transformations are and information enhancement processes designed to performed either to obtain desired end-products reduce noise in the input pattern, to eliminatere- directly or to facilitate further processing, recogni- dundancies of the input signals, and to sharpen tion, or classification.4-448 In addition, advanced selective areas such as either edges or "skeletons" experimental work includes the automatic analysis or, in another case, the "medial axis transformation" of three-dimensional scenes, even although repre- of a pictorial pattern to be described, identified,or sented in two-dimensional forms. Examples include classified.4-48

5. Preprocessing Operations and Pattern Recognition Closely related to the problems of character, are checked against threshold values i^ accordance shape, and pattern recognition for information with which itis determined that the particular input and sensing are operations upon the input subarea is "black" or is "white" (or, in other cases, patterns that are sensed, which we show as pertinent the determination is in terms of some set of specified to Box 3 of Figure 4 (p. 21). These include image gray-seale levels). In other cases, however, analog improvement operations such as boundary and values are first processed 5'3 because, for example, contrastenhancements;noiseand redundancy of Nadler's argument that black-white quantization reduction as in stroke thinning; extraction of salient of a character 'rivets the noise to the signal',5.3a or criterial features and the development and use and that therefore as much processing as possible of property filtering techniques. should be done on the raw video signal before There are,for many applications, important quantization. requirements for noise and redundancy eliminations, for filtering, and for feature extraction operationsas Boundary and contrast enhancement preprocess- close to the data source as is possible 5.1 That is, "a ing operations are of particular importance in input reduction isnecessary, for otherwise the tasks patterns involving multilevel gray scale.5.3b For one involved merely in obtaining input data in a suitable example,boundary andcontrast enhancement form for processing in concepts becomeover- techniques developed at the National Physical whelming the field of visual pattern recognition is Laboratory, Teddington, England are being applied a vast one in which certain problems have baffled topictorial data processing of fingerprint and the experts for years, and in which breakthroughs chromosome patterns. These preprocessing tech- come slowly." (Strom, 1965, p. 110). Then there are niques are based on Fourier transforms resulting processes designed to reduce or eliminate back- from apertures placed in the plane of an imaging ground noise, as by the use of infrared or ultraviolet lens to suppress selected spatial frequencycom- scanning techniques that will sense only the printed ponents. The basic pattern to be processed is in the characters to be recognized on paper forms that form, of a photo transparency, illuminated by a laser may have been overstampedor subjectedto or o,her monochromatic light source whose col- extraneoussuperimposednoisesuchasdirt, limated output is brought to a focus by a lens, which coffee stains and the like.5.2 re-images the input, via various Fourier plane and band limiting stops, to a processing-image plane 5.1. Preprocessing Operations, Image and such as a TV camera, whose video signals may then Information Enhancement be fed to pattern-processing or recognition system. Demonstration photographs for output results for Contrast enhancement in optical character recog- chromosomephotographsshow bothoutlining nition developments is often achieved by a quantiza- effects (elimination of areas of relatively constant tion process in which subareas of the input pattern gray-scale density by low-frequency-component sup- 27 366-107 0 - 70 - 3 pression) and reduction of noise and fine granular xx structure (by suppression ofhigh-frequency corn- xxx xx xxxxx XXXXXXX nonents).5-4 Similarly, with respect to a multispec- X X XXXXXXXX X X ?Pl tral imagery program at the University of Michigan, XXx X x xxx x x a primary goal "is to seeif the information content X XXXAX XX xjt xx x x xxxx be x xx §x in the multiple channels can in some way x xxxx§ xx x xxx processed and combined electrically to produce x x xx xxx xya xxx xxx xxx xx xxx xxx xxx animagewith enhanced target-to-background xx x x xxxxx §§§ 4 contrast." (Parker and Wolff, 1965, p. 24). lit xxx xxxxx3C Preprocessing operations of various types such as reduction of solid black areas to their outlines, were developed for pictorial data processing applications ,,riginal image k Scanned image Character after reduction at the National Bureau of Standards inthe mid-50's. converted into black- of noise (Kirsch et al., 1957). Extensions and further devel- and-white picture opment of these techniques have beencarried out by Moore (1964)withparticular referenceto FIGURE 7.Image enhancement for noise reduction. quantitative analyses of photomicrographs with the scanner on-line to the computer.More specifirr.liy, sensed in the lower right-hand vicinity of an input photographic material can be read directly into the image otherwise probably identifiable as an upper- computer, subjected to various logicalmanipulations case "0", there are veryreal questions as to and analyzed mathematically for 20 numerical whethercontrastenhancementandreductive parameters.5.4a Schreiber et al. (1968) indicate, for operationsshould be performed(inthe case another example, that pictures, including printed that the black area is a flyspeck or other accidental material, can be reconstructed from efficiently noise)or whether, on the contrary,gray-level- encoded contour information.5-4b threshold settings should be relaxed (in the case in a 1966 survey, Hart points to two major types that the black area results from broken connectivity of preprocessing operations for information en- for a defectively printed "tail" of uppercase "Q"). hancement as "dimensionality reduction" (where RabiLow points out, moreovor, that "an '0' and the pre-processors are designed to reduce the a 'Q' may only differ by5 percent in area, but computational load on the processor in its decision- the difference inelectrical output between the making by taking a large number of measurements correlatio_i circuits can be made to be 30 percent, and reducing them to a smaller number) and or more." (Rabinow, 1966, p.23). "feature extraction" (where the purpose is to make Similarly,inspeechanalysisand synthesis only those measurements that are likely to provide research, Liljencrantz (1965) reports of a filter a great deal of informationabout the pattern). bankspeechspectrum analyzer developedat In this survey, Hart covers slightly more than 50 the Royal Instituteof Technology, Stockholm, papers and reports concerned withthe theory that the band-pass filters in the bank "can be andapplicationofpreprocessingtechniques set in a number of different combinationsof center in pattern recognition.5-5 frequencies and bandwidths. In this way itis Preprocessing operations designed to enhance possible tr) make a gross analysis for orientation the information content of an input pattern typically as well as detailed analysis with ahigh resolution involveeitherreductivetransformationsand in either frequency or time." (p. 1). normalization, or both. Thus, in optical character Theappropriateinteractionbetweennoise recognition, stroke width normalization may be elimination and information enhancement operation applied.5-5a One example isa development at should obviously be subject to control by system- International Computers and Tabulators,Ltd., feedback inputs and threshold setting. Another where one, two, or three cell widths in the input obviousexampleisheight-and-width-parameter 4. pattern are processed through logical networks re-scanning in the case of character recognition to provide a three-cell stroke output. Forphoto- techniquesthatareintendedtodiscriminate interpretationapplications,itisreportedthat between upper- and lower-case inputs and between "another recent processing operation, known as ascender anddescenderindicia,whetherfor phototone, consists of the edge enhancement of individual character or for specific character-string an aerial photograph. The enhancedimage is then sequences.5-5b added to the image to produce a more legible Consider the feedback mechanisms involved , photograph; the total process can be considered in a gross scan followed by a morerefined scan as a type of pre-emphasis filter."(Tobler, 1968, pattern upon the detection of certainpre-designated p. 267). Image enhancement for noisereduction indicia. Such indicia may include the "finding" as developed by Gattner andJurk (1963) is illus- of a fiduciary mark and the subsequent alignment trated in Figure 7. of the detail scan to follow a tilted page orline It may occur that the information enhancement (FOSDIC)6 the "first-finding" of a black line and noise reduction processes are interdependent of type or print as the scan comes upfrom the to a high degree. Thus, if a smallblack area is bottom of an envelope in an early Farrington reader 28 for postal addresses s.7 and the adjustable scan recursive feature-orderingwillhavesignificant focus-defocus settings based upon relative height- advantages 5.11b and Chu and Chueh (1967) claim width determinations of an initial character of that if a sufficient number of well-selected features an address -line 5.8 is used in character recognition, the error probability Image enhancement operations may also depend of a Bayes decision function can be made arbi- upon feedback factors with respect to neighboring trarily small. cells of a given cell in a quantized graphic input Among the many techniques for feature extraction matrix. That is, a black cell may be retained as that have been investigated are those of stroke "black" in an improved image according to the direction encoding, edge and corner detection, number of orthagonally adjacent cells that are detection of "lakes" (fully enclosed white areas) similarly black, or it may be erased if a majority and "bays" or "inlets", and the use of other topo- of its immediate neighbors are "white". Thus, logicalpropertiesof input patterns .5.12Schultz such operations may serve both to thicken or (1963), for example, has expanded the number fill in a black line, area, or character stroke and of such features previously included in patented simultaneously to smooth rough contour edges systems of this type and has included the use and to eliminate small areas of apparent noise." of contextual information to separate alphabetic Distance transformations, based upon such local from numeric characters. neighborhood operations, have been explored by At Siemensin West Germany, for another Rosenfeld andhisassociates, among others.* example, the recognition philosophyisthat of Detection of clusters, edge or boundary enhance- the extraction of criterial features from character mentsortheirconverseskeletonizing,and shapes,largely independent of character size, noise elimination or smoothing are all possible location in the input image field, and, to some applicationsof thesepreprocessing techniques extent, orientation (Gattner and Jurk, 1963). Noise forpictorialdata ."°Forexample,Rosenfeld reductionisfirstapplied toeliminate isolated and Pfaltz (1966) describe sequential operations noise and to smooth ragged edges. As early as that label compone.Lts of a given subset of pic- 1958, a computer program involving the appli- torial data and compute distances from every cation of majority logic to a small, sliding sub- picture element to this subset. In terms of the matrix (as in some other local neighborhood pre- distance function, a "skeleton" subset can then processing operations) was developed. The features be defined that serves to minimally determine the (or "form elements") used include determinations original subset. of convergence or divergence of character strokes. Ledley's"Bugsys"proceduresareaspecial 5.2. Property Filtering and Feature case in which it is desired to locate and trace the Extraction features essential to pattern detection for digitalized photographic input data.5-13 Tou and Heydorn (1967) emphasize that the Then, at the M.I.T. Research Laboratory for preprocessing operationswhichreduce dimen- Electronics "a character-recognition procedure for sionality have as the goal the minimization of the identification of letters and numbers in news- entropythat is, to provide transormations from print has been developed. The approach is based many to few dimensions or properties, selecting upon automatic tracing of the outer black-white those most discriminatory for one class as against edge of each letter. After nonlinear smoothing, of all other classes." Similarly, Hughes and Lebo the 'gear backlash' type, the horizontal and vertical state that "the term 'feature' implies a combing *ion position-coordinate waveforms are processed to of two or more measurements which, taken together, detect the sequence of horizontal and vertical describe some pattern aspect relatively invariant changes of direction, together with low-resolution within one class, but with a different characteristicinformation on the spatial locations where such set of values outside that class. A feature is a set changes occur. The result is a 30-bit code word from of two or more measurements taken as a vector which the letter is identified by look-up in a stored measurement pair." (Hughes and Lebo, 1967, p. 19). list, in which there may be several different code Extraction of criterial features, feature detection, words associated with the same letter." ** and determination of class-membership properties Criterial feature techniques that combine various have the objective of isolating and identifying the topological properties and stroke direction clues most salient characteristics of input data in terms developed at a relatively early date include those of of pattern recognition or categorization require- Grimsdale et al. (1959) to determine the number, ments.5-"a However, as Hart remarks, "there size, curvature, relative length and orientation of is no good conceptual framework in which to view constituent segments of character patterns, Sher- attempts at feature extraction, and consequently man (1960) with respect to recognition of hand- a success on one special problem seldom sheds printed characters, and Frishkopf and Harmon light on a solution to a slightly different problem." (Hart, 1966,p.29). Chen (1966) suggests that Quarterly Progress Report, Research Laboratory for Electronics, No. 80, RLE, M.I.T., 218 (1966) (further references are given to: J. K. Clemens, "Optical Character Recognition for Reading Machine Applications", Ph. D. Thesis, Dept. E.E. M.I.T., Sep. 1965 and H. Teager, "Multidimensional Visual Information Processing", Conf. Ad- *A comprehensive study of pictorial data processing by Rosenfeld (1968), was vances Biomedical Computer Applications, N.Y. Academy of Science, N.Y., June unfortunately received too late for discussion in this report. 3-5, 1965).

29 (1960) with respect to constrained cursive hand- in the search for relative invariance for various writing. (See also, Sec. 5.4 of this report.) interesting sets of patterns. These he classifies Current projects under Steinbuch at the Tech- as: local transformations,including local averaging, nische Hochschule, Karlsruhe, further exemplify edging, and recognition of particular local con-. the importance of preprocessing operations in dif- figuration or feature extraction; global or holistic ferent areas of automatic pattern recognition. First, transformations,includingtranslation,rotation, investigations of automatic processing of pictorial expansion, contraction, filling in of hollow figures, data contained in aerial photographs are under location of optical center of gravity, and operations 5.13a The objectives include the location of todetermineconnectivity;and"functionals", streets and the recognition of objects such as or operations that count orencode, such as 'blob' vehicles and the determination of their coordinates. counters, moment of figure with respect to agiven The scanner provides a variety of scanning modes point, slope of line, distance between two points; and varying resolutions, under programmed control, and other transformations such as projections onto including line scans, point scans, rotating a spot of a line or axis, or the mapping of a patternperimeter variable diameter around a point, and the capability along some reference axis. The search for relative of illuminating a line on the screen and rotating it invariance includes investigation of area-preserving so that intersections withsmall, fine lines are transformations, shape-preserving transformations, detected. image enhancement, contour projection, contour- In the detection procedure, the presence of directionsequences, and selectionof criterial parallel lines is first looked for at relatively low feat ures.5-13 b resolution and then within the indicated area vehi- In many of these areas of development, ar, im- cles are looked for with high resolution. The scan- portant R and D objective is to identify those criterial' ning system preprocessing capabilities include an features that are relatively invariant under trans- analog-digital converter to process 64 levels of formations of size, location in the input image area, gray-scale information. Preprocessing operations and orientation. One approach is that of "auto- include evaluation of density gradients, detection correlation".5 "4 The CYCLOPS machine at the of contrast boundaries by detecting phase of signal National Physical Laboratory, Teddington, England, and determination of boundary directions. is an electronic version of the earlier Parks tech- Preprocessing operations of feature extraction nique for detecting criterial areas in character auto- may also provide means for somedegree of data correlations as originally realized in an optical sys- compression for either storage or transmission. tem by Clowes (1962). Analog techniques are used Thus, "the ability to reconstitute patterns from to carry gray-scale information along to the recogni- features suggests a means to transmit a set of tion logic. Extraction of criterial features of this high-resolution photographs over a low-bandwidth type in the case of whole words or multicharacter channel....If the total number of features sequences is exemplified in the work ofClayden for any set of patterns is less than the total number and Parks (1966), also at NPL.5.15 of retinal cells, then the feature-detection technique Other character recognition techniques that have can be a means for achievingbandwidth reduction been experimentally developed, such as that of in an image-transmission system." (Block et al., Alt (1962), provide considerable invariance to size, 1964, p. 83). rotations, or placement in the input image field. For another example, at M.I.T., "the nonlinear Hu (1965) also has proposed the use of moment processing of pictures, for elimination of all features invariantsforrecognition. These moments are other than the principal edges, is being investigated. derived by a summation of products of the density The aim is to simplify pictures to the extent that distribution and the distance from an arbitrary tactile display and perception of the principal edge axis.5-15a Various optical techniques have been patterns becomes practical." It is reported however, exploited for such purposes, including the deriva- that "edge extraction, which involves processes tion of the Laplacian function,51 or Hilbert trans- relatedtospatialdifferentiation of the picture forms,5-15c as well as several variations on the use of intensityfunction,isanontrivialoperation, Fourier transforms, although at a relatively high especially for noisy pictures. Edge extrapolation processing cost.5J5d and smoothing, which are desirable operations From these examples, it is evident that character from the standpoint of noise elimination, place and pattern detection and recognition techniques the problem in the realm where general pattern- are of interest with respect to the current stateof recognitionconsiderationsbecome important." the art in the computer and information sciences (Quarterly Progress Report, Research Laboratory on at least a two-fold basis.First, such techniques for Electronics, No. 80, M.I.T. 219 (1966).1 offer automatic means of sensing, transcribing, and Minsky (1962), considering the case of visual re-recording of data values or textual interpretative pattern recognition with respect to the moregeneral statements as originally marked in a prepared for- area of so-called 'artificial intelligence',has grouped mat for marksensing, visually encoded as by means together both interdependent and independent of a bar or color code, or handprinted, drawn, typed feature extraction possibilities in a listing of the or printed asalphanumeric characters and other operations and transformations that may be useful graphic symbols. Secondly, insofar as these tech- 30 niques incorporate adaptive or self-organizing fea- TheControl Data 915variable-formatpage tures, they are themselves fruitful areas for the reader was designed for use with the recommended development and application of advanced research optical character font (upper case) of the American principles. Standards Association 5'25, but Hustvedt (1967, p. 3) notes that "the CDC reader Rabinow is now building 5.3. Optical Character Recognition for Bank of America will operate with the loosest possible constraints of a controlled operation...it The area of automatic pattern recognition R & D is expected to read the output of all typewriters effort that has had the greatest practical application the bank owns and with very little or no constraint to date is that of optical character recognition on format and other variable factors." A recent (OCR), especially for relatively limited or standard- Japanese entry in the muitifont field may also be ized fonts and character sets. Looking first to the noted: the Toshiba OCR developments are directed techniques for the automatic recognition of printed, to the reading of six or more stylized fonts and typed, or machine-embossed alphanumeric char- handwritten numerics.5.28 acters, considerable progress is to be noted in the In general, however, itisstill reasonable to field of the limited-character-set category. Several assumethat:"Themultiplefontrecognition hundred optical character reading machines of problem is probably the most intriguing challenge this limited type are now in operation, largely for ofcharactersensing. When requirementsfor commercial data processing applications such as large quality tolerance, and cost compatible with insurance premium billing, credit card accounting value, are added to this problem, basic recognition and subscription renewals."8 Some 14 to 15 com- philosophy, as well as the means for implementation, mercial suppliers offer OCR equipment in the United must be re-evaluated." (Greanias, 1962, p. 145). States alone ."8a Nevertheless, multi-font optical character readers However, it is noted that "the scope of appli-are in use by or on order for such U.S. Government cations for character readers is currently limited agencies as the Social Security Administration,5.27 primarily by their inability to read a variety ofthe Army Finance Office, the National Security differentfonts, by their poor performance on Agency, the Defense Intelligence Agency, the Li- handwritten documents, and by the lack of standard- brary of Congress, and the Rome Air Development ization within the industry. Consequently, consider- Center. In addition, a special Philco reader for able development effort is being put into theseZip code recognition is being tested by the U.S. areas, as well as into improvements in reliability Post Office Department 5.28 The versatility of auto- and speed." (Feidelman and Katz, 1967, p. 0210:30). matic character readers is still quite limited, how- The special case of retyping to a specialized font ever, both with respect to the number of fonts that has been mentioned previously (Sec. 4.2.1). can be processed without extensive mrnual handling 5.29 and with respect to format differentiation 5.3.1. Multifont Character Reading for inter-mixed input of varied type items. On the other hand, there are potential break- Thus, a continuing R and D requirement in the throughs promised in the relatively recent opera- area of information sensing and input is the develop- tional installations of variable-format-page-reading ment of more versatile, large vocabulary, multifont andmultifont machines. Devices incorporating page readers 5.3° In a survey of the OCR field re- one or both features include both the pioneering ported in 1966, Feidelman concludes: "The char- andthe more recent Farrington machines,"7 acter recognition field is relatively new, with much Recognition Equipment Inc.'s "Electronic Retina work to be done on improving equipment perform- CharacterReader",5.18andPhilco'smultifont ance and developing character readers at lower page reader.5-19 Huntley (1964, p. 94) cites reject cost. Consequently, it is in an active state of de- ratesfortwooperationalalphanumericpage velopmental flux that can be expected to continue reader installations as 1 in 100,000 and 6 in 100,000. for several years. In the near future we can expect Among the advantages to be expected from multi- to see the multi-font capability in all standard char- font page readers are capabilities for reducing the acter readers. Further away, possibly in five years, amount of proofreading required in the preparation character readers able to read handwriting should of manuscripts and other copy 5.2° and for various be commercially available. By that time, we can decentralizations of input preparation operations.5.21 expect to see the character reader replacing punch Other important developments in multifont page cards as the primary computer input medium." reading are the operational IBM 1975 Optical (p. 52). Page Reader,522theexperimental Link Page For obvious reasons, it would appear likely that Reader developments at the L ..nk Division, General commercial interests in R & D in tip- computer and Precision,Inc.,5.23 and laboratory investigations information sciences are now and Nill continue to at Sylvania.5.24 Earlier investigations of multifont be largely oriented to those areas which promise the possibilitiesat IBM included the development relativelylargestcontinuing markets. Multifont of computer programs for the design of recognition page-reading automatic character recognition tech- logics to be applied to multifont inputs, including niques may therefore be expected to be achieved two Cyrillic fonts (Kamentsky and Liu, 1963). for a wide variety of preformatted schedules and 31 reports used exten.:iveh by business and industry charts, drawings, and photographs. In such cases, or as required by Covernmental practice, yet such a temporary solution may be to provide certain developments are unlikely to be extended to the manual preprocessing 'steps such as the sorting typical page of a scientific or technical journal, in of material by font type or the masking out of English; much less to those printed in Russian, graphic material and of footnotes appearing in Japanese, or Chinese. small point size. Programmed scanning and dim cted re-scanning There are also many unresolved problems in- as preprocessing operations are becoming available volved in the development of suitable rules so that for certain types of character recognition require- machine techniques can be used to distinguish ments, such as the processing of variable formats between title, heading, or page number and the when several different forms are inter-mixed.531 text itself, author's name in a cited reference and Further, "with a stored-program controller, the the title of the reference cited, and so forth. Auto- system can determine three very important things matic recognition of discrete characters at input is during a single reading pass: (1) Whether or not difficult enough for character and font vocabularies there has been a mistake in data preparation, (2) of adequatesizefortext-processing purposes Whether or not there has been an omission in data in any case, but Garfield (1965) points to "fantastic preparation, and (3) Whether or not the machine syntactic problems even if the machine has a uni- has read the data correctly. Also, the system can versal multifont reading capability. For example, edit,accumulate,balance,verifycheck-digits, in the citation, J. Chem. Soc. 1964, 1963, which check parity, and condense data to provide easier number is the year and which the page number? access and reduced storage costs. Exception docu- These are not trivial problems." (p. 189). ments can be marked and sorted during the single reading pass, and details can be printed on a pe- A particularly severe variant of this problem ripheral printer so that corrections can be made arises in the processing of texts involving inter- easily." (Philipson, 1966, p. 128.) mixed languages, alphabets, and meta-textual em- In other potential applications, incoming textual phases. The problem of subclass within a pattern materials may be scanned for the occurrence ofclass is especially acute in character recognition certain key words or phrases which are then used applications involving mixed alphabet input, where for mechanized indexing or for preliminary routing either many-to-one or many-to-many identification or classification. This example is analogous to transcription is required. Thus, it may be required the rough sorting of incoming mail, where a pre- to recognize the character shapes "A", "a", "a", liminary distribution by subject-matter or field-of- and "H" as all identifiable and transliteratable to interest is made by mailroom personnel. In some ASCII code 10000001 (= 'A'), or it may be required operations, machine detection of certain marks to separate the Greek and Cyrillic from the Roman or symbols, other than the text, can be used in versions of the same character as separate identifi- selective retranscription or re-use of data, such cation classes. The problem is aggravated by the as the reading and retranscription of those portions fact that the existence of orthographic identity in of teletype messages that have been bracketed character shapes does not necessarily imply identity by a human editor. of character denotation. Consider that the mix is Cyrillic and Roman. When does the input character Similarly, recognition of special symbols may pattern "H" denote 'H' and when 'N'? or "C"9 be made to result in self-adjustments of the reading indicate 'C' and not 'SI process such that the detection of underlining by the machine may be used to cause the machine The current state of the art in optical character to skip the reading of symbols that are underlined. recognition thus leaves much to be desired other Conversely, the reader equipment may be designed than for routine, massive operations involving large so that it reads only underlined material, such as volumes of simple, format-controlled messages and the specific words in a patent text indicated by records, which typically utilize limited, special- an analyst as those to be converted to appropriate purpose, standardized fonts and character sets. code symbols for subsequent mechanized search. Beyond the questions of multiple fonts and large However, much remains to be done before recog- character repertoires are problems of recognizing nition equipment can successfully "read" a typical noisy and degraded characters. documentary page.5-32 It is to be emphasized that "there is no known How is a machine system to be given the neces- shortcut to optical character reading, stylized fonts sary detailed instructions that will enable itto included. The real problem is reading degraded distinguish in automatic page reading applications material, not different or stylized fonts." 5.33a Vander between the same symbol (such as ".") being Lugt et al., point out further that: "In the design of used as a punctuation mark terminating a sentence, any practical system, one must know how the sys- as a denotator of an abbreviation, or an indicator tem performs when the input characters are noisy. of numeric fractions? 5.33 A printed page may Three kinds of noise are considered: changes in the also contain subscripts and superscripts, special signal size, changes in the signal orientation, and symbols, and graphic material such as mathe- changes in the signal quality." (Vander Lugt et al., matical equations, chemical structure diagrams, 1965, p. 131). 32 5.3.2. Performance Requirements and Quality Control niques are directly related to the objectives that are to be served inthat intended application .5-44a The problem areas of continuing concern there- Similarly, the operational requirements for a char- fore include those of paper handling, quality con- acter recognition device are related tothe charac- troland performance monitoring, dealing with teristics of the situation in which it is to be applied noise and with degraded characters, reject han- and to the standards of reference selected for dling, and error detection and correction tech- performance measurement. The evaluation of a niques.5.34 First, paper handling represents a major given character recognition system is therefore factor in OCR equipment cost and a constraint on directed as much or more to these factors as to the speed of automatic readin..5.35 Quality and the logical and mechanical characteristics of the performance controleffortsinclude the further system or device itself. development of measurements of print character- For example, in such massive paperwork activities istics 5.36 such as ink smudge and bleed,5.37 or of as the keypunchingof individual wage earnings light stability 5.38 and of instrumentation such as data from employer tax returns for social security the Kidder optical character tester5.39or the Optical accounting, a steadily increasing work load may Print Quality Monitor of International Computers rapidly outstrip available manpower. There may and Tabulators, Inc.5-4° be a high rate of turnover among personnel who The need for extensive quality control measures perform the necessary keyboard operations, and in OCR applications is further attested by Hustvedt a high training andreplacement cost. In such as follows: "The day of successful,economical situationsthebasic management objectivein optical character recognition is here, as several looking toward automatic reading devices isto hundred operating installations attest. In all suc- meet present manual outputstandards for an cessful operations, however, control is exercised increased volume of work. That objective would over the preparation of documents tobe scanned." then be of greater significance than increasing (Hustvedt, 1967, p. 2.) the speed or the accuracy or decreasing the cost Controls are also required in the selection of of the transcription operations, although obtaining good quality paper and in minimizing the effects these latter advantages would also be desirable. of carelesshandling.5.4' Paper with extraneous Overall systemrequirementsthat determine material likely to interfere with clear imprinting, what factors will be most significant in evaluating or even with watermarks,should be avoided if automatic character recognition techniques for a possible.5-42 Thesystemplannerconsidering particular application involve the various stages of possible OCR applications must next consider data origination, transmission, receipt, input to the the quality controls available in terms of line and reading-recognition process, output, storage, and characterregistrationtolerancesbothforthe subsequent use.5.44b In any specific case, the per- input items and for each of the equipments under tinency and weight to be assigned to the various consideration 5.43 factors must be determined by thorough fact- In experiments reported by Vitale (1965) the finding and analysis of present and desired condi- objectives are to obtain figure-of-merit measure- tions and procedures. In general,possibilities for ments on sampled characters both todetermine maintaining high quality input, limiting the size of degrees of degradation and to determine what the vocabulary, handling carrier items efficiently, error and reject rates might beexpected if recogni- and meeting realistic reliability requirements, should tion logics were adjusted to given levels of noisy outweigh questions of a position-dependent as printing. Uffelman et al. (1967) have:investigated 5.44 versus shapedependent recognitionlogic, of per- the character- and stroke-dependency of noise missible reject rates, and of speed or rate of recogni- However, it has been pointed out that "unfor- tion. In particular, the accuracy of the recognition tunately, procedures for handling rejected and error system output cannot be expected toexceed the documents are quite necessary atpresent. In accuracy of the data or information asit is initially all fairness, it should be noted that most causes generated and initially recorded. of rejects are due to the source document rather than the OCR device. Nevertheless, one must consider the entire operation with its good and 5.4. Recognition of Handprinted Characters, bad documents against the present operation." Handwriting, Geometric Shapes, and Line (Auerbach Corp., Source Data Automation, 1967, Drawings p. 3-133.) Moreover,while the IBM 1975 multifont reader operatingatSocial Security can read In terms of continuing experimentation there are between 50 and 60 percent of typical inputs, "any a number of R & D requirementswith respect to improvements intheportion read must come many other areas of potentialapplication of optical from improvement in the preparation of reports character recognition techniques. The case of con- byemployers a difficultmatter."(Hustvedt, strained handprinting may be noted first. The most 1967, p. 3). rigorous constraints are those involved in hand- The factors that are critical for the development printing the character within a specified area (such of performance specifications for any given appli- as a preprinted box) but witharbitrary embellish- cation of automatic character recognition tech- ments as in a patent issued to Dickinsonand 33 FOSCRIPT SCANNING FORMAT

FIGURE 8.EOSCRIPT decoding criteria.

Wheeler in 19415.45 or in a FOSDIC developinent.5-46 have been investigated with particular reference It is to be noted that, in the latter case, since the to constrained hand-printed characters, as devel- FOSCRIPT scanning technique scans only for black oped, for example, at IBM (Johnson, 1951) and at in specified areas outside, or inside the guide-line Bell Laboratories (Dimond, 1957).5.47 boxes (as shown in Fig. 8), these need not be in a In developments at Standard Elektric Lorenz, separate color of drop-out ink, as is the case in most Stuttgart, West Germany, Dietrich has also made of the other constrained handprinted or handwritten some investigations of possibilitiesfor recognizing character recognition techniques. constrained handwritten numerals, to be inscribed in boxes with two preprinted dots, in a manner 5.4.1. Constrained Handprinting similar to the IBM and Dimond techniques. Then there is a patent issued to Vernon and Walsh (1962) Typically, constraints involve the use of restrictive which considers the use of a six dot and a 'forbidden guidelines to control variations of size in characters area' constraint for handwritten characters.5.48 asprinted,including,for example, ascender- Somewhat closely related to some of these dot- descender lengths relative to other characters; the constrained techniques, especially those involving requirement that characters be drawn as closely as use of the sequence of stylus movementsfor possible to an exemplar shape (e.g., "4" rather than identification purposes, are several systems for "il", see also Fig. 16, App. A, p. A.116), or arbitrary on-line input of graphic patterns, specifically includ- centering and "crossing" requirements, as about ing handprinted or handwritten characters.548a preprinted dots, and combinations of these con- Groner (1966), for example, reports a relatively trols.5.46a The acceptability of such constraints to simple recognition scheme together with con- typical users must of course be considered.5.46b venient editing features.5.49 Earlier investigations Then there are techniques that involve what might in this area include patents issued to Crane (1964)5.5° be termed critical area or vector "crossings" detec- and Harmon (1964),5.51 and work by Roberts5.52 tion, e.g., that character strokes cross or do not cross and others at M.I.T. ?nd in cooperating organiza- superimposed lines or do or do not occur in certain tions such as Mitre Corporation and Bolt, Beranek, arbitrary areas of an image field. These techniques and Newman. 34 On-line handprinted or handwritten character tion have been mentioned previously (Sec. 4.2). (or word or signature) recognition techniques are Here, we are concerned with the property filtering of increasing importance in many man-machine- and recognition techniques. Pioneering approaches interactive experimental operations. One recent to the recognition of relativelyunconstrained hand- example is the "SHAPE SHIFTER".5-5a Another printed or hand-written characters by Selfridge example is to be found at S. D.C.5-52b (It should (1955, 1956), Dineen (1955), and Clark and Farley be noted, of course, that many of the techniques (1955) stressed the problems of extracting the involving stylus, light-pen, or equivalent on-line significant features from a background of noisy or input are highly dependent upon the sequence irrelevant detail. In simulations of machine recogni- of character or word or signature formation and thus tionof handdrawn characters they specifically are not generally transferable toother handwriting developed various types of filtering of the source recognition applications). pattern in order to extract such features,especially Commercially available equipment in the con- the detection of edges and corners. Later develop- strainedhand-printedareaisrepresented by ments at M.I.T. have included the work ofDoyle, IBM with a sales check application as early as Sherman, Roberts, and Teitelman, among others. 1965 5.53 (and more recently the 1287 reader has In a 1965 survey of automatic recognition of hand- been introduced; 554), and by Recognition Equip- writing, Lindgren covers the work of Earnest,562a ment, Inc.,555 Information International,Incor- Frishkopf and Harmon, and Mermelstein and Eden, porated(III),556Philco and CDC,557 and the Doyle (1959) provided an early example of computer Optical Scanning Corporation 5.58 A capability simulation of a scheme for the recognition of largely for recognition of handprinted numerics has also unconstrained handprinted alphabetic characters, been claimed for the Cognitronics ROCR (Remote using 28 criterial features. The machine simulation Optical CharacterRecognition) development of the technique resulted in computer recognition, previously mentioned. (See p. 17 and note 3'29. which was correct approximately 87 percent of the The REI equipment,inparticular,combines time, of "sloppy" handprinted characters. high-speed multifont typewriter or handprinted The emphases in this program were on parallel alphanumeric recognition at high speed with high processing of the observed against the master speed sorting of input documents. An entry by identification formulas and on derivation of probabil- Toshiba in Japan has been noted previously (see ities of occurrence of the results of criterial feature p. 31 and note 5'26). tests by extensive testing with samples ofthe char- For another example, at Olivetti in Italy, the acters to be recognized. The handpfinted char- basicrecognitionconceptbeinginvestigated acters, whether 'teaching' samples or unknowns involves a multi-topological system for the detec- to be recognized, were constrained by theframe tion of line directions, connections, intersections within which they were printed but otherwise were and relative positions of intersection as applied often badly formed and noisy. Further investigations not only to the ISO stylized A andB fonts but of the frequency with which a given character was to handwritten numerals as well.A distinctive confused with some other showed that many of feature of the Olivetti approach (i.e., in contrast the incorrect recognitions were for character pairs to multifont techniques involvingshifting from (`A', 'H') where the source pattern would be almost one font to another) isclaimed to be the capability equally ambiguous for a human reader. of handling a family of fonts, so that the system Sherman's investigations (1959) were concerned can recognize charactersvarying asto pitch, with the problems of recognition of hand-printed size, stroke widths and edge tolerances. characters, especially the search for character Other present or potential suppliers of OCR invariants where the source patterns may vary in equipment who have considered applications to size,slant, registration, rotation, and the like. constrainedhand-printedcharacters,include Since the use of holistic templates would be im- Sylvania,5-58 Farrington,5.60Bull 5.60a and workon practical because of the enormous number of pos- possibie. Zip Code applications by Burroughs, Control sible combinations of pattern parameters, Sherman Data (Rabinow), and Philco -Ford 5.01 as well as by turned to the field of mathematical topology, with NCR,862 but, in general, it is to be noted that: particular reference to graph theory, for his criteria "Some of the techniques currently being investi- of recognizability. The use of graph theory enables gated in connection with hand written documents the encoding of a given pattern in the form of a con- are curve tracing, detectionof selected features, nection matrix. The rows and columns of this and context recognition ....Although a number connection matrix correspond to the nodes of the of companies are working on the problem, most graph, while its elements correspond to the number of the work has been kept confidential." (Feidelman of connections or line segments between the nodes. and Katz, 1967, p. 0210:31). However, the connection matrix in this form would not discriminate between characters having 90° or 180° ambiguity, orotherwise having topological 5.4.2. Handwritten Characters and Handwriting equivalence (e.g., "S" and "2"). Stylus and light pen input techniques for the In Roberts (1960) experiments models of adaptive direct entry of handprinted or handwritten informa- networks were investigated with respect to recogni- 35 1 tion of the same handprinted characters used in joints, splits, and ends of lines together with their the Sherman tests of the quasi-topological method positions relative to one another.5.65a for character recognition.5-63 Other hand-printed Marill and Green (1960) have used a variation character recognition experiments of interest in- of a vector crossing scheme in a proposed model clude those of Highleyman (1961,1962),5.63a Spinrad of pattern recognition as applied to handwritten (1965),5.63b Uhr and Vossler (1962g.63C and Brill characters.Inthis scheme, involving ineffect (1968).5.63d a polar scan where the vectors dissect the image To date, the most common approach to feature- field at 45° intervals, they measure the distance extraction recognition of relatively unconstrained along each radial vector from the edge of the field handwritten characters has involved some form of until the first character portion crossing is detected. contour-following or of extraction of angles, slopes, Kelly and Singer (1960) have also investigated and directions of junctions and disjunctions ofcon- means for obtaining characteristics of curves, for tour strokes.5'63e For example, at IBM, in a hand- characterrecognitionpurposes,by measuring writing reader system, "the flying spot actually distances from the center of gravity with respect traces the outline of the character it is looking at. to radial vectors. These tracings are converted to voltage wave forms Various operations to reduce noise by local internally. These wave formsare then compared averaging, to standardize line widths, to extract against statistical probability tables that were gene- criterial features, and to ascertain the relative rated in conjunction with Tufts University [from] location and size of such extracted features,are exemplified in the system described by Bomba.5-66 unconstrained handwriting samples....About 100,000 numerals were fed into the system, and This has been tested on handprinted samples reliability and accurate reading of about 98per- of 34 alphanumeric characters, using a scanner cent was developed." (Merz, 1964, p. 84.)5.631 developed by Highleyman and Kamentsky (1959) and simulation on an IBM computer. Kuhl's (1963) approach, involvingan angular Sometimes not only features but relationships mapping transformation derived from contour fol- between the features are required to he detected, lowing, resulted in 95 percent correct recognition as in Sutter's 1960 patent for recognition of hand- of several hundred hand-printed numerics and written numeric characters, such as the following: upper-case block letters.5-64 At the British General Post Office Research Station, computer simulation "The initial stroke of the symbol or numeral investigations of the recognition of handwritten is inclined downward toward th.- right.... alphanumerics have been carried out using a tech- "Within the second zone there is a scanning nique that encodes directions of connections from line on which a second pulse occurs, this pulse a matrix cell for quantized, thinned character stroke being later in time than the first pulse or, in with respect to its eight nearest neighbors. (Deutsch, other words, there is a stroke to the right of 1966). the stroke being followed,... Variations on contour-following techniques have In Unger's SPAC, or "Spatial Computer" (1958), been or are being used.564a In discussions of apaper arectilinear networkoflogical modules was by Sprick and Ganzhorn (1959) at the International proposed in which each module has direct contact Conference on Information Processing, several with its four immediate neighbors, and in which techniques were described. For example, Elkind all modules simultaneously receive an identical obtained 85 percent accuracy in experimentalrec- command or instruction from the master control ognition of handprinted block capitals by deter- unit. Programs have been written and tested to simulate SPAC in the recognition of handprinted mining slopes of character lines, dividing the slopes alphanumeric characters and in the detection of into three categories, and determining the number L- and A-shaped features in sets of randomly of incidences for each category per character. Work patterns 5.66a at the Dahlgren Proving Ground has also been re- drawn For characterrecognition ported in which curve tracing was employed, but operations, the Unger technique consists first in the input pattern elements consisted of indications smoothing,imageenhancement, andclean-up of horizontal and vertical motion and of transfers operations. These operations fill in holes in other- from one mode to another. wise black areas or small indentations in otherwise . . straight edges, they eliminate isolated "black" A variation of contour following that isnon- cells including those that create small protrusions reentrant was used in a prototype reader for hand- from an edge, and under certain conditions they written numerics at Rabinow Engineering,a sub- fill in missing corner points. sidiary of Control Data (Holt, 1964). Greaniaset al. For 34 alphanumericcharacterstested,34 (1963), for numerics only, used a helicalscan to features or properties were used by Unger for trace large segments and a small circular scan to discrimination. These are primarily features that examine detail.565 A principal difference between can be detected by contour-tracing (horizontal the Greanias developments at IBM and the Control cavity open above, vertical cavity open to the right, Data approach is that, in the lattercase, Holt's for example), but the list includes some relative "watchbirds" involve theuse of conventional scan- position - dependent and proportion-dependent ners and tracking circuits that record beginnings, properties as well, such as "leftmost point of a 36 vertical cavity open to the right lies in the upper The problem of automatic signature identification two thirds of the figure", and "heightof the left leg is also being explored, for example, at the Tech- of a V-shaped figure less than half the height of the nische Hochschule, Karlsruhe, where an off-line right leg." Although the processing operations are input de-ice permits conversion of either light-pen carried out simultaneously and in parallel over the input or microfilm images to magnetic tape. The entire image field, the choice of the `next step', account number of the presumptive writer isknown given the outcomes for any one operation, follows a and the problem is to identify the signature for rec- decision-tree structure. Fritzsche (1961) has further oznition purposes 5.75 Studies reported byMauceri considered applications of the basic Unger approach. (1965) involve on-line verifications of signatures to Other character recognition experiments involv- prevent highly expert forgeries, andinclude con- ing hand-printed characters include those at the siderationsof relativeacceleration and pen-to- Astropower Laboratory and those of Fischler,5-68 writing-surface contacts.5.76 Marzocco (1965) has Lewis9569 Minneman,5.7° Spilerman,5.7' Uyehara,5.7 investigated the automatic identification of first-name and Weeks,5.73 among others. A particularly difficult signatures.5-76e case isindicated by reports of experiments in Longer range R & D concerns involve the prob- recognizing hand printed Chinese characters at lems of cursive handwriting in general. The original Purdue University.5.74 An approach to the recogni- research approach at Bell Laboratories involved tion of handwritten Katakana characters (table of input from the writer via a captive stylus so thatX Japanese word-syllables), which are composed of and Y coordinates were automatically generated by nearly straight strokes, has been described by his pen. The first significant feature detected in Noguchi et al. (1967). this system was that of relative vertical extent within Another variant on raster-interception techniques, or beyond a baseline and aparallel guide line above such as those of Weeks, is provided by Glucksman, it. A rough first sorting then provided groupingsof who proposes that characteristic locations of types characters with ascenders, those with descenders, of interceptions can be utilized for recognition of those with both, and all other characters. A second multifontcharacters,including thosethatare criterial feature was the presence or absence of handprinted.5.74aAnapproachalsosomewhat retrograde strokes. Abrupt changes in slope (or similar to that of Weeks is reported with respect `cusps') were also detected as were the presence or to developments for handprinted character recogni- absence of loops or near loops and special marks tion in the U.S.S.R., where iterative scanning such as the dot of the "i" or the cross-bar of the processes are applied to anormalized rectangular "t97.5.77 matrix covering the input character image, with Appropriate combinations of these criterial fea- changing inclinations of 'vertical' rows and correla- tures can be used for letter-by-letterrecognition of tions with each of the reference patterns are handwritten words where the word can be seg- obtained for each scan. For computer simulations mented so as to locate its letter constituents with ofthistechnique, 20 personscontributed 50 reasonable accuracy.5.77a Frishkopf and Harmon handprinted characters each and the results were (1960, 1961) also considered possibilities of rec- that "froo. the total of 1,000 characters about 96 ognizing the handwritten word as a whole, again percent v. ere correctly recognized."(Kovalevsky, emphasizing that the highly variable and non- 1966, p. F67). essential details of a particular source pattern should be eliminated as far as possible and that "Lean ing" models such as the various Percep- labora- thesignificant features should be isolated and tron models, those developed in Steinbuch's preserved.5.78 tory atthe Technische Hochschule, Karlsruhe, D. A. Young in 1960 suggested the automatic Gamba's 'PAPA' devices, and Conflex I, have been input and recognition of handwritten computer designed primarily to test various theories with programs, and current investigationsalong this respecttosimulationof human processes of perception and concept formation or to artificial line include work at the Technische Hochschule, intelligence problems, rather than with respect Karlsruhe, with respect to handwritten words as to the development of practical characterreaders. used in ALGOL programming. Also under investi- These will therefore be considered briefly in Sec- gation at Karlsruhe are techniques for the automatic pattern classification of handwritten characters, involving tion5.7, on theoretical approaches to consideration of various methods of determining the recognition and characterization. discriminating criteria as derived from large samples A number of R & D efforts directed toward pattern of characters. The methods used are similar to recognition and classification problems in general those suggested by Kamentsky (1961)5.78a and by have involved the use of handprinted characters for Uhr and Vossler (1961), and Simplex methods tests and demonstrations. Among the manypossible are also being tried. An integrated scanner-com- examples we note the following: puter system can be used to simulate linear classi- (a) McLaughlin and Raviv (1968).5.74b fying networks and to investigate automatic classifi- (b) Chen (1966).5.74e cation techniques under feedback contro1.5.79 (c) Prather and Uhr (1964).5.74d Eden and Halle (1960) have discussed both the (d) Chow and Liu (1966).5.74e synthesis of cursive handwriting and its analysis, 37 finding that 18 strokes appear to be discriminative of the input pattern. Topological relationships are for well-formed Latinscripts.5-79aNeisser and preserved under rotation. The problems of detect- Weene (1960) have also studied human recognition ing relatively invariant features for the recognition performance, using the same handprinted upper of geometric shapes have also been attacked, both case letters used in theSherman machine experi- by criterialfeaturesextraction and by "para- ments, to determine types of error, overall accuracy, propagation" techniques 5.8'2 and confusion data. In general, however, continuing R & D require- Mermelstein and Eden (N64) have based their ments for automatic character-recognition progress handwriting recognition experiments upon analyses require increased attention to: esoteric alphabets; of the intrinsic hand movements that are involved hand-printed characters other than the numeric in writing cursive script. Words were segmented set alone (and, here, for improvedefficiency for into strokes and the strokes were recognized in the numeric case such as Zip-coding); cursive hand- terms of the statistical probabilities of theirbelong- writing for both automatic signature and word ing to pre-selected classes.5-8° Still other examples identification; the further exploitation of context- in this area are provided by current investigations predictivetechniques; and detailed fact-finding at MIT,581 by Simek and Tunis (1967) 5.81aand .investigations as well as experimental development by Teixeira and Sallen (1968).5.811) of format differentiation and field-content-identification procedures5.82a 5.4.3. Shapes and Drawings The automatic recognition of cursive handwriting shares with automatic speech recognition, more- Techniquesbeingdevelopedforautomatic over, difficult problems of segmentation5.82b The- character recognition should also eventually be problems of automatic speech recognition com- applicable to the recognition of simple geometric prise a particularly difficult area of pattern recogni- shapes and of schematic stylized graphic material tion R & D requirements for other reasons as such as is found in line drawings and electrical well 5.82c circuit or chemical structure diagrams. The potentialities for automatic recognition of graphic infor- S.S. Speech and Speaker Recognition mation specifically int!ude the problems of machine encoding of chemical structures. The Perkin-Elmer Some of the difficulties of automatic speech Corporation, which has been active inpattern recognition relate to the physical, physiological, recognition developments for blood-cell identifi- and psychological phenomena of vocalization;5.826 cation, has also explored the problems of machine -others to cultural phenomena, such as broad varia- diagrams. tions in the evolution of spoken-languages and the recognitionofchemicalstructure continual emergence of dialects at many levels;-83 Investigations in this latter area by Cossum et al. and still others to differences which occur for one (1964) have been mentioned previously (p.23). The first approaches to machine recognition of and the same speaker.5-83a simple shapes in line drawings had already been In the first category, that of physiological factors. demonstrated in the 1950's, for example, by Shepard marked differences are found as between maleand. (1959), Harmon (1960), Hodes (1961), and Singer female speakers.5-84 In the second category, it is (1961) among others. At the Western Joint Computer noted that "another difficulty arises from the dif- Conference held in Los Angeles in May 1961, ferent dialects in common use. Here the same words Uhr and Vossler (1961) reported additional results or phrases spoken by differenttalkers will have from a pattern recognition program, including different phoneticcontent. Thus transliteration recognition of outline drawings of shoes, chairs, from a sequence of phonetic elements to English and comic strip cartoon faces. Fain (1960), a Soviet words may involve complex linguistic structure." scientist working in the field, has been reported as (David and Selfridge, 1962, p. 1098). The converse investigating possibilities for recognition of three- of this problem is that of homonyms words that dimensionalobjectbyatechniqueinvolving are pronounced the same althoughthey have dif- possible projections in terms of a grid mask. ferent meanings and may be spelled differently, Questionsof machine-usable techniquesfor such as "bear" and "bare". Thus Hogan remarks: shape-recognition in the abstract (e.g., triangles as "Problems of recognition from continuous text and vs. squares, regardlessof size or of position in the the linguistic resolution of acoustic ambiguities input-image-plane), have been explored by Harmon, (e.g., two, too, to) are problems that will be with (1960), Unger (1958, 1959), Deutsch (1955), Stevens us for a long while." (Hogan, 1966, p. 93.)

.(1961), and Glucksman (1965), among many others. Investigations into the possibilities for automatic Harmon, in particular, developed a special device speech recognition began early at the Bell Tele; for Gestalt-type recognition of line drawings of phone Laboratories (David, 1958, cites unpublished circles, triangles, squares, etc. With the use of work of Kersta in 1947 and of Galt in 1951 with a dilating. circular scan,similar transformations reference to recognition of spoken numbers from 'areobtainedfor geometrically similarfigures, sound spectrograms 5.85), at the University College with variations of source pattern size being trans- of the City of London (Fry and Denes, 1956, 1958) lated into time-of-arrival changes in the derivation and at M.I.T.5.85a Early work in automatic recogni- 98 tion of spoken digits includes the investigations of Tunis, 1966, p. 70). However, there is much less Davis et al. (1952) and Sebesyten (1962). At the agreement among speechrecognitionworkers NPL Symposium on the Mechanisation of Thought (and in fact there is much continuing controversy) Processes, held in Teddington, England, in 1958, a with respect to the choice of which features in recognition device was demonstrated that served, speech are significant and discriminating. The in effect, as a phonetically operated typewriter. 1963 remarks of Garvin and Karush are largely Other approaches to a phonetic typewriter principle still pertinent today: "The phonetic substance of include work at Kyoto University, Japan (Maeda and language clearly lends itself physically to both Sakai, 1960; Sakai and Doshita, 1962, 1963), and at reproduction and measurements, but very few of RCA (Meeker and Green, 1961; Meeker et al., 1962). the parameters required for the recognition of the A vowel recognition technique, using clusteringlinguistic content of the acoustic signal have as methods for the classification of presumably rep- yet been found." (p. 369) resentative samples, isdescribed by Dammann One of the most obvious attacks on problems (1966).5.851) of mechanized speech recognition is that of the Continuing R & D efforts in the United States in measurement and anlysis of "formants"."° How- speech analysis and recognition are exemplified by ever, while it has been argued that "the success projects at the Air Force Cambridge Research of speech synthesizers based on Dudley's model Laboratories (Petrick and Willett, 1960), at the has adequately demonstrated that itis the fre- Massachusetts Institute of Technology ( Forgie and quency spectrum (specifically, the energy concen- Forgie, 1959), the University of Michigan (Peterson, trations in the frequency domain, or `formants') 1963, 1966), and work at Philco-Ford,5.85c among of the speech waveform...that is the information others. Other Japanese examples include work at carrier" (King and Tunis, 1966, p.67), various the Electrotechnical Laboratory, Kokusa Deusin investigatorsarenot convincedthat formants Denwa Company, Tohoku University, and the Elec- are good measures for use in automatic speech tric Wave Research Laboratory.586 Further British recognition systems."' efforts in the field may be exemplified by work re- David and Selfridge (1962, p. 1098), point out ported by Lavington and Rosenthal.5.86a Some recent thattypicalrecognitiontechniques use either European examples will be discussed below. It is "spectral templates or parametic property lists", to be noted that a 1966 review by Bhimini et al dis- givingasexamples formantlocations,overall cusses both speech analysis and recognition prob- intensity,andvoice-unvoicedstatus.Earlier, lems in the context of man-machine systems. David had suggested that the appropriate param- In terms of experimentation and instrumentation eters of speech analysis should be functions of difficulties, there is first the question of analytical the voice pitch and that "pitch detection is funda- equipment available to detect and measure the mental to the whole problem of speech, analysis, properties and characteristics of speech in a given and quite likely to speech recognition as well, speaker's utterances, and there are secondly the for a variety of talkers and pitches." (1958, p. 307.) problems of adequate sampling to set up a compre-. Gold (1962) has investigated computer programs hensive basis for recognition of many vocalized for pitch extractions from recorded speech. Reddy sounds and many different speakers.7 Also, it (1967) describes a computer process that deter- has been noted that "the acoustic power available mines pitchperiods by the recognition of the for speech recognition appears to be rather small. peak structure of the speech waveform. This work at The significant fact here is that this small percentage Stanford University, supported by ARPA, should be tends to be masked by the rest." (Garvin, 1965, p. useful in speech analysis since changes in the pitch 112.) ofa soundcontaininformationabout phrase The effectiveness of preprocessing operations boundaries and points of stress. Moreover, it is such as is so important in character recognition claimed that "more intelligible and higher quality is thus at least equally critical in speech recognition. speech can be obtained in speech communication For detection and extraction of criterial features using Vocoders by transmitting pitch information in speech, there has been frequent use of filter along with the bank-compressed spectrum of the bankanalyzers 5.88andsonographequipment sound." (Reddy, 1967, p. 343. Further references which provides a continuous record of frequency are to Flanagan, 1965, and Schroeder and David, and amplitude as against time.9 A sound spectro- 1960). Elsewhere, Reddy (1966) has considered graph developed at the Communication Sciences some of the problems of segmentation of speech Laboratory, University of Michigan, incorporates sounds. special features including the automatic marking The approach of Kusch (1965) at the Telefunken onto the spectrogram of time, frequency, and Laboratories, Ulm, West Germany, is based on the amplitude scales. (Dunn et al., 1966). assumption that the formants (or characteristic In general, "feature detection performs a trans- frequencies) do not provide sufficiently reliable formation on the primary measurement space, informationfor speech recognition because of withtheintentionof producing a secondary frequencydisplacementsof sounds from one measurement space in which the effects due to ... speaker to another. Instead, Kusch is concerned distortions are largely normalized out." (King and with the derivation of characteristics from both

39 the fine and the coarse structures of the con- at the 90 percent levelfor vowels and 70 percent tinuous speech signal. for consonants with some exceptions.5.93a For the recognition of spoken numerals (where, In the U.S.S.R., also, it is to be noted that "auto- in German, a number of different pronunciations matic machines discerning sounds of speech, are are found), Kusch first uses twocoding matrices, being designed with the aim of searching general the first for detection of sound groups specifkld by peculiarities of the spectrum that are retained after .the binary formulas for presence or absem..e of various transformations." (Ivanov, 1961, p. 15).5-94 high amplitudes at five time intervals. The second A spoken-word recognizer involving adaptive fea- coding matrix relates the detection of presence or tures is also under development atthe Standard absence of sound-group indicia to the identification Telecommunication Laboratories in England. (Hill,' 'of the 10 spoken numerals, in either German or 1966).5-95 English. It is claimed that, for well pronounced Returning to the United States, relatively recent digits and for small differences in sound intensity, speech recognition research is to be noted at the the equipment gives excellent automatic recognition. Communication Sciences Laboratory, University However, in colloquial speech, many sounds may of Michigan (Peterson, 1966), the National Cash be pronounced indistinctly, too loudly, or too softly, Register Company (Otten, 1965), Lockheed's Palo often with large differences in sound intensity occur- Alto Research Laboratory (Bhimani et al., 1966), ring within a single word. Therefore, the one com- and elsewhere. Tappert (1966), uses computer' ponent is divided into two parts with low andhigh simulations to investigate neural models of the sensitivitiesrespectively.Inteststodate, for speech recognition process, incorporating preproc- experiments with 37 male and 37 female speakers, essing operations that detect time sequencesof an overall average recognition accuracyof 87 features. In an ONR-sponsored project at Columbia percent was obtained without adjustmentsof the University the objective is to investigate problems equipment for the different speakers, and without of automatic speech recognition and synthesis with compensation for wide variations of natural sound emphasis on spectrum analysis using Gaussian intensity and wide spreads in pitch. A second set of filters and subsequent computer processing of the experiments involved normalization of the sound data. (Harris, 1966.) intensities to the same level and achieved improve- A correct recognition rate in the 90 percent range ment in average recognition accuracy to 93 percent has been reported by King and Tunis [1966] for correct automatic identifications5.913 different male speakers using limited, arbitrary Von Keller (1966) at the Technische Hochschule, vocabularies of 15-30 words (i.e., those used in a .Karlsruhe, has also explored several alternatives, specific programming language).5-96 However, we 'including frequency, analyses, autocon.elation, and may emphasize the limited natureof results in zero crossings of the speech wave.He points out speech recognition research and development to that the measurement of formants automatically date by noting that these authors claim: "It may involves difficulties because they tend to lie be- be concluded that the techniques investigated in tween the lines of the frequencies. He istherefore this work are adequate for achieving a forced concerned with first zero crossings of autocorrela- decision recognition rate of at least 98 percent over tion functions for the first two formants. Other a range of male speakers andfor arbitrary vocabu- aspects of the zero crossing analysisapproach in- laries of up to thirty words." (King and Tunis, 1966, cludes determinations of the time intervals between p. 78 (underlining supplied).)Petrick and Willett zero crossings on thespeech wave curve and the (1960) claimed similar accuracy for a vocabulary number of such crossings that fall within certain of up to 83 words. A demonstration model of an time intervals.6-92 adaptive recognition device, "Cynthia III", suc- At the Institut fiir Phonetik and Kommunikations- cessfully recognized up to 25 words after training forschung, University of Bonn, under the general with a particular speaker. (Lesti,1963).5-96a By direction of Ungeheuer (1965), developments for contrast, it is suggested that "no serious forecast' speech analysis and recognition include phase about computer systems in the 1970's can omit spectrum analyses and derivation of auto-correla- voice recognition systems with several-thousand- tion functions of fricativt,z in the initial position; word vocabularies." (Diebold, 1966, p. 297). equipment to measure width of intervals between oice recognition, recognition of cursive hand- zero crossings for clipped speech.A system of wnting and the isolation of particular characters or Ungeheuer's design is used to separate high and shapes inmultiple-pattern or noisy background low frequency components, to integrate filter out- share problems of both segmentation and re- puts every three seconds, and to provide output integration. For example. Davies points out that one results to an XY oscillograph. A prototype word of the more difficult problems of speech recognition recognizer was begun in 1962 and demonstrated at resembles that in machine stenotype recognition the ICA Congress in Belgium in 1965.5.93 where it is necessary to go to the level of syntactic Sakai and Doshita (1963) also make use of zero- analysis before word separations can be made. crossing analysis, using the phoneme as thebasic (Davies, 1962, p. 67). Then we note that "as K. N. recognition unit. They report satisfactory recogni- Stevenssuccinctlyputsit,'thesegmentation tion with respect to Japanese monosyllables as being problem is the fact that you can't segment'." 40 (Lindgren, 1965, p. 51). [based on private com- The converse of the latter point (that is, to find munication]. the speech characteristics peculiar to a given Garvinpointsto"the verydifficult speech- speaker), is required in the case of automatic recognition problem of recognizing the boundaries st eaker identification which is aspecial, limited of stretches within a semicontinuous signal" and, case of voice recognition. Thus, atthe University further, that "the natural vocal signal is semi- of Bonn, Ungeheuer's system provides the capa- continuous. That is, the number of physically ob- bility for photographing the different domains for servable breaks in the continuity of the stream of movements (on the X-Y display) of coordinate human speech is much smaller than the number of positions specific for different speakers and of discrete elements into which the signal may be thecharacteristicallydifferentvoweltriangles decomposed in either alphabetic writing or linguistic (see Fig. 9). Further processes involve separations analysis." (Garvin, 1963, p. 112-114). of the first and second formants by density distribu- In addition, phonetic signal strings raise special tions of zero crossings, and determinations of the problems of segmentation and of junctures (or number of zero crossings per 3-second intervals, interruptionsto the continuity of a stream of giving other plottings. Using such data, it is possible speech). It is therefore suggested that: "Linguistic to identify speakers independently of the text and acoustic research on the phonetic character- uttered,andsignificantdifferencesbetween istics of junctures will undoubtedly make significant professional imitators and the public figures whose contributions to this aspect of speech recognition." voices they are imitating can be shown graphically. (Garvin, 1963, p. 113). A related problem is that of speaker or "voice Thus, the input to direct machine processing print" identification when the speakers are uttering of spoken messages, for other than a severely certain code words. Tillman (1965) has analyzed restricted vocabulary of approximately two dozen 100 to 200 classes of sonograms of a particular word words, awaits considerable further research and as uttered by 10 different speakers in order todetect development effort in the area of automatic speech differential features, particularly those of a general recognition.5-96b In particular, it can be emphasized nature. Then itis noted that "Prestigiacomo at that "a practical device should be capable of Bell Labs ha:: produced contour spectrograms that automaticallysegmenting acontinuousstream show relative intensities that do not show up on of speech, correcting for the individual's speech theconventionalspectrograms.These relative characteristics and dialect patterns, and recognizing intensity patterns are claimed to be the clue to the information content of the incoming speech individual speakeridentification."(Lindgren, signal." (Bhimani et al., 1966, p. 279). 1965, p. 127). In speech recognition, also, as in sophisticated Earlier investigations at the Air Force Cambridge typesof character recognition and ingraphic Research Laboratories suggested a possible pro- pattern recognition generally, there are continuing gram whereby "an unknown speaker says his word research requirements for the design of appropriate into the computer and obtains one of the following measurement spaces 5.97 and increased appreciation types of responses: 'That was John Jones speaking is developing that many of the problems will not the digit three; I don't know who you are, stranger, yield to solution until far more attention has been but you spoke the digit two, I don't have the slightest giventounderlyingsemanticandcontextual idea who you are or what you said; speak more factors.5-99Specificresearchrequirementsare distinctly and repeat your word again, please'." suggested by King and Tunis (1966, p. 78) as follows: (Petrick and Willett, 1960, p. 15). "Other areas which require additional work include Another example is provided by investigations the problems of word segmentation, and discrimi- at Scope, Inc., which stress the importance of nation of voiced, unvoicee, and mixed speech nasals. The approach uses 25-element vectors to frombackgroundnoise." To theseproblems represent spectra of 30 different speakers with Bhimani et al. (1966) add those of time normaliza- respect to a 50-word vocabulary. Test results to tion of speech. date show 45 percent recognition accuracy for single test utterances ranging up to 93 percent for We note again possible avenues to cross-fertili- 10 utterances. Pierce points out that "progress is zation in the fact that in many language-data- being made in the automatic identification of 41. processing operations (to be discussed in a later speakers by voice, as a substitute for a hand- report in this series), the problems of determining proper segmentation as between word rootsand written signature, and this could play an important affixes are often critica1,5-99 but may yield to the part in banking and other credit transactions." same types of solutions as may bedeveloped for (Pierce, 1966, p. 148). speech recognition and the recognition of cursive A commercial approach to the solution of voice- handwriting. King and Tunis point to still another print identification problems is represented by continuing research difficulty as follows: "Trans- the recent purchase of Voiceprint Laboratories, formations that will eliminate the effects of varia- Summerville, New Jersey, by Farrington Manu- tions in speed of talking and other differences in facturing Company.5-m format structure between speakers must still be A 3-year program to explore the use of voice-. found." (King and Tunis, 1966, p. 78). prints in criminal identification has been initiated

41 w Abb.2: Sprecher S Abb.3: Sprecher Kr

Abb.4: Sprecher R Abb.5 Sprecher U

FIGURE 9.Patterns for different speakers. by the Justice Department in cooperation with the At the Argonne National Laboratory, the CHLOE Michigan Department of State Police."°a equipment, consisting of a flying spot scanner under We, note of course that to identify a new "sample" the control of a small general purpose computer, of handwritten signature or vocal utterance of a has been used for experiments of machine classifica- password or impression of a fingerprint as being tion of fingerprints (Shelman, 1967).5.b0" sufficiently"like" prior samples as to warrant The achievement of rotational invariance in finger- access, for example to privileged files, is theoreti- print scanning and processing by optiCal or holo- cally far less of an R & D problem than to detect graphic means is discussed, for example, by Marorn, and identify patterns of speech, or fingerprints, 1967 5-100g and by Horvath et al.,1967.5-mhThe or cursive handwriting, as such. continuing interests of the F.B.I. in this area of pattern recognition are discussed by Voaker..51"' 5.6. Other Pattern Identification and Semi- automated systems are also to be considered, Recognition Areas as in the case of the New York State Identification and Intelligence System (Kingston,1967).53(44 The difficulties with respect to the automatic' . pattern identification of fingerprints relate on the Gaffney, of the Federal Systems Division, IBM, one hand to the extreme complexity of these has developed a method of encoding fingerprints patterns and on the other hand to the very large based upon information about bifurcations (ridge numbers of patterns on file (i.e., typically, in the splits) for processing against a similarly encoded tens of millions) against which an input pattern or file.5'°1 The SDC project is similarly concerned with set of patterns is to be searched. Some relatively ridge characters but in the context of man-machine recentapproachestoproblemsof fingerprint interaction in the process of location and identifica- identificationare represented by the work of tion.5.1°2 In terms of continuing R & D concern, it Gaffney at IBM, Cuadra at System Development is well to note the following comments by Cuadra: Corporation, and Wegstein at the National Bureau "Various researchers concernea with the auto- of Standards, among others. Other investigators mation of fingerprint identification have been include Fulton,5mb Trauring, 5.104c Van Emden,5.md exploring quite different approaches to the prob- and Hank ley and Tou.5-me lem. Some are attempting to use fine details of 42 the fingerprints, automatically identified by opti- electroencephalographic analysis have been cal scanners, as the basis for fingerprint descrip- mentioned previously.* Additional examplesare tion; some propose to use automatically derived provided by Weihrer etal.(1967), who discuss measures based on the slope of the ridges in computerized electrocardiographic analysis various areas of the print; and someare attempt- developments at the Instrumentation Field Station ing to derive measures based on the fingerprint of the Public Health Service.51"6 European examples pattern as a whole. What is significant about all include the work of Sneddon at the University of this work, little of which is described in theopen Glasgow 5.1°7 and of Karlssonat the Royal Institute literature, is that there has not yet been sufficient of Technology in Sweden.5-1 " Related medical data research to provide a rational basis for choosing processing applications include x-ray and radio- a particular approach." (Cuadra, 1966, p.8, isotope interpretation,5'"" phonocardiographic,5-j°9 underlining supplied.) and radiocardiographic analysis.5-"° Ledleyet al. The approaches of Freedman and Pietnanen at (1968) have used machine analysis toprepare mitotic Bendix 5-102" and Wegstein (1968) at 1\,3Sare also indexes(i.e.,measuresoftheextentof cell based on minutiae of the fingerprint pattern, specifi- division).5-' I" cally ridge ending and bifurcations (Fig. 10), and We also note thecase of pattern detection in with constellations or groupings of these minutiae. seismic recordings (Rosen, University of Uppsala, Descriptors corresponding to the identification of for example), and that pattern recognition in thecase constellations are computer-generated and degree- of bubble chamber data has been reportedat the of-match scores can then be computed for thesame Centro Italiano di Studi ed Esparienze in Milan. and different fingerprint patterns. An important (Swanson, 1966, p. 9). However, it is to be empha- advantage of this technique is that identification sizedthat"bubble chamber picturescanning can be achieved from minutiae occurring in only a defines a context for visual data processing and portion of the total print pattern, so that searches pattern recognition in which the concepts 'proto- based on chance or latent partial printsmay be types' and 'images' become virtually meaningless." successful.5-102b (Narasimhan, 1966, p.167). Particularly difficult Hank ley and Tou (1968)are investigating tech- arerequirementsforthree-dimensionaldata niques involving elementary spatial filtering, image processing and for "scene" analysis suchas have enhancement operations to close ridge gaps or to been explored in the CYCLOPS 1 project at Bolt, separate contiguous ridges, and topological encoding Beranek, and Newman 5.11', and in continuing work of features detected in the center region of a single by Minsky and others at M.I.T.5-"2 print. These investigators point out that "the chief Looking forward, in the area of automatic pre- obstacletoautomatic fingerprint processing is processing operations upon input, we can foresee print quality. Even good quality fingerprints are increasing development and use of techniques of subject to a variety of imperfections....Most automatic analysis as applied to graphic, auditory, common are ridge gaps, usually caused by skin three-dimensional, and other types of source data folds, and contiguous ridges, which may be caused inputs. This area willthus include extensions by spreading of ink (or skin oils) due to finger and improvements to character and pattern recog- pressure, by particles on the skin, or in the worst nition methods and equipment, to developments cases by excessive inking or by smearing during and extensions of voice recognition techniques, rolling of the finger.5-1 °2c Somewhat less bothersome and to continuing research in the areas of pattern are light prints, thin ridges, and excessive pore recognition, detection, and identification. Wooster structure, which appear as dotted or perforated suggests that "the most exciting step of all will ridges. Print processing must also allow for damaged come when we are able to study pattern recognition or scarredfingers." (Hankley and Tou,1968, in text....How does a scientist realize that a pp. 417-418). piece of work in, say, psychoacoustics contains the Then there are some areas of R & Dconcern with cluetosolvinghisproblemsincloud cover respect to relatively less obvious applications of analysis?" (Wooster, 1964, p. 13). pattern recognition techniques. Some quite exotic examples include recognition of plant species from leaf venation patterns,3 terrain and cloudcover 5.7. Some Theoretical Approaches to Pattern analysis,5-ma therecognition and identification Recognition and Categorization of audio waveforms in coughing,5.i°4 the recognition of printed sheet music,5.i°5 or even the rumps of Finally, in the area of preprocessing operations migrating baboons! 5'105° In addition, various other and pattern recognition we may consider a few of types of "pattern" detection, identification, analysis, the problems and prospects for automatic pattern arid, in some cases, generation are to be noted in detection and patternclassification. There are various types of musicological, as inliterature, problems of both pattern detection (emit analysis research.-5-mb For example, for purposes of struc- orclassificationconcentrating upon similarities tural analysis of atonal music, Forte (1967) manually among evt :itsor sequences of events) and of encodes music scores in a special notation .5-1°5c pattern differentation (etic classification)." 3 The application areas of electrocardiographic and Sec p. 24.

366-107 0 - 70 - 4 43 MO... 0.- 01.14411. ^4INIMIINI 4 R & D attacks on the problems of pattern identi- with efficiency, tractability, and generality, and with fication and detection, generally, have to do with the minimal amount of prior information." (Cooper, problems of automatic means for the detection of 1966, p. 1).5-"8 At the Ballistic Research Labora- memberships inclasses,5-"3abased both upon tories, U.S. Army Materiel Command, investigations properties that are similar for members of the same have included problems of pattern separation by class and upon properties that differentiate between linear programming techniques (Taylor, 1967) and members of different classes 5-"3r' Observedprop- consideration of mixtures of variates as a separate erties of both tyres may be considered to define a class in pattern recognition problems. (Sacco, 1967). multidimensional or metric space in which distances between given property vectors of different input Adaptive principles and feedback mechanisms patterns can be used to separate one class from are employed to a greater or lesser extent in most of thesetechniques. Inthesimplest systems, another. For example, questions of feature selection, involving mask- or template-matching, adaptation featureeffectivenessestimation,andvarious is especially to be found in training sequence or measures of distance between pattern vectors in n- `teaching sample' procedures in which a number of dimensional feature space in terms of recognition- exemplars of each character in the set of characters error probabilities are explored by Fukunaga (1968). to be recognized are analysed in terms of the For another example, work by Bledsoe (1966) particularly discriminating or criteria! areas of each concerns the mathematical treatment of problems proposed mask, so that these areas may be differen- inmulticategory patternr: cognition operations. A continuing R & D question is the extent to which tially weighted for optimum discrimination between hardware considerations may dictate the measure- the various members of the set. To the basic adap- ment space.' "4 tive mechanism based upon sample data there Secondly, there are both practical and theoretical may be added other decision-functions, such as problems with respect to the choice of the prop- those designed to minimize risks of misrecognition. erties to be measured and used in defining the (Chow, 1959, various references, and others) .5."8a recognitionspace. The methods involved may The most obvious criterion of choice for a minimal range from designationof probably significant set of properties or features would be the applica- features on the basis of intuitive, a priori grounds tion of information theoretic considerations."8b to extensive computer analyses of the similarities Lin (1964), for example, considers the problem of and differences among many samples of each of the generating weighted area correlation masks for various classes to be recognized. A generalized alphanumeric character recognition. The solution consideration of measures of similarity in terms proposed uses an information-theoretic, measure of pattern classification tasks of various types is to identify "high-information" points from among provided by Baker and Triest (1966). all retinal points, and a random selection procedure Next, thereare the theoretical problems of for picking out five to seven points at a time from minimizing the number of property measurements among the high-information points. The mask ele- to be made on input patterns (reduction of the di- ments are assigned plus or minus values in such mensionality of therecognitionspace) and of a manner as to maximize the value of the informa- optimizing the discriminating power of the set of tion measure. The final selection of recognition properties that are retained.5"5 Important in this logics..."is made on the basis of a distance area of theoretical pattern recognitionresearch is measure between classes." (Hart, 1966, p. 15). the increasingly recognized need to use machine Unfortunately, for practical applications, such techniques for the systematic evaluation of possible patterns of interest as character shapes and sounds criterial features held in common by exemplars of of human speech reflect a variety of other factors, members of useful sets or classes.5-"8 An important some of them of physiological or evenaesthetic example of recursive random generation of feature- origin which do not necessarily accord with the detection operators for automatic character recog- For nition is provided by Uhr and Vossler (1961). requirementsofinformationtheory.5-"8c Representative of relatively recent approaches to example, in the Roman alphabet, the upper-case solution of some of the problems in this area is an character "0" is distinguished from the upper-case ONR-sponsored program at Information Research "Q" only by the presence or absense of a descender Associates, Inc. Objectives are to determine those tail.In the Cyrillic, the characters "Schah" or feature, property, or character extraction techniques "Schahchah" are similarly distinguished, but both that will best exploit differences between classes, characters are wider than the normal character that will exhibit minimum sensitivity to variations matrix requirements. within a pattern class, and that will be of maximum For such reasons as these, information theoretic efficiency as class distributors.5 "7 considerations do not necessarily apply to optimal The RADC-supported program of Cooper at design of character recognition techniques. These Sylvania Electronic Systems has as its objective reasons are, in effect, additional evidenceof the the development of "a mathematical theory of current lack of techniques to isolate, discriminate pattern recognition, with present emphasis on tech- between, identify, and use effectively,pattern- niquesfor performing nonsupervised'learning' discriminatory and selection criteria. 45 As has been pointed out, co-occurrences of properties among the members of R & D require- "Estimation isbasic topattern recognition a presumptive class. A continuing where a class of observation waveforms or vectors ment is indicated by the fact that apparently not describe a pattern. This is because the number enough is known as yet about the comparative of classes, the statistics for each class, and the a efficacy of various statistical association formu- priori class probabilities are in general unknown. las,5-121 of which more than 50 can be identified,5122 Both supervised estimation (where samples are and that finding a suitable method for making com- verystickyproblem." classified)andunsupervisedestimationare parisonrepresents"a important. One reason unsupervised estimation is (Baker, 1965; p. 153). important is that the performance during oper- Maron suggests that"it would be useful to ation can improve if and only if unsupervised strengthen the theories (which presently are not always clear) behind some of the current techniques estimation is utilized." (Monds and Carayanno- in order to provide logical justification for their polis, 1968, pp. 9WI. preference (over alternatives); i.e., to have some Chow and co-workers have suggested that the measures of the goodness of alternative association detailed structure of a recognition system can be techniques." (Maron, 1965, p. 12). systematically derived in the functional form of Adaptivepatternrecognitionsystems based probability distributions."8d Thenearest neighbor uponlearningmodelsand"forcedlearning" dependence method goes beyond the assqmptions of procedures are another important area of continuing statistical independence 5 "8e The recognition net- research and development. For example, Greenberg work consists of three levels: a layer of and/or and Konheim describe experiments in the computer gates, a set of linear summing networks inparallel, generation of separating hyperplanes, on 32X32 and a maximum selection circuit. Formulas for the scanner output for training sequences of alpha- weights of the recognition parameters are derived numeric typed characters, with error rates for test as logarithms of ratios of conditionalprobabilities. characters ranging from 0.7 percent (original typing) These formulas then lead to a procedure for esti- to 1.7 percent for inputs including carbon and Xerox mating weights from sample characters which are copies. A second experiment on similar training and then used in subsequent recognition. (Chow, 1965). test data, with comparable results, generated binary Other measures proposed include the use of search separations. "Here linear functionals were statisticaldecision and probability theories.519 formed to successively split the alphabet into two As previously noted, some techniques under con- groups, each of those two more, and so on until tinuing investigation involve systematic searches individual letters were identified. Linear separability for features of properties that remain relatively was demonstrated." (1964, p. 306). invariantundertranslationsofmagnification, Historically, the first investigations into adaptive rotation, or translation in the input image area.5-12° or self-organizingmethods of pattern detection and For example, Giuliano etal. (1961) developed a recognition are exemplified by Uttley's "conditional computer simulation program based onthe measure- probability machine" experiments, where the device ment of the first ten momentsof the character "learns" to detect and follow black-white boundaries. pattern, normalized as to positionand scale, and (1956, 1958, 1959) Gamba's learning-recognition on discrimination by meansof a distance-measuring systems (the "PAPA" devices) arebased on a algorithm to give quantitative measures of similarity criterial-crossings technique, where the interactions between characters. They concluded that the ex- sought between input patterns and stored reference perimental system showed reasonable discriminat- patterns are those of intersectionsof the input ing power in the presence of print disturbance image with a pre-established intermediary pattern. in addition to being insensitive to variations in In this case, a configuration of randomlines is positioning, scale, line thickness, or heavyness of produced by a computer-based pseudo-random print which, in this system, are compensatedfor by generator,andassociationalorprobabilistic- normalization. inference weights are built up from a training In the area of pattern detection and classification sequence of sample "characters"in terms of sub- research, an obvious possibility is the investigation units of the random configuration.Specifically, of adaptive systems. Techniques of continuing the "features" or "properties" of the input image R & D interest include the use of statistical inthat are extracted include such possibilities as ference methods and the use of learning models. whether the total number of intersections are even Both techniques have been studied, for example, or odd, the number of intersectionsin the first by Kanal et al. (1962) for potential applicability to half of the image area as compared to those in the recognition of underwater sounds, speech with second half, whether more or fewer intersections respect to a limited class of utterances,transmitted occur to the left of the centerof gravity of the complex wave forms, and multifontcharacter input image than those occurring to the right, recognition. whether there are more or less than n intersections, Statistical inference methods may generally be and so on."3 based on various statistical association techniques In the United States, pioneering efforts in the that provide measures of similarity derived from area of self-organizing areusually associated with 46 the "Perceptron" research begun by Rosenblatt for some but not all samples, or "good" for some, but in 1957. Extensions of this technique by Rosenblatt not "high", and to convert them to "high and good and his colleagues and by other investigators go for all" by logical combinations of "and" and "or". to multi-layer devices capable of feature extraction (p. 356).5-132 and to multiple levels of training and "learning". Of particular interest to future automatic pattern For example, Block et al. (1964) describe a pro- recognition capabilities, therefore, are experiments cedureforadjustingtheassociationalweights aimed at developing methods for adaptive extraction (built up as a result of reward-punishment feed- of significantfeatures or properties of sample backs) in the first layer of a two-layer Perceptron patterns, especially where those properties likely model in such way as to improve detection of to be most useful in discriminating between members significant features, without regard to themariner of different pattern categories are not known, or of training and weight adjustment in the second are not well-known, in advance. In such cases, a layer as related to desired responses.5124 At M.I.T., training sequence of representativeor sample Minsky and Papert are concerned with similar patterns may first be used. A continuing R & D problems (1967). requirement of a fact-finding nature is, therefore, Adaptive majority -logic schemes are represented for studies of the true representativeness of samples not only by multi-layered Perceptrons, but also and of the applicability of results based upon such by the Conflex devices of Scope, 111C. 5'125 and by samples to real-life situations. Beyond these ques- developments at the General Post Office in Great tions are serious technical problems as to the extent Britain."26 Kovalesky (1965) and Barus (1962, 1966) to which classes are in fact separable and/or of have been concerned with nonlinear transformations whether such classes as can be separated constitute realized when the association units have many useful or meaningful groupings of their respective inputs during the "teaching" process.5'27 i nembers.5132" A "learning" model for a machine system to In adaptive systems using linear threshold deci- recognize three freehand characters ("5", "R", sion elements,5.133 iterations with varying pattern and "I") has been developed at Stanford Research samples over time either produce convergence to Institute. These techniques are also being con- optimal system performance or else the lack of sideredforapplicationtothe classification of convergence indicatesthatthese patternsare cloud photographs and toproblems of speech linear) inseparable.5-'34 Problems of linear separ- recognition and speech intelligibility. (Hall, 1964). ability, then, are important areas for further in- At Bolt, Beranek and Newman, an ONR-sponsored vestigations of useful classes of patterns or sets of project involves the design of a pictorial input pattern property measurements.5-135 system and researchincharacter recognition, Similarly, for the areas of automatic classification, including work on handwritten characters and or categorization generally and for those applicable simple pictorial diagrams. (Strollo, 1966). to documentary items in particular, it has been A large number of additional examples might also suggestedthat "the most fruitful approach to be cited, including Hughes Aircraft's `Multivac' automatic classification research is not so much (Multi-Variant Adaptive Computer) which has been in the development of new models for classification used experimentallyto recognize alphanumeric as in investigation of how useful classes of objects characters, to play simple games, and to differentiate aretobeefficientlyconstructedfromlarge between different kinds of sounds,5-'28 the Minos II universes." (Lehmann, 1964, p. 5-1). at Stanford Research Institute and the ISODATA There are several major causes of difficulties with (Iterative Self-Organizing Data Analysis Technique respecttooverlapping and non-separability of A) procedure for its use,5-129 or CHILD (Cognitive classes. The fir, t and most obvious is the incidence Hybrid Intelligent Learning Device) at the Informa- of noise in the case of inputs from the real world. tion Processing Laboratory, Rome Air Develop- Then there are problems of irreducible ambiguities ment Center.5-'3° between input patterns, whether because of additive Bonner (1962) describes a multi-level program to noise or multilation or because of the fact that the derive or enhance automatically recognition logics character was poorly formed to begin with, as in for determining the significant properties of samples most handwriting examples. of the various members of the population to be Then there is the problem of the use of context discriminated. The method involves: "1) Use of to resolve problems or ambiguous or indeterminate nonexhaustive heuristic algorithms to find 'valuable' pattern identifications. We shall, in general, reserve properties...2) Employment of various devices discussionof context expectancy and context- to make the requirements expected of heuristic predictive-analyses to later consideration of natural method (1) less stringent. These include a method language processing techniques and problems (in of dividing a problem into a number of simpler another report in this series), but we note here that, problems equivalent to the original problem. ..and with respect to automatic speech recognition "the a decision process which can operate effectively same spoken form may well correspond to more than without requiring properties of very high `value'." one written form, and this ambiguity then has to be (p.353).5-13'Bonner'sprogram,further,takes resolved by context searching, which is a syntactic properties found to have "high" and "good" values operation analogous to its equivalent in machine 47 i 1, 1.1

translation" (Garvin, 1963, p. 114), and that "even gated by Vossler and Branston (1964), combining a with very legible handwriting, machines are bound small dictionary (3,737 words) and digramdata, to make mistakes, which theywill need contextual special algorithms were proposed which would information to resolve." (Lindgren, 1965, p. 105). "collect in groups all words which were similar in Some other examples of the importance of context most of their letters but yet differedfrom dictionary checking for purposes of resolving ambiguous words. Such an algorithm might group together the possible pattern identifications were given in the garbled words: CAMBODIA, CAMBIDJH, 1961 NBS state-of-the-artreporton automatic CAMEODIA, CAMEOOIA, CHMBODIA. By de- character recognition. (Stevens, 1961). Dictionary termining the most frequent letter in each position lookup techniques have been proposed by Bledsoe the de-garbled word CAMBODIA can be con- and Browning (1959) and Baran and Estrin (1960), structed."(VosslerandBranston,1964,p. among others, and have beendisclosed in a Rabinow D2.4-7 ) .5'13"e patent.5-'36 Further considerations were raised by Another intriguing example of context checking Alt (1961), and extensive investigations have been is in the case of recognition of chemical structure reported by Thomas and Kass ler (1967). diagrams." In general, however, an Auerbach An intriguing use of context is provided by an survey has concluded:"Context recognition will IBM experimental system for automatic reading of certainlyinvolve an enormous increasein the natural language text, involving cluuer-seeking to storage capacity and logicalcapabilities of charac- identify the distinct types of characters encountered ter readers, but this may bejustified by the increase and a cryptogram-decoding procedure involving in efficiency that can be attained. However, the .pre-storeddigramfrequenciesandpermuted- economicsofcontext-recognitionreaderswill identificationlettertranslationfrequencies.5-136a remain highly speculative until considerably more Edwards and Chambers (1964) simulated a recog- development work has beenundertaken." nitionprogram, introducingnoiseby random (Feidelman and Katz, 1967, p. 0210:32). Similarly, number generator with respect to Alt's calculated "althoughthesyntactic, semantic, and higher moments to provide distorted momentvalues to the order constraints on the language would seem to be point where erroneous or ambiguous identifications applicable to our problem, we do not yet have any would be made. "The same program was then rerun very good ideas on how to usethem." (Earnest, by inserting the digraph statistics. In the latter case 1962, p. 462). the computer narrowed the choice down to two or Another factor that may be considered in an am- more letters, using thedistorted moment values. biguous identification decision is that of expected It then noted the letter it had identified, correctly single character frequency. That is, where the or incorrectly,immediately preceding the letter results of input-pattern-to-reference-pattern match- under investigation and picked the letter of those ing are ambiguous, final choice may bemade under consideration which was most likelyto on the basis of thecharacter that occurs more fre- succeed the preceding letter." (p. 468). quently, for a given language and subject matter Edwards and Chambers note human use of source. An example of considerationsof this type context to correct nosiy characters, e.g.,CW must is reported by Blokh (1960). be either COW or CAW if it occurs in English text. Chow (1957) considers the case where previously They continue: "While it would no doubt be difficult determined noise statistics (specifically including to 'teach' a computer enough facts toperform this the ways in which various characters and noise are type of reasoning, there is gooddata available on frequently combined) are used to determine the the occurrence of digraphs or combinations of two conditional probability density with respect to a letters in the common languages." (p. 465). corrupt and noisy input pattern.Moreover, by The use of digram and trigram frequency statistics associating the derived conditional probabilities has also been proposed by David and Selfridge wita estimates of loss or risk involved in the substi- (1962),5.'37 Harmon (1962),5.137a Sebestyen and Edie tution of a specific character for another, Chow (1964),5.138 Steinbuch and Piske (1963),5.138a Sakai suggests a means tooptimize the recognition and Doshita (1963) for Japanese speech recog- system performance by minimizing therisk-cost. nition,5.138b and the Astropower Laboratory 1964 Finally, in the area of theoretical approaches to report, among others. An intriguingexample is pattern recognition that we have reviewed,there presented by Carlson (1966), where a computer are the problems of automaticdetection of "classes" technique was developed to replace garbled charac- whose definitions and extensions are not known in ters in genealogical records based upon trigramfre- advancethat is, the area of automatic categoriza- quencies for proper names.5.'39 An accuracy of tion.5.14oa In this area, a number of techniques better than 90 percent error correction is claimed. have been experime. tally investigated for a variety Other proposed techniques for contextchecking of applications, e.g., factor analysis, discriminant and error correction in the case of ambiguous analysis, and clumping or clustering methods.5.1401) charactersand"misspelled"words,combine A taxonomic approach to automatic classification digram or trigram frequency statistics withdic- is exemplified by Tanimoto (1958, 1962). (We will tionary lookup procedures5.139aor cyclic permu- defer discussion of some of these techniques to a tations of the input text.5.139b In a system investi- later volume of this report in connection with docu- 48 ment grouping and automatic indexing applications, ing results in terms of selection of clumps corre- briefly discussing here only the latter technique). sponding to recognized diseases with the notable First, Le Schack provides some useful definitions exception of Hodgkin's disease. In the latter case, and distinctions: "A particular case of the classi- however, it was noted that "patients show a chaotic fication problem is considered here. The basic raw collection of symptoms, some being practically data are estimates of the pairwise similarity be- complementary. The only unifying characteristic tween members of the population. It is desired to is that there are a lot of positive symptoms. I can identify subsets in which the members are so scarcely conceive of a dump definition that would similar to one another, and so dissimilar to non- be likely to group these patients;I am unsure members, that it is most useful to consider them as a whether this is a reflection on clump theory or on unit. This process has been termed 'clumping' or Hodgkin's disease." (Needham, 1961, pp. 44-46). `clustering'. It is clear that an exact formulation Techniques of automatic classification or cate- of the clustering problem must be in terms of gorization based upon the clumping methods of optimization; for the goal is the best balance be- Parker-Rhodes and Needham have continued under tween loss of precision, which is the inevitable investigation, notably at the Linguistics Research consequence of representing each individual by Center at the University of Texas.'44 Dammann the typical or average characteristics of his cluster, reports on the use of clustering techniques for ex- and gain in economy, which is the result of replacing periments in voice recognition as follows: "Within many individuals by a few clusters." (1964, pp. the limitations imposed by the number of points XIV-3). within the clusters of the sample space, this experi- Further, "the intuitive notion of perfect clustering ment resulted in the satisfactory separation of eight is best explained by considering a population of unknown clusters of complex patterns in an arbi- elements which can be partitioned into nonover- trary sample space, identified aberrant samples, lapping subsets, such that within each subset there and suggested an output code." (Dammann, 1966, is a nonzero linkage between each pair of elements, p. 88). Karlsbon and Arvidsson at theRoyal Institute but no linkage between a member of the subset of Technology, Sweden, have applied cluster-seeking and an element which is not a member. To gen- techniques to EKG recordings. eralize to the idea of strong clustering, assume that A marked increase of interest in cluster-seeking we are able to distinguish between strongand weak is to be noted in the recent literature, in the broad- links;the definitionisthen modified to read: ened areas of application being explored at CLRU ..Within each subset there is a strong linkage (Sporck -Jones and Jackson, 1967), and in proposed between each pair of elements, and at most a weak practical applications for "autonomous" text-read- linkage between a member of the subset and an ing machines (Rabinow and Holt, 1966; Casey and element which is not a member." (Le Schack, Nagy, 1968).5.144a 1964, XIV-3). The selection and retrieval systems in which It is to be noted first that the linkage or connection statistical associations between text-word-associa- strengths are derived between items on the basis tionfrequencies, index term associations, and of the degree of similarity or association between document-document associations may be imple- them such as the sharing of some number of com- mented to offer additional possibilities of useful mon properties. Thus, for example, "it is appropriate clusterings. 5.145 to take as a connection measure the 'Boolean near- Present constraints on the practicality of the ness'; that is the connection between two objects application of clumping and other techniques for is the number of columns in which their rows agree." automatic categorization and classification include (Needham, 1961, p. 9). thoseof computationalrequirements, memory Initial work on clustering techniques assumed requirements, adequacy of sample size, problems that the clusters would be known in advance5 -'41 in the manipulation of large, sparse matrices, and The clumping techniques of Kuhns, Parker-Rhodes questions of whether the material does in fact lend and Needham of the Cambridge Language Research itself to meaningful partitionings. Thus Doyle re- Unit (CLRU), however, are largely addressed to the marks: "As to the future of research in automatic finding of unknown clumps.5.142 Then we note also classification, we can certainly predict that it is that "Bonner, Hyvarinen; Rogers and Tanimoto; going to be very interesting and productive. As to Firschein and Fischler; Glazer; Stark, Okajima, and its application, the future is less brightthe im- Whipple; Jakowitz, Shuey and White; and Ball mediate future at least. The major barrier is that and Hall have considered the problem of unknown of expense. The required cluster analysis and docu- clusters." (Mattson and Dammann, 1965, p. 296) 5'142a ment .grouping techniques arecharacteristically Computer programs to apply clumping techniques voraciousinconsumptionof computer time." to the properties of a variety of objects have been (Doyle, 1964, p. 29). used in experiments in automatic classification of Beyond costare questions ofcomputational archaeological artefacts such as pottery vessels and capabilities adequate for the handling of realisti- shards both at CLRU and elsewhere.5.'43 Also at cally large universes. It is further to be noted that CLRU an experiment was conducted with respect "certain combinatorial problems are beyond the to classification of medical symptoms, with promis- capacity of even the largest and fastest computers." 49 (Hayes, 1963, p. 284). Then, with respect to realisti- Future prospects for even more sophisticated and cally large universes of items to be categorized or complex pattern detection and pattern classification classified, the following "main point to be made techniques in a wide variety of fields can certainly is that theoretical elegance must be sacrificed to be foreseen. Examples of areas of continuing R & D computational possibility: there is no merit in a challenge are as follows: classification program which can only be applied (1) "The pattern recognition problems of the to a couple of hundred objects." (Needham, 1963, futureareapttoinvolvea tremendous 8).5.146 latitude of variation in the patterns. Design Finally it is to be noted that over-simplifications by humans will be difficult because of the large of pattern categorization problems are to be found amount of data that must be manipulated. in various analytical approaches to problem solution. Computers will haye the ability to handle these For example: the usual assumptions of independence data and can do so if we can tell them how." and/or normality crop up continually to restrict the (Bonner, 1962, p. 353). usefulness of the results obtained. "One is reminded (2) "Research in this area [information recogni- of Tukey's injunction to the effect that "what is tion] is needed to acquire knowledge of such needed is not optimal solutions to carefully formu- subjects as: the interrelationship between the lated, ideal problems, but merely 'good' solutions analytic and Gestalt aspects of visual-pattern for nonideal situations." (Hart, 1966, p. 11). recognition; how and what subsets of point An unfortunately not-well-coordinated thinking stimuli are perceived as unitary entities; the still persists in the areas of character, voice, and figure-figure, and figure-background separation othertypesofpatternrecognition.Thus,in mechanizms; and the meaning of the direction Kovalevsky's words: "Designers of reading devices and limitation of attention." (Office of Aero- and speech-recognition systems sometimes were space Research, 1965, p. 26). not interested in theory, and mathematicians disregarded applied problems. This resulted in a (3)"It has been shown, then, that any talk of an number of important tasks being left unsolved." ideal adaptive language teaching machine (1965, p. 42). must remain premature until a number of Another point to be observed is that, despite the serious problems have been solved. It is clear wide differences in the language used to describe that most of these problems are inherently particular recognition techniques, there is a far theoretical rather than technical: their solution greater commonality of approach among different is tied to the development of a complete and systems than might be suspected at first glance. Of satisfactory theory of language and of speech. particular promise for progress in automatic tech- The considerationof what is involved in niques for categorization, generally, therefore, are developing such a machine highlights the possibilities for increasing convergence of pattern complexity of language. Fundamental is the recognition and pattern detection techniques in fact that language allows a great deal of many fields.5.147 redundancy, but any specific speech event "Pattern analysis, as provided in the proposed contains only a selection of those features. system, can be used to discover relationships which Thus, speech recognition depends on using pertain to vast bodies of data which would be almost those features that may be present to decide impossible to obtain through any known thanual on those that are absent. It is, for instance, techniques.Intelligentuse of patternanalysis possible to understand speech in the presence should assist in evaluating new trends in crime, of considerable noise provided that syntactic testing new approaches to the administration of and semantic information is available to the criminaljustice,anddiscoveringpatternsof listener; but specification of exactly what this structureandactivityof criminal and special information is has proven to be extremely organizations." (Geddes et al., 1963, p. 266). difficult." (Spolsky, 1966, pp. 494-495).

6. Conclusions As we have seen in the preceding Section; of questions of linear separability, determinations of this report, research and development efforts in the membership in classes, and the like. Other problems broad area of automatic pattern recognition are of under attack include those of voice recognition, particular importance to the improvement of infor- automatic detectionof patternsof interestin mation sensing and input processes. This area electrocardiograph and bubble chamber data, and extends from successful techniques for magnetic simulation of observed human or animal perception ink and optical character recognition to research phenomena. investigations of potentialities for automatic detec- The scope of the field of pattern processing tion and classification of cloud formation patterns research is indeed surprisingly broad. The term and to theoretical developments with respect to "processing" is chosen deliberately in order to 50 beg terminological controversies with respect to (1) Research on format clues e.g., to pattern"identification,"pattern"classification," skipinterspersedgraphics,to and pattern "recognition," as variously used in the ignore page numbers and short literature." In the pattern-processing R & D area, titles,todifferentiatebetween we are concerned either with the identification of a period, decimal point, and the like, given input pattern, or with the meeting of member- priortomulti-format,multi-font ship-qualification requirements for a pre-established page reader design and applica- class of patterns, or with the properties or features tion." that a given input pattern shares or does not share (2) R & D efforts on thedifficult with some one or more classes of previously en- alphabets not likely to be tackled countered patterns; and with the determination of by commercial interests: Cyrillic, which properties or features of some set of patterns Arabic,Chinese,etc." Special significantly distinguish them from other possible problems to be investigated include sets or classes. the common European practice of Input problems for automatic analysis of text (to spreading out the letters of a word be considered ina later reportinthis series) or proper name in lieu of italiza- involve questions of multifont, exotic alphabet, and tion." multilevel character recognition and difficulties in (3) Developmentoferror-detecting, more advanced techniques for speech recognition error-correction techniques for in- and for pattern recognition generally. Thus, as put likely to be noisy or difficult, Wyllys points out: "When character readers reach e.g., RCA and NBS-Chinese con- that level of development [multi-font page reading text-checking operations. of a typical scientific journal], it will be possible to b. For conventionalized symbols, we need: utilize the positional and editorial features of printed (1) R & D investigations to provide for text e.g.,paragraphing, sub-heading,captions, size, translation, rmation and other italization, and capitalizationto furnish clues for symbol invariances. automatic abstracting and other purposes." (Wyllys, (2) Development of "picture-language 1963, p. 7). equivalence", e.g., how to name Pattern recognition as an area of R & D concern shapes that have been recognized; is directed to the identification or the detection of for example whether, given the arbitrary patterns, some of which can be visually name, we can find a corresponding perceived (such as photographs, charts, textual symbol or shape." patterns and three-dimensional real-world objects) (3)Research concerned with two-di- and others which require nonvisual perception such mensional well-formedness, e.g., for as speech and electromagnetic signals. Character structural and circuit diagrams, also recognitionisconsidered as a specific case of forcharacter-componentsasin pattern recognition and is often dealt with as a Chinese ideographs, and two-dimen- separatefieldof investigation because of the sional grammars. emphasis that has been placed upon it. Much of the effort to date has been limited to the development of 2. Pattern recognitionincluding speech, EKG, print readers which are usually limited to small EEG, star plates, cell tissue, microphotographs, alphabets, single font, and often with a specially regularitiesinsignal sequences, Stenotype designed type style. records; Morse code signals, maps, drawings, In addition, computer experiments have enabled gray scale images, photos, three-dimensional demonstrations and simulations of much more objects, and so on. sophisticated pattern recognition schemes including a. Research on limited multi-speaker speech at least elementary target pattern detection, electro- input, e.g., dictation of data from instru- cardiograph analysis, some automated photo inter- ment observations. pretation, cloud pattern detection, detection and b. Research on programming languages and tracing of bubble chamber tracks, limited speech compilers for manipulation of two and recognition, recognition of contrained handwritten three-dimensional graphic data. and some limited identification of words in cursive c. Research and development of automated handwriting. photointerpretation, specifically including Research requirements in the area of pattern photomicrographs. recognition were summarized for a Task Force of d. Research and development on acoustic COSATI in 1965 by Davis and Stevens, substantially and speech analysis, recognition, and as follows: synthesis." 1. CLracter recognitionincluding convention- e. Research on techniques of recognition alized shapes and symbols, as in chemical of three-dimensional objects, on deriva- literature, circuit diagrams, flow-charts, and tion of dimensional data, on reconstruc- the like. tion of two-dimensional representations of a. Fortheprintedcharacterreading three dimensional objects, on normaliza- systems: tion of perspective drawings. 51 r

3. Word and item patterns detection and recogni- 6. Research and development of self-organizing tion: systems for pattern recognition. a. Research on patterns of citedness, citing- 7. Research and development of techniques of ness and of overlap of indexing terms simulation of human behavioristic modes of previously assigned to co-cited items. pattern recognition including experiments in b. Research on patterns of word co-occur- automatic deductive and inductive inference, rence with sample assignments of cate- and concept formation. gory names, descriptions, etc., of prior term-term or word-term associations in For the Iong term, the combination of scanner the collection, and automatic indexing developments, languages for pictorial data process- based thereon. This includes devices such ing, and a "syntactic structure" and "context as ACORN for display of patterns auto- expectancy" approach will be important for appli- matically detected. cations such as patent searching where information c. Content analysis, e.g., of verbal protocol is found in drawings as well as in text." More records using both clue-word and tem- generally, such research points to the handling of poral patterns. a wide variety of information where the meaning or 4. Automatic categorization and determination cf message is conveyed as much by two-dimensional membership in classes: arrangements,orderings,and juxtapositionsof a. Research on mathematical theories of symbols as by the identification of the symbols determination of membership in classes, themselves. linear separability and threshold logics, Looking forward, in the area of automatic opera- factor analysis, and the like s-7 tions upon input, we can foresee increasing develop- b. Research in automatic classification and ment and use of techniques of automatic analysis indexing, where syntactic and semantics as applied to graphic, auditory, three-dimensional, are not involved and where it is equally and other types of source data inputs. This area applicabletoprop:_rtiesor attributes will thus include extensions and improvements to other than words or text. character and pattern recognitions methods and c. Research on detection of patterns of equipment, to developments and extensions of signals embedded in noise. voice recognition techniques, and to continuing 5. Research and development of adaptive systems researchintheareasof patternrecognition, and techniques for pattern recognition. detection, and identification.

52 Appendix A. Background Noteson Research and Development Requirements in Information Acquisition, Sensing, and Input

In this Appendix we present further discussion and background material intendedto highlight and emphasize currently identifiable research and development requirements in the broad field of the computer and information sciences. A number of illustrative examples, pertinent quotations from the literature, and reference:.- current R & D efforts has been assembled. These background notes have been referenceci, es appropriate, in the text of this report.

1. Introduction 1.1In a provocative and challenging discussion He says further that "thesenew problems, and of a research team's findings with respect toa the future of the world dependson many of them," desirablemechanized documentation system of will require science to make a third great advance the future, Licklider (1965, p. 40) states: "thereason that "must be even greater than the nineteenth- for setting forth such a plan is not to guide research century conquest of the problems of simplicity or ., and development, which would be presumptuous, the twentieth-century victoryover problems of but to provide a kind of checklist or scorecard for disorganized complexity." use in following the game. If the technology should 1.3Some examples are as follows: "One field take care of most of the items in the plan but fall of science now moving forward ata rapid pace can behind on a few, then it might be worth while for be singled out as offering unusual scientific and an agency interested in the outcome to foster special economic challenge. This is the broad area of infor- efforts on the delinquent items." mation manipulation. Included are the broad areas 1.2"Information processing systems are but of information theory, computation and communica- one facet of an evolving field of intellectual activity tion. To the basic scientist there is presented the called communication sciences. This isa generic question as fundamental as the nature of light or term which is applied to those areas of study in gravitation, namely, what is information; what is which the interest centers on the properties ofa knowledge; what constitutes consciousness of exist- system or the properties of arrays of symbols which ence; are there immutable laws of the physical come from their organization or structure rather universe governing the nature of information and than from their physical properties; that is, the study knowledge as yet undefined and perhapsunsus- of what one M. I. T. colleague calls 'the problems of pected that await discovery in the next 5 decades?" organized complexity'." (Wiesner, 1958, p. 268). (Bowie, 1962, pp. 609-610). Theterminologyapparentlyoriginatedwith "One difficulty in the endeavor to match mechani- Warren Weaver (1948). Weaver noted first that the cal to human behavior...is that it is far from areas typically tackled in scientific research and clear what sort of activity counts as recognition of development efforts up to the twentieth century a pattern by a human being. It is not clear conse- were largely concerned with two - variable problems quently what sort of behavior is to be duplicated or of simplicity; then from about 1900on, powerful approximated by mechanical devices. techniques such as those of probability theory and "Although some investigators seem aware of this statistical mechanics were developed to deal with problem, there is little evidence of any sustained problems of disorganized complexity (that is, those attempt to sharpen our understanding of human I. in which the number of variables isvery large, the pattern recognition. The deficiency at point is not individual behavior of each of themany variables the admitted lack of a comprehensive psychological is erratic or unknown, but the systemas a whole theory of recognition, but rather the lack of a clear has analyzable average properties).Finally, he description of what humans do in various situations points to an intermediate region "which science which normally would be taken as instances of has as yet little exploredor conquered, where by pattern recognition." (Sayre, 1962, p. 27). contrast to those disorganized or random situations "The next fifty years will see electronic informa- with which the statistical techniquescats cope, the tion processing systems brought to a stage of so- problems of organized complexity require dealing phistication comparable to that of today's electronic simultaneously with a considerable number of communication systems. This will require two variables that are interrelated in accordance with things. In the first place, the basic principles under- organizational factors." lying the brain's method of digesting information

53 11 must be discovered and translated into simple (i.e., sets of words which are synonymous or related electronic systems. This is as necessary as our in meaning), each of which stands for one descrip- simple theory of electromagnetism was for develop- tor; the result will be a set of descriptors forming ing detector, amplifier, and transmission systems. the search pattern of the text in the IRS concerned. In the second place, capacity for electronic storing The difficulty here is to distinguish the homographs, and processing of information, on a scale equal to but the experience has shown that this difficulty is that of the human brain, must be developed. The well resolved by lexical analysis of the context electronic information handling systems of the environmentofthehomographs."(Mikhailov, future will be able to receive information in visual 1968, p. 35). and auditory form, just as humans do." (Van 1.4b"Laboratory of Electromodeling (LEM) of Heerden, Proc. IRE 50, 621 (May 1962).) the All-Union Institute of Scientific and Technical "Concern with meaning introduces a parameter Information (VINITI) of the Academy of Sciences of ignored in the simplifying assumptions of communi- the USSR. Development of a reading automation was cation theory. Such Air Force problems as automatic begun here in 1959. The operational reader... abstracting and indexing of scientific papers, infor- reads typewritten texts and inputs them to a mation storage and retrieval, mechanical translation, Ural-4 computer." (Stephan, 1967, p. 5). and even automatic distribution of messages depend "An operational automation for reading type- upon a deeper understanding of the meaning of writtentexts and inputting them to a Ural-4 meaning than has heretofore been necessary or computer to solve a number of problems of an possible. This understanding may even come from information processing nature has been developed suchesotericfieldsasanalyticalphilosophy, in the Laboratory for Electromodeling of VINITI, epistemology, and taxonomy." (Wooster, 1961, p. under the Academy of Sciences of the USSR. . 17). "Pattern recognition in the operational reader is 1.4Goldberg, in his 1931 U.S. Patent, disclosed based on a method...for comparing the pattern this principle as applied to statistical operations under investigation against a set of fixed image-field such as selecting and counting particular records regions which yield invariant descriptions of patterns identified by some specified combination of indicia, in binary or any other coding system. The informa- forexample,variousalphabeticand numeric tion distinguishing areas were called distinguishing symbols. In a suggested embodiment he visualized fragments and the method, correspondingly, the the records as being stored on a positive photo- fragment method... graphic transparency and a 'search plate' with the "The technical implementation of the fragment negative images of the selection criteria processed method, developed in the Laboratory for Electro- modeling, has made it possible to devise an analysis a_ s follows: unit by extremely simple means. The unit for storing "If the transparency containing the various sta- the fixed symbol variants (that is, the fragments) tistical indications is now run through the machine is a ferrite matrix in which the fragments are in- in such a manner that the negative coincides with scribedinthe formof output-threadedcore the transparency a complete coincidence impene- wirings...The device turns out to be small, trable to any light or heat radiation will only be simple, and reliable in operation... possible in one defined case; this will only occur "A coded description of the character being when the negative bears exactly the same charac- scanned, selected from the matrix of fragments, is ters, marks, figures, etc., as the transparency in fed into the decoder which contains code descrip- question, the only difference being that in the tions of all characters of the set being examined and negative these records are light on dark ground, previously obtained as a result of statistical process- while in the transparency they are dark on light ing of large masses of observed characters in an ground. A certain combination has thus been picked actual reading device and digital computer system. out of a large number of others with extra-ordinary The encoded description of the symbol is compared speed and reliability hitherto not obtainable. In with all the encoded descriptions of symbols in the order to obtain the coincidence of the negative with decoder, and a minimal signal is obtained on the the transparency they can be either brought into decoder bus for the character most closely corre- contact (direct superposition) or be projected one sponding to the encoded character. The index of this upon the other (optical superposition); the latter bus is transmitted to the output unit, where this method being more advantageous as the mechanical index, with the aid of an encoder device, is trans- features of the machine are simplified." formed into Ural-4 machine code and is then relayed 1.4aFor example, "for automatic input of docu- to the computer's memory and onto papertape (for ments to an automatic IRS, the descriptor type monitoring purposes)... language without grammar is used. Chernyaysky "The reader's speed is determined by an electro- and Lakhuti...apply the first approach rigor- mechanical system of character scanning and the ously; the text of a document (a short abstract) is operationalcharacteristicsof theferrite-diode compared word-by-word with a descriptor dictionary modules; at present, the rate is 10 characters/sec." prepared in advance (the subject being that of (Avrukh, 1967, pp. 20-23). Electrical Engineering), in which the significant "Three cabinets painted in traditional computer- words are grouped together in equalizing classes grey stand in a small room of the Laboratory of 54 Electromodeling...thisisthe reading auto- people, analog and digitalcomputers, and a large maton. variety of input-output devices." (Ware, 1%5, "In the first cabinet there isa page of typewritten p. 472.) Ohlman, looking forward in 1963, suggested text on a revolving drum. A greenish band of light that "another interesting possibility is audiointer- from a cathode-ray tube shines downon a line. To rogation and reply, involving telephone lines and the side is a photomultiplier whoseeye is focussed tape recorders...Any combination of these modes on the line running under the band of light. Scanning is possiblea store might be interrogated orally and the letters, the ray of light is constantly transmitting a reply obtained visually, etc." (Ohlman, 1963, signals to the photomultiplier: 'white, white, black, p. 193). white...' These signals transformed into electrical 1.7"In 1936, Dudley [1936] invented the vocoder, pulses enter the second cabinet which is tightly a device for compressing the bandwidth of speech packed with an intricate ligature of varicolored signals in order to transmit themover channels conductors connecting ferrite and semiconductor of very limited capacity. The vocodermeasures the elements. speech power in a number of frequency bands "There are frames covered with an openwork and transmits thesemeasures as signals over a screen in the cabinet and there are miniature ferrite series of narrow low-frequency channels (Fant and cores at the intersections of small wires. In several Stevens, 1960). At the receiver the speech isre- places the screen has been 'darned' by additional constituted by modulating the spectrum ofa broad intersections of small wires. The electrical pulses band sourceitsaccordance with the frequency from the first cabinet are stored and analyzed in region and ampli:=Ide of each of themeasure- this screen-matrix. The automation has features signals derived fron the original speech. Normally stored in it by which it is possible to describeany thisreconstitutedspeechsignalispresented typewritten alphabetic character. Oneor another acoustically for a listener. Alternatively, vocoded combination of features allows the machine to speechsignalscanbedisplayedvisuallyor recognize the letter being read. tactually." (Picket, 1963,p. 2.) "The complex cobweb of electronic circuits in "The first compater applicationwas probably in the third cabinet examines the results obtained by the design of the so-called `vocoders'. The vocoder the recognition matrices in the second cabinet. is a telephone system that utilizes long distance Here the final determination is madeas to which transmission lineswith greater efficiency than letter has been read. From here the lettersare fed conventional systems. At the sending end, thevo- by cable into the memory of a Ural-4." (Koltovoj, coder extracts and transmits only certain features 1967, pp. 38-39). of the speech wave pickedup by its microphone. An actual example is shown in Figure 11. It will be At the receiving end, ituses the transmitted infor- noted that the font uses large-sized and well-spaced mation to reconstruct a simplified version of the upper-case alphabetic plus numeric characters. original speech wave. The transmissioneconomy is Stylization has been applied tonarrow some char- achieved by not transmitting the whole soundwave acters that are normally wider than others, to add a -but only those parts of it that contribute to speech small connecting bar to the character 111' andto perception. As we co not fully understand the accentuate punctuation marks. A relatively poor essentials of the speechprocess, we are not always print quality may also be noted, leading to obvious sure which features to extract or how to extract variations of stroke thickness and to problems of them. A wide variety of circuits has therefore been correct discrimination between members of some tried in attempts to achieve acceptable vocoders." character subsets. To date,over 2,000,000 char- (Denes, 1966, p. 250). acters have been read by optical character recogni- 1.8They report: "Amore promising 'measure- tion equipment and fed to a computer for experiment space' for speech recognition is suggested mental author and title index preparation. by a consideration ofa theoretical model for the 1.5For example, in work at the William Alanson generation of synthetic speech sounds." (King and White Institute by Jaffe and his colleagues. "Jaffe Tunis, 1966, p. 66). also indicates the necessity for analyzing nonverbal 1.9"The equipment consists ofa voice spectrum as well as verbal behavior. For this purpose he and analyzer acting as online inputto an IBM 1620-11 his coworkers developedan automated device for digital computer system. The basis of thespectrum the extraction of time-sequence behavior in the analyzer is a contiguously tuned bank of bandpass transaction between client and therapist (Cassotta, filters whose instantaneous outputsare continu- Feldstein and Jaffe, 1962)...the new device per- ously compared in sucha way as to locate the mits the recording of time-sequence data (not actual instantaneous peaks in the envelope of the speech content) from microphone inputs to punched cards spectrum. The output of the spectrum analyzer is without manual intervention." (Ford, 1963,p. 5). a binary coded representation of the peaks of the 1.6See account of the Chrono-Log Model 3,000 envelope of the frequency spectrumas a function Oscillographic Time Code Generator, Data Proc. of time; it serves as input to the IBM 1620." (King Mag. 7, No. 2, 54 (Feb. 1965). and Tunis, 1966, p. 65). 1.6aThus, "this is the era of the interface, See also Section 5.5 of thisreport. especially the custom interface. The industry is 1.10In KWIC indexing, the texts of titlesare willing to couple anything to everything including prepared for input to the computer and thistext 55 5 rI48 0 Il,k1OPOBORTEXHIAKECIIIITUBAH1/1f10 AHAII43A KOHTYPOB/IIOJIHICOB B.r./ 5 risi PACHO3HABAHItIE PYKOHIACHNX 3HAKOB C HOM0141310 CJIEAfilliERPASBEPTICl/IfCEMEHOBCKIiiitif 8 rI83 II tINTAIOIIIEE YCTP02CTBO/HAPIIITA AK1f1PA 8 1'176 K BOHPOCY340:DEKTI/IBHOCTI/1 HAPAMETPOB OBPA30B/TAJIIIKCHMC A.A./

8 1'175 IC BOIIPOCY01103HABAH1/1}1 OBPA3OB ME TOA014 ILACOK/EYAOBAC B.B./

G 148 About Numerical Technical Computations and Analysis of Contour s/Polyakov, V. G. G 151 Recognition of Handwritten Symbols With the Aid of Curve Following/Semenovsky G 183 Reading Device/Narita Akira G 176 The Problems of Sensitivity to Parameters of Shapc..-;/Tale3mis, L. A. G 175 The Question of Recognizing Shapes by a Method of Masks (?)/Bulovas, V. V.

FIGURE 1].Material read by Soviet print reader

A I 3 4 is then prut.-:essed against thestop list in order to canoe is difficult to predict, it is more practicable to eliminate from further processing themore common isolate it by rejecting all obviously nonsignificant words, such as "the," "and," prepositions, andthe or 'common' words, with the risk of admitting like, ss well as wordsso general as to be useless certain words of questionable value. Such words for indexing purposes, suchas "introduction" or may subsequently be eliminated or tolerated as "progress" or "preliminary," which will, however, `noise'. A list of nonsignificant words would include be printed out in the titleas it appears in the final articles, conjunctions, prepositions, auxiliary verbs, index. certain adjectives, and words suchas 'report', In the original proposals of the late H. P. Luhn `analysis','theory', and the like." (Luhn, 1960, for KWIC indexes, he stressed that "since signifi- p. 289).

2. Information Acquisition, Source Data Automation,and Remote Data Collection

2.1"A new system from 3M allows EKG's to be `The primary measurementsare processed by a transmitted and viewed at long distances in seconds. remote computer, and the measured and derived Heart data is transmitted via telephoneor trans- variables are printed outon an output typewriter mission lines to a Model 1260 recorder. The images in the ward, and displayedon a screen at the are displayed on the unit's CRT where they can be patient's bedside'." (Davis, 1966,p. 248). viewed at once or photographed bya microfilm "It appears there is a need for hardware assist- processor camera." (Data Proc. Mag. 9, No. 6, ance in caring for the critically ill in order to reduce 10 (1967).) human error and to provide permanent records and "An FM data transmission systemwas designed displays which the humansenses are not trained to to transmit a fetal electrocardiogram utilizing the assimilate entirely. This would provide for better normal telephone circuits." (Hagan and Larks, care of patients as well as production of hard-copy 1963, p. 147.) records." (Brown et al., 1967,p. 77.) "The major work in the medical field in 1965 took place in three areas: machine-aided record keeping, "Valid statistics on physiologicalprocesses re- machine-aided physiological analysis, and machine- quire substantially more data thancan be manually aided medical communication. Vallbona, Caceres, recorded. As a result,many separate biomedical and others are trying touse detailed physiological instruments, as well as integrated physiological data derived either directly fromsensor or indirectly monitors, are being built to collect useful data. An from sensor tracings as suitable input toa medical ultimate, but difficult, objective is theuse of real- computer system. Caceres utilizes a recording cart time data-acquisition systems to provide patient or remote telephone link to go directly from the EKG data that can be analyzed for immediateuse by the leads to the computer, while Vallbona telemeters attending physician." (Aron, 1967,p. 65.) the physiological data to apaper tape punch from 2.3"There is need for the monitoring of patients which cards are prepared thatare then fed into for intra-hospital use, but also for networkuse in the computer." (Baruch, 1966, p. 258.) terms of exchange of information (withother 2.2"A computer system using an IBM 1620- hospitals) on techniques, etc." (Brown et al., 1967, 1710 for high speed monitoring and real timereport- p. 77.) ing on critically ill patients...[has been developed 2.4"In a broader sense, telephone telemetry by the Shock Research Unit, USC School of may help to interconnect research laboratories and Medicine at the Los Angeles County Hospital.] even universities, making possible that degree of The system currently accepts analog signals from inter-university cooperation which seems indicated 11 transducers on the patient's body...Addi- in this epoch. Several of Marquette's neighboring tional information...is obtained by manually universities have expressed great interest in this entered data." (Rockwell et al., 1966,p. 357.) field[telephone telemetry of low-level biologic "Greanias [1965] points out that 'primary data signals] and a regional net, including at least four from the patient himself may be processed by the states, is planned." (Hagan and Larks, 1963, p. 150.) computer...on-line....The on-line data 2.5"In an experiment for the Chevrolet division such as electrocardiographs and respirationrates of the General Motors Corp., the Boeing Co., used can be automatically monitored by the computer computer graphics last year to run off sketches of and directly fed into storage without human inter- Chevy'sSting Ray...Proximityprobesre- vention.' He mentionsone of the earliest projects portedly were used tomeasure the length, width to develop an on-line system for direct patient and height of a clay mockup. These coordinates monitoring, being carried out in the Shock Research were then punched into cards and fed into a com- Unit of the University of Southern California's puter. Punched tape was made and used torun a School of Medicine and the Los Angeles County numerically controlled drafting machine,or engi- Hospital under the direction of Dr. Max H. Weil. neering plotter..." (Gomolak, 1964,p. 65). 57 2.6"The General Electric Co., reports...that only onceUse it, and use it, anduse it'." (Kroger, it has provided the Ford Motor Co., withan experi- 1965, p. 267). mental NC system for making automobileparts. "It is hard to conceive of a great expansion in A model of the part isphotographed and an management information systems unless there is a electronic system converts the photos into anumer- breakthrough in source data automation." (Voltin, ical control tape, which controls the machining of 1967. p. 8). the part." (Electronics 38, No. 20, 25 (1965). 2.7a"The Bureau has developeda new model of Other examples are as follows: "In a steel fabri- FOSDIC (Film Optical Sensing Device forInput cation detailing application, the computer accepts to Computers) to assist the Bureau of the Census basic design criteria and determines the exact in processing the data to be received fromthe 1960 dimensions of steel members, how they should be decennialcensuses.Thiselectronicmachine cut, and the size of the connecting members. The rapidly reads microfilmedcensu., documents and number of bolts and lengths of welds to be usedto transcribes the dataon magnetic tape for direct interconnect members are determined. input to an electronic computer. Itwas designed by "The CONSTRUCTS (Control Data Structural M. L. Greenough and C. C.Gordon of the NBS System) seeks out possible obstructions andcon- staff and E. S. Stein of the Bureau ofthe Census. nectionproblems, andautomaticallyprovides "The central element of FOSDIC III, likeits special solutions as required. predecessors, FOSDIC I and II, isan electronic "Once calculations have been completed, full- scanning assembly. Light from thescreen of a sized drawings are producedon a plotter." (Bus. cathode-ray tube is focusedupon the microfilm Automation 12, No. 7, 55 (1965). image, and the transmissivity of small, discrete "An advanced concept in the calculation and areas on the film corresponding to the hand- transmission of machinability data was demon- written marks on the original documentismeas- strated by GE's Metallurgical Products Dept. ured with a photoelectric cell. Bymoving the elec- Machinability problems were posed toa remote tron beam around on the face of the cathode-ray GE-225 computer via a Datanet-760 terminal. tube, any selectedarea of the image can be ex- Near-instantaneousanswersonrecommended amined. Control of the position of the illuminated feed, speed and carbide tool grade for turningopera- area, and interpretation of the signals from the tions were displayed on a TV monitor." (Data Proc. photoelectric cellsare functions performed by the Mag. 7, No. 11, 12 (Nov. 1965).) associated electronic circuitry. Through- positional "An integrated computerized system has been control, or scanning, the point of light travelsin developed at he Sun Oil Company for the produc- prescribed manner fromone point to the next in tion of isometric drawings with accompanying bills turn. Since there is only one electron beam in the of material, requisitions, purchase orders, and cost cathode-ray tube, the scanningover the image is estimates for. .piping configurations....A signifi- serial in its timesequence. Then, as soon as one cant feature of the operation is a single input sheet image is scanned, the film is automaticallyad- initiating the process whether it be for the entire vanced to the next frame. procedure or only part of it. These procedurescan "FOSDIC can read offany part' 'of all of the in- produce fabrication sketches for piping fabricators formation on a document, dependingon what the and/or construction site personnel. They will yield programer instructs the machine to do. The instruc- cost estimates for cost engineers as an aid in the tions are in terms of the horizontal and vertical evaluation of material, line sizeor configuration coordinates of the index marknext to the desired changes for a single drawing or any combination information. Program control is through a plugboard of drawings describing a section of the plantor into which the detection logic has been ,wired." the entire plant. Requisitionsare provided for ( FOSDIC III to Assist in 1960 Census, 1959). ordering material and updated lists of the status "A key factor is...FOSDIC, or more properly, of items ordered are produced for project engineers. Film Optical Sensing Device for inputto Computers. This equipment transforms microfilmpictures of The purchasing department by writinga cover special paper forms into impulses letter can use the requisition asan appendix and, on magnetic tape ...One FOSDIC sheet takes the place ofa dozen thus, produce a purchase order. The material punched cards and FOSDICcan process 100 of inventory control group at one of our refineriescan these sheets, in the form of microfilmimages, in use output cards from this system to interrogate one minute." (Census Bureau Slashes Record their computer as to the inventory status of items Handling Time, 1964,p. 32). being requisitioned. In an integrated system such A related development, FOSDICIV, is designed as this, justification and advantages accrue from the for programmable scanning of the microfilmed. integration over and above any of the singlecom- images of punched cards: "An improved modelof ponents. In this paper the discussion will be limited FOSDIC...has been completed by the National to the isometric drawings and bills of material Bureau of Standards foruse with the computers phase of this system." (Richer, 1967, p. 134). of the National Weather Records Center... 2.7"In commercial data processing systems FOSDIC IV reads dataon past weather conditions where there is a strong profit discipline, the cardinal from microfilms of punched cards from the Center's rule is 'Capture data at itssource Capture it archives. The machine performs logical operations 58 on the data it reads and also selects certain data to "The Leo-Parnall Autolector isan automatic be recorded on magnetic tape for later input to optical scanning device which reads both computer digital computers." (FOSDIC IV Reads Micro- printed and hand marked forms directly intoa filmed Weather Data for Computer, 1967,p. 63). computer." (Dig. Comp. Newsletter 16, No. 4, 2.7b"With advancing technology, input from 26 (1964). authors may be in machine-readable formor may 2.12This technique "changes electrical impulses be converted directly to this form byscanners. into numerical values. These numbers are fed intoa Electronic display and light-pen techniquesseem computer which then examines the series of num- likely to play important roles in composition and in bers to produce a standard pattern for each motion. making editorial corrections and also in the trans- The pattern can then be used to design control mission of stored information to users." (American circuitry for complex movement patterns." (Bus. Institute of Physics Staff, 1967, p. 372). Automation 70, 12, No. 10, (Oct. 1965).) 2.8"A record from a relatively simple input 2.13"Techniques are being developed in several typing operation can be processed to form a multi- laboratories which will permit theuse of optical plicity of high typographic quality, complex output character recognition in industrial control applica- products...This procedure allows the primary tions, for example checking labelson bottles during record to be processed into a variety of secondary packagingprocessesandmaterialcontrolon products, regardless of their typographic forms. A assembly lines." (Daniels, 1967,p. 10). computer processing system has been developed to 2.13a"Kartrak ..utilises an unmanned, produce the primary journal typesetting tape, an weather-proof, trackside scanner which sends out author index typesetting tape, and an abstract beams of white light and receives coloured reflected journal entry typesetting tape." (Buck land, 1965, light from strips of durable reflective material. pp. 2-20). The strips, attached to each car, are arranged to 2.8a"Miller Electric co. uses Control Data 180 form a colour code representing the identification data collectors on the shop floor of its plant in number of the car and its weight when empty." Appleton, Wis., to capture data in punched tape. (Data and Control 2, No. 12, 9 (Dec. 1964).) The collectors are portable and have self-contained "Automatic car identification adopted nation- punches. Data on production status and attendance wide by America's Railroads. Thomas M. Good- is entered into the devices by the people who work fellow, president of the Association of American on the welders and welding equipment which Miller Railroads, has announced the adoption of industry- sells. Plastic identification cards are used to input wide automatic car identification system to monitor employee identification. Pre-punched cards identify the 1.8 million freight car fleet. The KarTrak auto- each part being manufactured, dials are used matic car identification system, selected by the by the employee to enter variable information, and AAR, is designed and manufactured in Bedford, a clock module automatically records time of input Mass., by the Commerical Electronics Division of on the tape." (Menkhaus, 1967, pp. 33-34). Sylvania Electric Products Inc. "Major components in thenew system are a 2.9"Automatic data collection (ADC) implies trackside electronic 'scanner' made by Sylvania the recording, in machine-readable form, of the andthe 3M Company's`Scotchlite'reflective pertinent data about a transaction at the time the sheeting. Strips of reflective sheeting are coded by transaction occurs. Some data collection systems color and design in such a manneras to represent collect and record the transation data in machine- numbers to the scanner. A light beam from the readable form for later batch processing, while scanner 'reads' numbers from bottom to top. The others feed the data directly into real-time computer retro-reflective sheeting 'bounces back' numbers systems to provide up-to-the-minute information for to the scanner which feeds them into a centralized operationaldecisions."(Hillegassand Melick, computer. The system works at train speeds of up 1967, p. 50). to 80 MPH' and is unaffected by weather conditions. 2.10"Honeywell's modular, multipurpose Data The system makes possible instantaneous location Station can optically scan documents, read punched of any freight car in the country. cards, print at high or low speeds, and punch and "The TeleRail Automated Information Network read paper tape and is designed as a remote terminal (TRAIN) recently established by the AAR will be for Honeywell 200 and 2200 computer systems... tied in with the nationwide ACI system and the The optical scanning unit reads up to 50 documents advanced information systems of individual rail- a minute imprinted in a special bar code." (Data roads. TRAIN, based at the Association's Washing- Proc. Mag. 7, No. 2, 47,1965). ton headquarters with computer links to all Class I 2.11"Hewlett-PackardCo.,RedwoodCity, railroads, will provide the AAR Car Service Division Calif., unveiled its desk-top optical mark reader ..." with complete reports of car locations by railroads (Commun. ACM 9, No. 6, 468 (June 1966).) and car flows through principal rail gateways, and British examples of optical mark reading equip- will permit more rapid distribution of the equipment include an ICT (International Computers and ment to meet shipper needs for freight cars." Tabulators, Ltd.) device and the "Autolector" of (Computers and Automation 16, No. 12, 61 (Dec. English Electric, Leo. 1967).

366-107 0 - 70 - 5 59 2.14Davenport (1965) summarizedsome of the scan, 1(4 and alarm, spill calculation, and operator's advantages of this techniqueas follows: information programs." (Computers & Automation "(1) Cost is relatively- low, only twoor three times 14, No. 7, 38-39 (1965).) that of an ordinary telephone. 2.17 "A new data-message communications (2) The remote input station can be used filter- systemisnow being marketed nationally by changeably for either voice or data trans- Digitronics Corporation...The 600 Dial-O-Verter, mission. collects information from many remote sources and (3) The keyboard is as easy touse as a small automatically separates the computer-bound data adding machine. from the message traffic." (Commun. ACM 9, 782 (4) The station can be connected to both local (1966).) and long- distance telephone facilities. "Digitronics is currently marketing three distinct (5) The input device can be portable and capable product lines, each of which represented roughly of sending from any phone jack. one-third of the company's bookings or about $8 (6) Transmission can be either topaper tape million last year. These three lines are: and card punches or on-line toa computer. TheDial-o-verterlineofcommunications (7) Effectuallyreal-timeoperationcanbe terminals; achieved at low cost." The Data-verter line of data acquisition and 2.15"The data that is accumulated and stored is transmission devices; and information relevant to flow of work and accumulated A component line consisting of paper tape charges in a production facility. The information is readers and spoolers. entered by workmen at numerous data collection devices in the time keeping and productionareas. "The Dial-o-verter line is oriented toward the The production facilities are physically separatedtransmission of large volumes of data over the public by several hundred miles from each other and telephone network or a half-duplex leased voice- corporate offices but because of the competitive band line. The Data-verter line is designed for nature of the business it is important to providecomparatively lowcost, low-speed collection and status reports three times daily. The amount oftransmission of data using specialized components. data accumulated at each facility duringan eight- The devices in the Digitronics component line are hour shift, including material purchase requisitions, sold primarilytoother manufacturers of data distribution of hours worked by product and stock processing equipment." (Axner, 1967, p. 45). status, adds up to 100,000 characters of information. 2.18See Data Proc. Mag. 7, No. 2, 41 (Feb. This data is accumulated during each eight -hour 1965). Further, "the IBM Hardware Monitor isan shift and then transmitted from each facilityone example of a general data acquisition system seek- at a time. Actual call time for each facility at 200 ing to capture a continuous stream of events for characters per second is nine minutes. The five off-line analysis." (Estrin et al., 1967,p. 647.) facilities can be polled in less thanone hour to 2.18a"Data analysis is preceded by dataac- provide a computer compatible tape tostart a com- quisition. During analysis, intermediate resultsare puter report cycle." (Hickey, 1966, p. 177). the basis for the proper choice of strategies, during 2.15a We note with Clippinger that: "There isa acquisition, for the choice of experimental condi- steady trend to move the computer interface with tions. Proper conditions for an experimentmay be data collection closer and closer to thesource. short-lived, hence the need for interaction must be The use of interrupt processing makes it feasible recognized while the experiment is still in progress." to process these data as collected, to keep files (Lockemann and Knutsen, 1967,p. 758). updated, using a negligible amount ofcomputer "The PACE (Precision Analytical Computing time." (1965, p. 209). Equipment) System provides one solution to the 2.15bMoreover, itis noted that "if current analytical data processing needs of industry. PACE trends Continue, the system and circuit designer is a computer-based system which automatically will have to solve manynew problems associated and simultaneously processes the electrical signals with an extremely low-temperature environment. generated by a variety of analytical instruments. This environmentoccurs in space but there are Therefore, the system can be put on-line with: (1) many terrestrial applications for devices requiring chromatographs; (2) massspectrometers; (3) lowtemperaturesmasers,lasers,parametric spectrophotometers; (4) NMR spectrometers; and amplifiers and infrared detection devices." (Allen (5) physical testing machines....The data reduc- and Niehenke, 1965,p. 75.) tion program examines, classifies and reduces the 2.16"The electric utility industry's first inte `raw' input data from each instrument into a com- gratedcomputer/telemetrysystem...atth pact form. The condensed data contains just the Public Utility District No. 1 of Chelan County, information needed to characterize the particular located in Wenatchee, Wash....The telemetry instrument." (Lichtenstein, 1966, p. 10). and data collection system is under directcomputer "A digital data recording system that automa- control...Important on-line programs include tically converts signals from laboratory instruments load frequency control, actual and estimateden- into computer language is available from Beckman croachment, telemetry control, pondageaccounting, :instruments, inc., Fullerton, Calif. 60 "The versatile system readily translates output crinic Unit (TU) and Arithmetic Unit (AU) to from a variety of atomic absorption, ultraviolet an provide summary information on photographs and infrared spectrophotometers into digital form. Out- complete their analysis. The PAU is of cellular put can be visually displayed or recorded punched design and reduces input images to idealized line paper tape, magnetic tape or standard IBM cards. drawings or graphs and has been designed in some Identification data may be introduced and recorded detail. The TU and AU are computers of more by the setting of 10 thumbwheel, parameter switches. conventional design a-id have not been described "The system can be economically expanded as in detail. We shall describe the PAU as a device users' needs increase. Eight operating models are having properties similar to CAM.... available. Up to three encoders may be used simul- "The PAU containsa network of identical taneously, permitting wide latitude in the handling processing modules organized on a 32 by 32 grid. and calculating of analytical data. A one-bit principal register (PR) with each module "Capabilities range from the simplest digital is capable of accepting the contents of any of its recording of ordinate values on manual command, to automatic and simultaneous recording of ordinate eight nearest neighbors. The set of PR's forms a and abscissa data while detecting and recording two-dimensional shift register whose contents may spectral peaks and valleys." (Computers and Auto- be shifted in any of eight directions. Associated mation 16, No. 11, 60 (Nov. 1967).) with each PR is a "stalactite" or multibit memory 2.19 McGee and Petersen, 1965, p. 90. The register which communicates only with the PR. immediate objective in their research has been the Logicaltransformations may be performed in automatic measurement of bubble chamber tracks parallel on contents of stalactites under central previously recorded on film. control, the transformations being identical for all "PEPR is a computer-controlled CRT scanner stalactites. Local logical transformations involving used to automatically measure bubble chamber points in a discretized binary pattern and their tracks which have been recorded on film. The PEPR nearest neighbors are performed by first shifting cathode ray tube defocuses the electron beam into a neighbors in each PR, shifting them down into short line segment whose angular orientation and stalactites, performing transformations on stalactite location can be independently controlled by the contents and finally storing results in the PR." system ... (Fuller, 1963, p. 25). . "Control of any CRT scanner would be similar to 2.21"ThePhysicsDepartmentof Rutgers the PEPR scanner. In other data collection systems, University will use a PDP-6 computer for ele- counters, pulse height analyzers, telemetry con- mentary particle analysis. The Rutgers System verters, etc., may all serve as sources of inpv. t data. will evaluate the photographic record of an experi- The controller output registers may be used for a ment on four film-scanning machines. Two will variety of external control purposes as well as data search for useful tracks and two will define their sources for display and printingdevices." (McGee locations by positioning crosshairs along and at and Petersen, 1965, pp. 82-90). the ends and vertices of a track in each of three 2.20"Theon-lineman-computer systemis views." (News and Notices, Commun. ACM 7, employed in the measurement on pictures of nuclear 557 (1964).) particle events obtained from high-energy accele- ratorexperiments...An operator guides an 2.21a"The work...is based on the use of a ingenious light sensitive measuring device, SMP mechanical flying spot digitizer to automate the (Scanning-Measuring Projector), across an image initial scanning and measurement steps. The in- of the film projected onto a flat table. The SMP creasing volume of bubble chamber film being device transmits a stream of measured data to the produced at many high energy physics laboratories CSX-1 computer for validation and analysis... around the world makes some form of automation Through a question and answer procedure con- essential. Two experiments currently in progress at trolled by the computer, ancillary data are fed con -. Brookhaven National Laboratory and Columbia cerning the nature of the vent, the film number, the University...[use] the flying spot digitizer... on- operator number, etc. line to an IBM 7094 computer." (Rabinowitz, 1968, p. 159). ...A calibration procedure is performed to measure fiduciary marks on the film... 2.21b"There are specific instances of produc- ...The measurement of three separate stereo- tivity improvement by automating a normally human scopic views of the same event is conducted." graphical interpretation procedure. One example is (Brown, 1965, p. 82. [U. Illinois] PIP, a spark chamber photograph interpreter. Scan- "A second pattern recognition computer using a ning and interpretation of 5000 frames per hour was cellular organization has been proposed by McCorm- possible, as compared to 50 frames per man hour." ick and coworkers for particular use in analyzing (Wigington, 1966, p. 88). photographs of particle tracks obtained from spark 2.21c"PEPR (Precision Encoding and Pattern chambers and bubble chambers used inhigh Recognition') is a device proposed in 1961 by Pless energy physics. The pattern recognition computer and Rosenson for the scanning and measuring of requires a "Pattern Articulation Unit" (PAU) for photographs of events in bubble chambers. It uses a preliminary processing of photographs and a Taxi- high precision cathode ray tube connected on-line 61 to a computer." (Watts, 1968, p. 207).[Reference is that otherwise would be too big for him tohandle." to an internal M.I.T. report dated Feb.1%11 (Cooper, 1964, pp. 20-21). 2.22 A computer-controlled x-ray diffractometer 2.25"GasChrom-8,anew computer-based (CCXD) system developed and operated by the system for analysis of data frommultiple gas Thomas J. Watson Research Center (IBM) provides chromatographs, has been announced by Digital structuralinformation oncrystallinematerials. EquipmentCorp.,Maynard, Mass. The new (Cole and Okaya, 1965, p. 32): "This advance in the system includes the company's latestgeneral pur- speed of crystallographic analysis was achieved by pose computer(PDP-8/I), a specially designed making all necessary x-ray measurements auto- chrornai..graph interface, easy-to-use conversational matically with an experimental system, developed software and Teletype. by scientists at the IBM Thomas J. Watson Research "Other types of instruments can be connected Center two years ago. Under the control of an IBM to the computer using interfacesavailable from 1620. computer, the crystal is turned so that the DigitalEquipmentCorp.Speciallydeveloped x-rays are reflected by differentplanes of atoms in software packages, designed for the PDP-8/I, the errtal and the intensity of the reflected x-rays is extend the computer's capability to serve other measured. The system is run 24 hours a day, without instrumentation such as mass spectrometers, NMR supervision. The computer is operated in a spectrometers and a variety of otherinstruments. sharing mode, so that it can perform many prelim- "The new system detects peaks and shoulders, inary computations while the experiment isin calculates peak areas and peak retention times, progress." allocatesoverlapping peakareas,correctsfor "Several years ago after a crystallographer had baseline drift, and calculates component concentra- joined the staff at the IBM Research Centel, the tions. It also identifies peaks: applies responsive question arose as to how best to get the data for factors, and. types out a complete analyses report him that he needed in order to work at a high pro- immediately after a rm. is finished. In addition, fessional level.he solution seemed to be to auto- GasChrom-8 allows parallel operation of a strip mate an x-ray diffractometer, an instrumentwhich chart recorder and attentim switching." (Computers could produce the needed data, and finally, the and Automation 17, No. 4, 63 (Apr. 1968).) nature of the data taking taskseemed to indicate 2.26"Small defects in multilayer printed circuit that the instrument should be on-line to a computer boards can be swiftly and accurately repaired but for real-time feedback and instrument control. An first they have to be found. To pinpoint such IBM 1620 computer was rented, the control program common faults as short and opencircuits, the written, and the system has been in steady operation Autonetics Division of North American Aviation, since late in 1963." (Cole, 1966, p. D. Inc., has developed a computer-controlled, infrared "The data taking and justification is not the only radiometer that scans and analyzes a board's computer oriented part of thecrystallographic thermalradiationpattern." (White and Jones, work. The data analysis, the determination of the 1965, p. 96.) atomic positions from the diffraction data, involves 2.27"The objective of the unique and ambitious extensive calculations. And finally, with the output program is to construct a highspeed electronic increased as it has been, the graphic display of scanner-computer for automatic 'visual' observa- intermediate calculations and the drawing of the tion and analysis of medical specimens and bio- views of the molecular configuration become bottle- logical system." ("PHS Grants ...", 1965, p. 8.) necks if use is not made of computer plotting. Thus Further details of the project plan under the direc- computers play a role in each step ofthese experi- tion of Glaser of the University of California are mental studies." (Cole, 1966, pp. 2-3). See also as follows: "With this automatic systemitis Cole et al., 1963, and Abrahams, 1963. believed that identification of the causative organism 2.23"Special peripherals can be on-line to the and determination of their drug sensitivity or resist- jobs...The Manchester Installation has an x-ray ance can be made much fasterand with higher diffractometer and an audio I/O device in this reliability, than is possible with plesent hospital .category." (Morris et al., 1967, p. 68.) techniqueu. It is expected that essentially the same 2,24"The surprising thing is the trend to auto- techniques will be useful for monitoring levels of mation in the research laboratories, you find it in contamination of food, water, and medical supplies those experiments that are characterized by large where it is important to know how many living organisms ale present and of what kinds." ("PHS numbersofrepetitivemeasurements ...But you find it also in experiments that arecharacterized Grants....", 1965, p. 8.) more by complexity than by repetitiveness.There 2.28"The systems assembled at NBS auto- the hub of the automation comes in the analysis of matically obtain and record data from unattended the datait's done on line, in real time, most experiments at rates up to 50 datum points a second often with a small computer located right at the The system can also be programmed to vary experi- experiment and, in fact, incorporated into it. The mental par, .zs, such as voltage orwavelength." analyzed data are displayed before the experimenter ("Transistorized Modules for Data Logging Sys- and guide him in the performance of experiments tems", 1966, p. 42.) See also Stein (1966). 62 2.29"Argonne National Laboratory's CDC 3600- came in 1946, when the United States Army Signal based system represents user-manufacturer co- Corps first succeeded in bouncinga signal off the operation for adaptation to local needs. ANL is lunar surface. Even then, some of the advantages modifying CDC's SCOPE system to use the 826 radar might have over optical methodswere evi- disk for scheduling and handling input/output dent. For example, certain quantitative changes in functions. ANL developed programs that link a the polarization and absorption of reflected signals cathode-ray-tube display and camera recorder to resulting from transmissionsatdifferent wave- the system for recording graphic and tabular data. lengths might give us insight into the depth, and ANL's Standard Peripheral Processor, a 160A even the nature, of the surface layer; extremely package, includes an ANL modification for introduc- accurate determinations of range and surface fea- ing a microwave link for high-speed data trans- tures, as well as of lunar periods and possible un- mission. The Processor allows tape-to-tape trans- expected perturbations, might be made; and perhaps missionineither direction and tape-to-printer of greatest eventual value since neither distanceto transmission from the user to the computer system. the target nor atmospheric conditions constitute A data link from an ASI-2100 computer allows real- major problems, the methods developed for studying time interrupt to users of the 4.5 MEV and 12 MEV the moon could be readily adapted to planetary use." Tandem Van deGraaffaccelerators.ANL's (Wells, 1967, p. 26). PAULETTE is a computer-controlled device that 2.31See Meissner et al., 1963. Figure 12 illus- counts tracks made by subatomic particles in a trates the process flow for the automatic preparation finegrain photoemulsion. A hybrid computingof "smooth sheets" (i.e., charts containing sub- system recently installed joins two PACE analog stantially all the readings froma particular survey computers to a PDP-7. The hybrid enables analysis trip).Digitalrecordingsof thedata collected of dynamic systems that yield discrete and con- include information on the sounding in feetor tinuous data. CHLOE, an automatic film-scanning fathoms, including the phase number for the echo system,digitizesdatafrombubble-chamber sounder; a serial number identifying the reading; photographs and oscilloscope traces. The system the time of the reading; the position fix data; the isbeing evaluated for fingerprints, radiological heading o5 the ship, and data entered manually, photographs, optic nerve signals, galactic shapes, such as the ship's draft, tide corrections, and sound etc." (Swanson, 1967, p. 37.) velocity. 2.30"It thus appears that standard radioactivity 2.32"A new system, called the Laboratory measuring instrumentation (and this is intended Data Collector and General Purpose Interface, was only as an example) encodes information in a form designed to provide a reliable, flexible and yet that can be transmitted to and processed by the inexpensive digital data acquisition and recording computer in a simple, practical and very inexpensive capability for those who must collect, record, and way." (Neilsen, 1965, p. 635.) process digital data from laboratory instruments 2.30a"Multiple graphic display drivers, radar or other sources, and record this data in digital form. scanning equipment, map digitizing devices and The standard system allows the interfacing and earthquake sensor arrays are typical examples of control of a wide variety of digital output devices data acquisition or distribution systems. Data are and instruments with almost any digital data record- received from sensors, formated and stored in mass ing device or direct computer interface. It eliminates storage for acquisition, and then prepared for the need for a new expensive "black box" for each distribution.Periodically,dataaretransferred new or different connection between such devices from mass storage to other media. Here there is that becomes necessary in a laboratory or digital one basic functionto compile data but as many dataacquisitionsystem. The Laboratory Data processes as there are independent sensors or Collector can be programmed and operated by distributors." (Gaylord, 1968, p. 27). laboratory technicians in a few simple steps, and "Current effortstoincrease the power and does not require the experience of a computer sensitivity of radar observatio;:is, to reduce process- programmer. It also allows automatic identification " ing time and enhance pictare resolution, and to of data and automatic formatting and separation of significantly improve microwave and data processing data samples. It has built-in error checking and hardware will no doubt have major effects on the maintenance features which guarantee the validity future of radar astronomy. of data processed and allows rapid repair of the unit "In the data processing area, the immediate goal by the technicians if this becomes necessary. The is obvious: to develop total display systems capable basic Laboratory Data Collector unit which will of permitting the continuous display of radar data. interface standard analog-to-digital converters and This, together with the need for solutions to the digital voltmeters, and control punched paper tape special problems imposed by planetary observations or incremental, magnetic tape units, sells for $6,400. and the constant search for better resolution, should Various options including built-in analog-to-digital provide enough work to keep researchers busy for converters, digital multiplexors, digital data com- a long while." (Wells, 1967, p. 32). pressors, and digital clocks can be provided on a 2.30b"The first real indication th?..*. radar prob- modular basis. Berkeley Scientific Laboratories; ing might yield significant information on the moon Berkeley, Cal." (Computer Design 5, 68, Feb. 1966). 63 TRANSDUCERS ENCODERS INDICATORS

MONITORING PERSONNEL

SHIPBOARD < DATA

DATA LOGGER

TYPED RECORD :.: - -t PAPER TAPE

BOAT SHEET

ADDITIONAL INFORMATION DATA CONVERSION GEOGRAPHAND I POSITIONS COMPUTER

PROJECTION COORDIN'TS

PUNCHED CAR DS

PREUMINARY SMOOTH SHEET

Y Y LOCAL OFFI C E INSPECTION PRECISIONY PLOTTER

IAPPROVED CARDS

FINAL SMOOTH SHEET

tr TO CENTRAL OFFICE AUTOMAT ICPREPARATION OF SMOOTH SHEET

FIGURE 12.Automatic preparation of smooth sheet.

64 2.33 "An improved mobile data acquisition at the top of which a solid-state parametric pre- system was designed for maximum flexibility in amplifier is located. Buried cable connectsgroups obtaining data and producing computer compatible of 25 seism-meters into clusters or `subarrays'..., tapes at remote test sites. This completely portable In a concrete vault near the center of each subarray system allows engineers to easily check operating a Subarray Electronics Module (SEM) multiplexes conditions in industrial or scientific equipment in and digitizes the 25 seismometer outputs intoa isolated plant locations or in the field. Inputsrepre- single bit stream which is then transmitted to the senting pressures, temperatures, strains, flow rates, LASA Data Center in Billings bymeans of open speeds, and other variables are recorded at rates up wire and microwave circuits, terminated at sending to 20,000 samples per second. The recorder is an andreceivingendswithsuitableModulator- integral part of the system, called Mobidac, the Demodulator (Modem) units. At the Data Center system accepts fifty low-level analog inputs, seven the bit streams from the 21 subarraysare combined, high-level analog inputs, and two digital inputs processed by digital computer, and the results which may originate either internally or externally. appropriately displayed or transmitted to remote Patch panel programming permits selection of locations for further processing." (Green et al., 1966, either binary or BCD formats, any one of several p. 328). data acquisition rates, and allows high-level and low- 2.35"The Advanced Record System is being level inputs to be intermixed in the program sam- used in connection with remotesensors to monitor pling sequence. Lower sampling rates than those water pollution in certain rivers on a 24-hour-a-day obtained by subcommutation areavailable by basis." (Johnson, 1967, p. 20). start/stop recording. Nixie displays are provided for 2.36"The AMOS (Automatic Meteorological read-out of data and for record count. Data values Observation Station) systems have been developed are recorded in gapped or gapless format for direct in a cooperative program between the National analysis by a computer. System Engineering Labs, Bureau of Standards and the U.S. Weather Bureau Inc., Fort Lauderdale, Fla." (Computer Design 5, for the purposes of storing and processing weather 74 (Feb. 1966).) data gathered at a number of remote stations." "An improved Mobidac (mobile data acquisition) ("The AMOS IV Computer fora Prototype system designed for maximum flexibility in obtaining Automatic Weather Station, National Bureau of data and producing computer compatible tapes at Standards, Tech. News Bull. 45, 13-15, Jan. 1961). remote test sites has been introduced by Systems "Plans arecurrently being generated fora Engineering Laboratories Inc. In the portable sys- dedicated network toprocess nationwide weather tem, inputs representing pressures, temperatures, information which includes file maintenance for strains, flow rates, speeds and other variables are subscriberquery.Messagesareperiodically recorded atrates up to 20,000 samples/sec." received from a worldwide weather sensing network. (Information Display 3, 63 (Mar /Apr. 1966).) The computer will format incoming data according 2.34"The Large AperatureSeismic Array to preprogrammed needs of a subscriber and will (LASA) project is being planned as a real-time also maintain an updated data baseso that sub- data-acquisition system for detecting and evaluating scribers may randomly query the center for the earth tremors (earthquakes or man-made tremors)." latest weather at any point in the world. Although (Aron, 1967, p. 65.) this network is presently plannedas a limited "Montana Array. This network is used for detec- access system, expansion of the service to total tion of earthquakes and nuclear explosions by government operations thru general purpose com- linking together 525 seismic detectors coveringan munication facilities is foreseen." (Jacobellis, 1964, area of 30 thousand square miles. Large quantities p. N2. 1-3). of information are sent to M.I.T. for processing. 2.36a"There are a variety of potential applica- One reel of magnetic tape is generatedevery 10 tions for a real-time analysis system. Examplesare minutes." (Brown, et al., 1967, p. 20.) See also air traffic control using Doppler radar techniques, Maguire, 1965, and Press and Brace, 1966. Doppler radar measurements of meteorsor terres- "The large aperture seismic array,or LASA ... trial orbital objects, vibration measurements during is a system of seismometers, signal transmission non-destructive testing, and medical monitoring facilities, and signal processing hardware designed during intensive care or surgery. Inmany other to provide a greatly improved detection and identifi- applications rapid availability of results froma cation capability for remote seismic disturbances number of data sensors is needed for efficient (teleseisms) of small size. Figure 1 showsa map of operation, as in seismic exploration." (Larson and the area around Miles City in eastern Montana Singleton 1967, p. 665). where the experimental LASA has been installed, 2.36b A few details from the Mariner mission to and it indicates schematically the disposition of photograph the Martian landscapeare as follows: the 525 seismometers over the 200 km aperture, "Spacecraft Mode 4 was Television Playback. as well as the 500 miles of open-wire telephone Mode 4 consisted of approximately 40,000 words, line and 275 miles of microwave facilities that or 240,000 bits of information per picture. The provide the signal transmission....Each seis- televisionformatcontaineda31-bit PN se- mometer is at the bottom of a 200-ft deep hole, quence .. then 30 bits of identification followed 65 by 200 6-bit elements for each TV line. Two hundred for earth scientists to have up-to-date base maps, TV lines constitutes 1 complete picture." (Gian- to include distributions of dynamic features, and opolos and Curl, 1966, p. 38). to readily assess changes in distributions by com- 2.37"Scientists at the California Institute of paringrepetitivecoverages."(Fischer,1968, Technology Jet Propulsion Laboratory, using an pp. 45-46). IBM 7094 computer, have been able to improve the "In the study of vegetation the presence of the quality of the most recent photographs of the high, near-infrared reflectance spectrum a funda- moon's surface by a factor of two... mental feature of leaf moisture in all healthy plant "Each picture [from Ranger VIII is made up of growth enables sensors of various types to define about 90,000 points and there are about 64 possible the vigor of growth. The visible and near - infrared shades of gray for each point. The signals are re- regions, combined with the time of the seasons, corded on magnetic tape and fed into the IBM enable the crop type and vigor to be predicated computer in digital form. The computer then plays with high precision." (Lyon. 1968, p. 46). a brightness-level matching game ...It is possible "At the national and international levels, hydrolo- for the computer to correct smudges in the picture gists are initiating studies in the use of remote where signals were not picked up evenly. It does sensing and data relay for a wide range of hydro- this by comparing one scan line with neighboring logicalpurposes ...Currentstudiesindicate lines and correcting the extreme variations caused that present space technology can provide sub- by the electromagnetic interference. After the stantial benefits. These include basic studies of compute): has processed the information, a visual the hydrologic cycle, snow and ice mapping, real- image is recreated by a film reproducer." (The Com- time communications of groundbased hydrological puter Bull. 9, 62 (1965). See also Data Processing 7,. data, survey of coastal hydrologic features and large 267 (July 1965).) lakes, and remote sensing of soil moisture." (Bock, 1968, p. 46). 2.38"Two experiments are under consideration, "Purdue University engineers, working in the one ranging, the other photographic. For the ranging NSF-funded LARS (Laboratory for Agricultural experiment, a 10-joule ruby-laser pulse is trans- Remote Sensing) program, have developed remote mitted from the ground through a special-purpose sensing and pattern-recognition techniques which telescope with a 1.5-meter aperture. It is returned can be used to gather information on crops from from the lunar reflector back through the same the air. The system entails sensing and recording telescope to a photomultiplier ..For the photog- of infra-red reflections from plants and could be raphy experiment, a 40-joule laser signal is trans- used with a low-level satellite (such as a Tiros) mitted through the special-purpose telescope but equipped with a spectrophotometer and a tape it is returned by the reflector to a separate general- recorder. From the tapeinput,the computer. purpose telescope...displaced approximately recognizes signal patterns and maps the area, 1.5 km to the east of the transmitting telescope. designating, for example, what crops are grown in This displacement compensates for relative trans- each field. This technique reportedly could be verse velocities between the earth and moon. modified to aid such projects as water and oil "There are many uses for precise information explorations, planet fly-bys, and icecap and drought on the range to a specific point on the lunar surface. investigations." (Datamation 14, No. 5, 113 (May For example, laser observations throughout' the 1968). lunar libration cycle could give independent checks on the moon's mechanical figure and provide obser- It is noted also that: "Aero Services and Spectra vations useful in calculating an improved physical Physics are incorporating a gas-laser ranging device ephemeris of the moon. Such observations could and a barometric-pressure transducer into a system also provide an independent check on the earth's that has demonstrated a capability for accurately equatorial radius, and on lunar theory dealing with profiling the surface over which the aircraft is being the eccentricity of orbits." (Hunt, 1967, pp. 5-6). flown..." (Parker and Wolff, 1965, p. 30) and that 2.39For example, "the idea behind multiband "...the U.S. Army Electronics Laboratories... reconnaissance is that by comparing two or more reported sounding through 1500 ft. of cold ice from photographs of the same object made in different low-flyinghelicoptors with a...pulsed radar regions of the spectrum we may learn something system ..." (Parker and Wolff. 1965, p. 28.) about the object we could not learn by studying 2.40"...Theoriespertainingtocrevasse the tonal values on just one photograph... detection have been substantiated by infrared "Infrared mapping of large forest fires... measurements over snow surfaces in Michigan "The U.S. Army Engineer Waterways Experiment and in the Arctic, and an airborne technique has Station...[used] a radar at four different fre- been devised for finding snow-bridged crevasses quencies (as well as with other sensors like gamma- that often cannot be seen by visual means." (Parker ray spectrometers) to analyze soil conditions as and Wolff. 1965, p. 20.) far down as 18 inches." (Parker and Wolff, 1965, 2.41"Long radio waves are helping prospectors pp. 23, 26, 27). search for water in arid regions of the Soviet Union. 2.39a"The quality of near-orthography, coupled The system, developed at Moscow State University, with the great speed of spacecraft, makes it possible operates on the principle that damp ground absorbs

.66 _ . radio waves better than dry ground does." (Elec- data filtering operations. Associated tronics 38, 216 (1965).) programs are carried out in network theory, Mathematicalanalysis 2.42"Graphic 'fingerprints' of water frost and and pattern recognition." (Mazzarese,1965, p. 69.) of carbon dioxide 'frost' have been obtained in "What is needed isa means of programming our recent measurements at the NBS Institute for sensors to provide the interpreter only with the data Basic Standards...by H. J. Keegan and V. R. he needs for his particular purpose."(Parker and Weidner...The fingerprints clear patterns of Wolff, 1965, p. 31.) the light-reflecting and absorbing properties of the 2.44"The militaryis becoming increasingly two types of frost at wavelengths in the infrared interested in multiprocessors organizedto exhibit region of the spectrum will be checked against the property of graceful degradation.This means infrared spectra to be obtained from Venus in future that when one of them fails, the otherscan recognize space probes." (NBS Tech. News Bull. 50, 20 this and pick up the work load ofthe one that failed, (1966).) continuing this process until all of them havefailed." 2.42a"Adaptive sampling is a means by which (Clippinger, 1965, p. 210.) the sampling rate of a givensensor can be adjusted "Systems are designedto provide either full to correspond to its information rate...Most of service or graceful degradation in the face offailures the time present telemetrysystems greatly over- that would normallycause operations to cease." sample the data. In nearly allcases, the sampling (Bonn, 1966, p. 1865). rate is set on the basis of the fastest expected "From our experiencewe have concluded that response from the source and not on the basis of system reliability in the medical community the quiescent or normal value. must Howev,-r, to match provide for sevei al levels of failure,leading to the the sampling rate to the data activity would require term'fail-soft' rather than `fail-safe'." (Baruch, an extremely complex machine and to date a 1967, p. 147). practical mechanization has not been developed." "On-line systemsare still in their early develop- (Korb-Ilan, 1967, p. 133). ment stage, but now that systemsare beginning to 2.43"Much of the work [with AFCRL's Experi- work, I think that it is obvious thatmore attention mental Dynamic Processor DX-11 has beento find should be paid to the fail-safeaspects of the better techniques for on line control of real time problem." (Huskey, 1965,p. 141).

3. Communication Systems and Data TransmissionLinks 3.1"A type of pseudo-noise modulation has been equipment in order to effecta system which is both developed recently foruse in random-access com- efficient and responsive." (Hittel, 1966, munications p. 395). systems.Here, many users may 3.3"...Many workers in both communications communicate simultaneously without going through and data processing havecome to. realize, in the a central switch or control...Code-division course of making further progress, that the two multiplexing using pseudo-noise modulation is... fields are becoming more interdependent.... [a] technique in which each transmissionis modu- "The most obvious bases for such interactions lated with a code that identifies theuser and to have been the physical extension of data processing which the addressed receiver isset to respond." systems and the increasing complexity of traffic in, (Andrews, 1965, p. 227). and control of, communicationssystems... 3.2"Two basic types of codesare found suitable "The major areas of interaction [communications for the burst type errors. The first is the forward- and data processing systems]can be identified as acting Hagelbarger code which allows fairly simple (1) high-speed digital data transmission,(2) digital data encoding and decoding with provisionsfor modulation techniques, and (3) computer-controlled various degrees of error size correction anderror communications." (Andrews, 1965,p. 226). size detection. These codes, however, involveup to "Data processing systems and communications 50 percent redundancy in the transmittedinforma- systems are being created in such a fashion that it tion. The second code type is the cyclic code of the is no longer easy (nor proper)to differentiate. be- Bose-Chauduri type which again is fairly simpleto tween the two." (Franklin, 1965, p. 182.) encode and can detect various error burst sizes with . "Both a single computer utility anda network relatively lowredundancy. This "code typeis involve a conjunction of two fields,computer tech- relatively simple to decode forerror detection but nology and communications technology." (Pyke, is too expensive to decode forerror correction, and 1967, p. 163). makes retransmission the only alternative." (Hickey, "Although data transmission predates electronic 1966, p. 182.) data processing bymore than a hundred years, the 3.2a"The ever increasing utilization ofcom- blending of the two technologies isa relatively new puter complexes from remote sources imposesa new art; a demanding one with few practicing specialists. dimension to the almost overburdeningtask of Each of the various aspects ofan information gather. systems design. Remotely accessed systems require ing, processing and transmission system hardware, a tight integration of communication services and software, communication channels, forms,etc.has 67 its share of specialists. But the man knowledgeable companies (such as Collins Radio Co. and Lenkurt in all phases of the blended technologies is still a Electric Co.) manufacture data sets for use with rare individual." (Menkhaus, 1967, p. 30). private communications facilities or, in some cases, "Western Union estimates that within five years with leased common-carrier facilities." (Reagan, some 60% of all computers will be tied into the 1966, p. 66). nation's communication networks. ... "DATEL. Thisisa new international data "The Bell system predicts that eventually half of transmission service being offered by ITT World the information transmitted over its network will Communications Inc. DATEL 100 is a full-duplex be data.... 100 bits per second facility. DATEL 600 can be used "Data transmission lines thrust the common for 600 to 1200 bits per second transmissions. carriers to the forefront as prime candidates for Additional DATEL servicesarebeing made entry into the information utility field." (Irwin, available." (Hittel, 1%6, p. 397). 1966, pp. 22-23). 3.5Thus, "the operational IntelsatI(Early 3.4"The Automatic Map Compilation System, Bird) and Intelsat II satellites and associated earth developed and operatingat Thompson Ramo- stations that comprise the present commercial Wooldridge, abstracts terrain altitude information satellite communications system provide a capacity from aerial photographs by correlating the imagery of close to 720 voice circuits between major earth appearing on stereo pairs, and outputs contour stations. All but a limited portion of the inhabited information and new photographs in which the globe is now within the line of sight of a commercial imagery appears in 'true' orthographic projection satellite relay. In addition, it is expected that by position." (Bertram, 1963, p. 108). 1968, with the operation of Intelsat 111, a truly global Further, "the input data for a given compilation, is international satellite communications system will in the form of a pair of aerial photographic trans- exist." (Martin and McKee; 1967, p. 63). CC. parencies together with pertinent camera data; i.e., ..A series of demonstrations last June in position and attitude of camera for each trans- which an RCA computer and the Relay I com- parengy, focal length of camera and distortion munications satellite were used to transmit news characteristics of the lens.... copy between the U.S., England, and Brazil. In "The Automatic Map Compilation System... some of the experiments, the 301 computer justified outputs a chart showing the altitude contour inter- type lines, and the resulting signals were fed directly vals over the stereo area and a new photograph in into linecasting machines." ("The (R)evolution in which the imagery has been moved so as to appear Book Composition...IV", 1964, p. 71.) in correct orthographic projection position to a "The heart rate, brain waves and sleep pattern of selected scale. The system utilizes a combination of a U.S. Marine patient in a hospital near Tokyo were digital and analog techniques to achieve the required transmitted to Houston by satellite yesterday in a accuracy and speed of operation ... demonstration of how medical data can be sent "Data printout takes two forms: (1) An altitude around the world. chart exposed by computer control of the brightness "The data was sent from the Marine's bedside at of one scanner in accordance, with the ineasured Camp Drake Hospital in Japan to Intelsat, a satellite altitude. Three brightness levels are used in a rotary in orbit 22,300 miles above the Pacific. Signals from sequence to show successive contour intervals; (2) A the satellite were relayed to a U.S. receiving station new photograph exposed by reproducing the image at Brewster Flat, Wash. picked up by one of the photograph scanners and "The demonstration was arranged for the Na- imaging it appropriately on the photosensitive film tional Telemetering Conference. Itwas believed to sheet." (Bertram, 1963, pp. 105-107). be the first time that several medical measurements 3.4a"The dual-mode Automatic-Picture- on the same person have been transmitted simul- Transmission (APT) recorder...has the unique taneously by satellite. capability of being able to receive transmissions "Regular telephone circuits fed the signals from from meteorological satellites, but between satellite Brewster Flat to the conference here and tocom- transmissions it can be switched ever and operated puter centers at the U.S. Public Health Service in as a standard weather-map recorder at 120 rpm and Washington, D.C., and the University of Texas. 166 lines per inch. It presents and prints out a large "The demonstration showed that the best medical display in sepia-tone on a special Alfax paper ... care group potentially can be extended to any person Transmission can be received over land lines 9r via in the world by bringing his data to any physician radio with an appropriate converter." (Bliss, 1966, or to a computer for analysis, said Dr. Fred Vogt. p. 3). "Vogt, department chairman of biomedical engi- 3.4b"Data sets are available from the common neering at the University of Texas School of Bio- carriers for use with a wide range of standard com- medical Sciences, said the cost someday may beas munications facilities. Among the most widely used little as $10 to have a patient's records displayed to are the Bell System Data-Phone 200 Series Data any medical expert anywhere. Sets, which permit use of the public telephone "Dr. Cesar A. Caceres of the U.S. Public Health network for data communications at speeds of up to Service suggested that hospitals begin thinking 2,000 bits per second. In addition, several other about computers to watch patients in surgery. He 68 noted the development of tiny electronic instru- sending up to 200 words a minute, the company ments to monitor such medical functions as heart- has centralized much of its general accounting beats and mental activity. by transmitting payroll and similar information "The principles and techniques employed to between distant points." (Marks, 1964, p. 16). monitor astronauts in flight can be used effectively "Another approach... isthe Data-Phone. to monitor patients undergoing surgery, Caceres Currently the Associated Pressistransmitting said." (The Washington Evening Star, Apr. 10, justified stock market copy all over the United 1968, p. A. 18). States. This is on the wire minutes after the markets 3.6"Dual routing, using both microwave and close in New York." (Blondeau, 1963, p. 89). cable, can be used to improve the reliability of any "There also are DATA-PHONE data sets avail- telecommunications systems, 10: required. This has ablethatautomatically establishcallsfor the been successfully accomplished by the Power Au- transmission of data between business machine thority of the State of New York and others." (Eld- terminals without the aid of any operators. This ridge, 1967, p. 179.) is possible by using a device called an Automatic 3.7For example, O'Sullivan (1966) describes Calling Unit. When directed by the associated Raytheon's private telephone branch exchange for business machine, this device automatically dials varied use of several different commercially avail- the telephone call. These units are capable of able time-sharing services. dialing telephone calls by using the conventional 3.8"A service provided by one of the common rotarydial technique or the new Touch:Tone carriers is called WATS which stands for Wide dialing system." (Goettel, 1966, p. 195.) area Teiephone Seivice. This provides for a flat There are conditions which require extensive rate for some number of calls from one point within data manipulation such as pay-roll transactions, one of several sizes of geographical areas. It is while others involve very little calculation at all possible at some point to show that the cost of the but consist, primarily of a message store and flat rate is less than the cost of some number of forward operation. Some of the machine aids individual calls. The use of the automatic subscriber to communications include scheduled automatic calling feature of the A-1200 allows this decision to calling or polling of remote transmitters from a be made anytime, because the central terminalcentral console, with automatic error detection places all poll calls. The remote stations normally and retransmission. In other cases the hardware do not initiate calls although provision is in the does little more than prepare a magnetic tape equipment for them to do so. The frequency with in computer usable form that is then mailed to which the central terminal polls subscribers auto- the central site." (Hickey, 1966, p. 176). matically can be selected on the console." (Hickey, 3.12"The touch-tonetelephone, already in 1966, p. 178.) service in many localities, is finding ever-increasing 3.9"Telpak Service. Telpak is a private line use as the key ingredient of low-cost data collection service that provides communications capacities of and management information systems... various sizes suitable for large-volume point-to-point "There are several outstanding advantages of transmission of voice, data, or other forms of com- using touch-tone telephones for gathering data: munication. Telpak capacities can be arranged so 1. The cost of each transmitting station would that they constitute a group of voice channels suit- be only two or three times that of an ordinary able for use for voice or data communications or a telephone. group of channels suitable for use with teleprinter 2. The station can be used interchangably as equipment. The Telpak capacity can also be used a telephone or as a data transmission device. as a single large channel for high-speed data services 3. It is as easy to use as a 10-key adding machine. such as magnetic tape, computer memory and 4. It can readily be connected to any existing facsimile transmission." (Goettel, 1966, p. 197.) telephone system, local or long-distance. "Cost of a telegraph channel in a private line is 5. It is easily made portable, and can send from $1.10 per mile per month; in Telpak, it is about 10 any phone jack. cents." (Titus, 1967, p. 63.) 6. It can transmit to a wide variety of machines, 3.10"A new compatible communications adap- ranging from paper-tape or card punches tor ...that permits UNIVAC 1004 and 1005 sys- to computers. tems to be used a high-speed data communications It approaches real-time performance at the terminal with non-Univac computers...The new lowest possible cost. DLT-9 uses the 4-out-of-8 code and format at 8. Its applications can range in size from a synchronous data transmission rates of 2000 and small in-plant system to a complex nation-wide 2400 bits per second. A higher speed model of the network." (Davenport, 1965, p. 36.) DLT-9 operates up to 40,800 bits per second using a TELPAK service." (Commun. ACM 9, No. 9, "With touch-tone telephone, people can call 707 (Sept. 1966).) a computer and, after the initial connection has 3.11"Thefirstmajor Data-Phone business been made, use the calling buttons as an input installationwas made for Firestone Tire and keyboard to communicate unambiguously to the Rubber, Co. in October, 1958. Using data sets computer,therebytemporarilybypassingthe 69 i ...,-4, foNt_y-c.;,741,,,,,,,o_A 3;.',-N-,...---,,,,,-.

very difficult problem of speech recognition. With techniques. First the digital information has been computer-generated speech, we can foresee the efficiently packed into an analog signal format well use of touch-tone telephone sets as the only remote matched to the channel (a technique called multi- terminal device for large and complex information level vestigal sideband). Second, compensation for retrieval systems." (Lee, 1968, p. 333). distortion introduced by the channel has been made 3.13"A 12-button version of the new pushbutton possible by the application of automatic adaptive telephone might be used in a rather simple way as equalization. In addition, the application of a simple a mixed alphabetic-numeric input device for modern but efficient forward-acting error control method computer systems, without interfering withits well suited to the noise statistics of the channel is ordinary use in placing telephone calls." (Davidson, possible on an optional basis. The redundancy asso- 1966, p. 27.) ciated with it reduces data rates by about one-sixth, 3.14"Bell System's Data-Phone service now but it improves the average error-free interval by enables business machines transmitting at speeds factors from one-hundred to one-thousand. The up to 2,000 bits-per-second (about 2,700 words-per- costs expected to be higher for these data set minute) to converse with one another without any arrangements than for those operating at lower human intermediary. Operating in conjunction "a- speeds, In t studies show that many systems with Bell System's 801-type Automatic Calling Units, the long-haul requirementswillfind the additional 201A Data-Phone data set now is compatible with speedeconomicallyattractive."(Simms,1968, computer-to-computer transmission as well as with p. 23.) machines that transmit punched paper tape, mag- 3.14b"At the wide band of the data spectrum netic tape, and card media. With this new feature of wefindrequirementsfor the transmissionof the 201A Data-Phone data set, companies with occasional high speed bursts of datato update several branch offices will be able to poll these computer memories,toload-share a computer offices automatically over the regular telephone or just for beck -up in case of computer failure. snitched network for sales, production, and other While these and other applications need wideband information. By malting such polls in off-hours, channel capacity, they can't justify the costs of companies will be able to take advantage of lower full-time wideband private lines. Recognizing this evening telephone rates and also reduce the trans- need, we are now introducing on a trial basis a mission traffic which their telephone service must servicecalled DATA-PHONE 50 a 50 kilobit handle during regular hours. The typical equipment per second common user switched service. Ini- configuration would consist of an automatic calling tially,it'sbeing offered in Chicago. Additional unit and a data set at the headquarters location trialofferings may be made in other cities in and a data set at each branch, along with the busi- the near future." (Quirk, 1967, p: 521). ness machines. The telephone numbers of the loca- 3.14c"Score, carrying Signal Corps electronics tions to be polled would be stored in the computer in an Atlas pod, went up in December, 1958. system. Bell Telephone System, New York, N.Y." Although Score is best known perhaps as `Eisen- (Computer Design 5, 69, (Feb. 1966).) hower's Christmas Card', the firstvoice trans- "Data phones now available for switched tele- mission from space achieved withitsdelayed- phone networks provide approximately 2,000 bits repeatercapability,italsoprovidedreal-time per second of [display] data, but within the near relay. Score spanned the continent as a real-time future this rate may be approximately doubled..." system, and alsocarried some 140,000 words (Haring, 1968, p. 38). ofdelayed-repeater Mail Bag')trafficbefore 3.14a"Innovations are also being planned to re-entry.Itseight-wattoutputat UHF served improve the effectiveness of our existing plant to ground stations based on the then-current missile carry data signals. For example, a new DATA- telemetry technology." ("Communicating by PHONE data set, capable of transmitting 3,600 bits Satellite", 1966, p. 8). per second over the dial telephone network, is now "Through the advent of the communications in the technical trial stage and will be available in satellite techniques modern communications the near future. A similar data set has been technology has made itpossible to have rapid, developed for private line service, but this one is adequate, and secure communications throughout able to handle 7,200 bits per second. Even higher the world." (Johnson, 1966, p. 99). speed versions are being planned for later on." "Satellitecommunicationstechnologyoffers (Quirk, 1967, p. 521.) thepotential of essentially unlimited wideband "A new data set is being developed to permit communications." (Johnson, 1966, p. 100). substantially higher bit-rates on voice grade lines 3.15"...Synchronoussatelliteswiththe than is possible with the presently available data capacity of 80,000 voice telephone channels or sets. One arrangement intended for use on the net- 160 one-way television channels, available to and work will operate at 3,600 bits per second with from rooftop antennas without complex switching simultaneous 150 bit per second reverse channel. arrangements... On private-line voiceband circuits, the data set "Coherent light from lasers will provide a revo- will operate at 4,800 bits per second, and there is lutionary increase in the volume of communication a possible extension to 7,200 bits per second. These that can be sent over a single pathway." (McManis, increases in speed have been made possible by two 1966, p. 28). 70 -1.,4-;4...P51(11,,,,,,,,trAliry,,,e11, ,

"Television signals have been transmitted over categories of distance over which itis applied. infrared beams from diodes and diode lasers Local loop transmission connects subscribers to of GaAs. Of course the objection to the infrared the common carrier wire and repeater centers and laser for transmission stems from its atmospheric local exchanges up to 15 miles in distance. Tribu- absorption properties, which make it a fair-weather tary transmission spans up to 150 miles to larger device. We may consider the possibility, however, centers or tributary exchanges. Medium haul trans- of transmitting the radiation through tiny hollow mission extends over distances spanning continents. waveguides...with laser diodes as repeaters, Long haul covers up to 5,000 miles of transmission such an arrangement may provide the basis for between continents." (Kaplan, 1968, p. 121.) alarge-capacitycommunicationcable."(Lax, "The microwave technology for tributary andme- 1965, p. 72). dium haul transmission has permitted increases in "To accommodate this expected volume, the data carrying capacity through theuse of extended capabilityofexistingchannels wire,cable, radio spectrum and digital modulation methods. microwaveandsatellitesisbeingupgraded Digital baseband microwave systemsare capable of and new channels are being added. Bell, Raytheon, transmitting data time division modulated at a GeneralElectric,InternationalTelephone and rate more than an order of magnitude greater than Telegraph Co., IBM and others are experimenting conventional microwave frequency division modu- with laser beams as possible data-carrying channels. lated. As an example, a modern type microwave There's a rumor that IBM will demonstrate the system operating within a 30 MHz channel band- use of light beams to transmit data during the width in the 6 GHz carrier band has the approxi- international fair in Canada this year. Guesses mate capability as shown in Table II. The micro- as to When this technique will emerge from the wave analog transmission system has a greater laboratories to be applied to everyday use, on a capacity for analog signals." (Kaplan, 1968, p. 121.) significant scale, vary widelyfrom three to 15 "The advancing technology in transmission has years." (Menkhaus, 1967, p. 37). been in the direction of large capacity digital 3.16"A second development which promises to baseband systems.Digital baseband microwave be useful in the future is a pulse code modulation and T-systems are important advances, allowing system now being introduced for voice transmission. greater than an order of magnitude increased This system digitizes a number of voice channels datatransmissioncapacityoverconventional and places the resulting information from each frequencydivisionsubdividedchannels.The channel in a time division way on a single one and high data rate systems should contribute signifi- one half megabit per second channel. In the future, cantlyto reduced transmission error rate and this high speed data channel may provide a lower costs. The application of these wideband data cost form of high speed data transmission and may systems ought to be accelerated because analog make possible a high speed facsimile system ofa transmission can be included as economic users more economical type: However much work re- on some of these high data rate facilities." (Kaplan, mains to be done in this area." (Wier, 1965, p. 107.) 1968, p. 129). "The advantages of planned, systematic innova- "The advancing digital multiplexer and modem tions are demonstrated by the Bell System's T technologies are making it possible to more effi- Carrier program. We are now installing T-1 digital ciently use the existing standard voice, base group, carrier systems to meet growing communications and super group carrier channels for data com- needs both voice and data. These systems, by munications. The data communication switching using pulse code modulation, can transmit up to 1.5 systems of the future are expected to be a hybrid megabits per second or 24 simultaneous voice calls combinationofacomputercontrolledcircuit over two regular twisted cable pairs. We are adding switch and a message store and forward system. these facilities wherever the needs warrant it. To- Efficient time division multiplexing systems can day these digital systems are available only in be extended to perform concentration and switching, selected locations, but eventually a large part of obviating the need for complex, relatively unre- our plant will be of this type. Even higher speed liableelectromechanical hardware inthedata digital systems are now being developed. A 281 communication exchanges. The store and forward megabit per second T-type facility is planned for operation, depending on the speed and length of 1971." (Quirk, 1967, p. 521.) message, might involve storage for one bit or one 3.17"Scientists view the millimeter wave guide or two characters, in which case the system will as another interesting potential means of trans- function as a circuit switch, or it might store a mitting data and telephone conversations. The wave block of characters or the entire message itself. guide is a precision-built hollow tube capable of Programmable common control equipment can carrying a wide spectrum of radio waves. Its po- be used to flexibly provide the convenience and tential for data transmission is great, but the devel- load leveling services." (Kaplan, 1968, p. 129). opment of this type of transmission channel is still 3.18"ITT World Communications is working some years away." (Gentle, 1965, p. 93.) on a high speed overseas voice-data communica- 3.17a"The transmission technology has devel- tionsnetwork.Speedsupto9,600bitsper oped over the years in areas associated with the second..." (Data Proc. Mag. 9, No. 6, 60 (1967).)

71 "Experimental work is under way on the trans- Wind by transmitting it from Washington to New mission of more bits per second over existing York within a few minutes." (Fussier, 1953, pp. long-haul coaxial and microwave systems." (Pierce, 225-226.) 1966, p. 154). 3.20"Another trend emphasizes the develop- "The ADEM concept has been successfully ment of new types of communication channels. demonstratedfordata modem operation over These channels not only will have the advantage of schedule 4B quality channels at data rates of making higher speeds available, but also will reduce 2400 and 4800 bits per second. The modem error the cost of data communication channels. One sys- rate of better than 1 in 107 at 4800 bits per second tem currently being developed will be capable of represents a two order of magnitude performance transmitting 220 million bits per second. This ex- improvement over presently available modulation tremely high rate of transmission is more fully techniques, most of which require manual equal- appreciated when one realizes that such a speed ization of the channel prior to data transmission. would make it possible to transmit the entire con- Manual equalization of a channel is a lengthy tents of Webster's New Collegiate Dictionary in and cumbersome process which is not at all amena- less than one second. Of course, there are few users ble to the practical operational environment and who would generate a volume of data sufficient to does not compensate for variations in channel justify this speed. But there would also be benefits characteristics during data transmission. ADEM, for smaller users. This new type channel would on the other hand, provides initialdata equalization mean that more data messages could be carried on automaticallywithinseconds and continuously one channel with no increase in circuit cost. It maintains this equalization throughout data trans- would enaDie the communications common carriers to use their channels more efficiently for common mission." (Northrup et al., 1968, p. 8). user circuit arrangements." (Goettel, 1966, p. 192; "MILGO ElectronicCorp.,receiveda $1.8 million contract from University Computing for see also Gentle, 1965, p. 93.) aspecializedversionof itsnarrow-band high 3.21"The AUTODIN telecommunications speed data transmissionset, Modem 4400/48." system, developed to meet the data needs of the (Data Proc. Mag. 9, No. 6, 60 (June 1967).) Department of Defense, has been called 'a unique "General Electric Company has announced its marriage between data processing devices and full-scale entry into the digital communications communicationsfacilitiesinacomputer-con- business with the unveiling of eight new data trans- trolledsystem.' AUTODIN representsa joint mission products. The ` DigiNet Series' is the trade effort of Western Union, RCA, Teletype Corporation name for the broad new family of solid state data and IBM; the latter three were selected as sub- setsto be manufactured and marketed by the contractors to furnish equipment to Western Union. GeneralElectricCompany'sCommunications This equipment in AUTODIN switching centers Products Department, Lynchburg, Va. routes data with requisite speed and accuracy "The new family of products will be used to trans- over transmission paths to the correct destinations. late computer and business machine language into Conventional wire, radio, ocean cables and micro- special radio and telephone signals for long-distance wave beam facilities are all at the disposal of the transmission, and then reconvert them at the other world-wide network." (Marks, 1964, p. 14). end. The series includes voice speed desk sets "Autodin (Automatic Digital Networks). World- which permit computers to call automatically and wide in scope, used by the U.S. Government fOr `talk' to other machines; acoustic couplers for con- defense purposes.. .Automatic switching necting portable teleprinters by telephone to central uses common carriers; also Autocon, uses voice time-sharing computers without special wiring; and network." (Brown et al., 1967, p. 20). high speed devices for sending computer informa- 3.22"Today's communications load is twice tion long distances via microwave or satellite. as heavy on the Army Strategic Communications "The product line covers speed and bandwidth Command (STRATCOM) asitwasinWorld requirements ranging from 300 to 230,000 bits per War II; about 60 million messages were trans- second. The DigiNet 100 and 200 Series operate in mitted in 1965. The STRATCOM network extends single voice channel circuits offering various options into more than 30 nations and connects with systems of speed from 300 to 2400 bits per second, simplex of the Department of Defense (DoD), Department or duplex, direct or acoustically coupled, with or of State, other federal agencies including Civil without a reverse signalling channel. The DigiNet Defense, AUTOVON (automatic voice network), 400 and 500 Systems operate at bit rates of 50,000 AUTODIN (automatic digital network), and satellite to 230,000 bits per second with means to get these systems. AUTODIN hassixswitching centers signals over single wire pairs to local exchanges in the United States; three others are planned where they can be picked up to be carried further in the U.Sthree in Europe, and seven in the within a multiplex group or supergroup." (Com- Pacific, including Alaska and Panama. The DoD puters and Automation 17, No. 1, 58 (Jan. 1968).) expects shortly to realize limited communications 3.19"A few years ago, a transmission system capabilitywiththeInitialDefense Communi- referred to as 'Ultrafax' conferred even greater cations Satellite System. The system has benefited distinction upon the complete text of Gone With the from NASA and Communications Satellite Corp., 72 experience and will be used instead of commerical 3.24 "A nationwide voice and data network systems only for unique and vital and national has been established to support Social Security securityneeds.Indiscussing DoD'sdecision and Medicare.Computer serviceshave been to proceed with its own system, Lt. Gen. Starbird interlinked with FederalTelecommunication itemizes some of the stringent factors imposed System communications networkso as to provide on military as distinguished from civilian systems: rapid transmission and processing of information survivability,reliability, flexibility,remote-area in `.this area." (Johnson, 1967, p. 17.) mobility, and security." (Swanson, 1967,p. 19). ..About 225 million words per year between 3.23"The new FBI National Crime information the [7251 field stations and Baltimore...Each System operates on line in real timeso as to comple- field station is equipped with a Model 33 (or 35) ment computerized systems already in operation or teletypewriter(s),withautomatictransmitters, plannedforlocalandstatelaw enforcement that transmit or receive at 100 words per minute. agencies." (Johnson, 1967, p. 15). The fieldoriginatedtrafficis concentrated by "January 27, 1967, marked a major milestone in the ARS Message Switching Computer and written the history of law enforcement. On that date the on magnetic tape at either Romney or Berwick, complex electronic information system knownas or both. From here it is sent via high speed, dedi- the FBI National Crime Information Center (NCIC) cated circuits to Baltimore, where itis received became operational in a testingor pilot phase. For on magnetic tape ready for input to the Social the first time local, State, and Federal government Security Administration's computer without agencies throughout the nation were linked inone conversion." (Johnson, 1967, p. 18). computerized network to serve a common need... "A nationwide network is in operationto support "Low speed leased lines are used fnr the initial the flow of information required in Medicare and phase of the system. These lines, handling trans- Social Security. This system doesnot yet provide 4 missions up to 135 words per minute, link each for the most advanced time sharing techniques, terminal agency directly to the center's comput- like time service and direct access, that has been erized file. One or more linesare assigned to each incorporated into the law enforcement medium. terminal device so that no contention factor exists. Further improvements aresure to come." (Johnson, Each terminal is able to communicate directly and 1967, p. 14.) immediately with the NCIC computer. A constant "The Social Security Administration also main- polling of these terminals is conducted, andre- tains magnetic tape to magnetic tape transmission sponses to all messages are made on a 'one for one' systems from the National Blue Cross Headquarters basis, with no buffering or queuing of incomingor to Baltimore." (Johnson, 1967, p. 18). outgoing messages... 3.25"Thevariouscommunicationsmedia "Terminals now in use range from IBM Model (e.g., digital signals, voice and facsimilemessages, 1050 and teletypewriter Model 35 devices tocom- one- and two-way video) were extensively con- puters of various makes including the RCA 301, sidered. Each medium presents itsown difficulties IBM Model 7740, and Univac 418. With respectto in cost, time schedule, and technical and adminis- the latter terminals, direct computer-to-computer trative problems." (Brown et al., 1967, xiv). interfaces exist, and the terminal computersare in EDUNET listings of available communications turn serving numerous remote terminal devices con- services show voice grade 4 kc rates at 2000 bits nected to their own systems. By using thisarrange- per second; dial network with 2400 bits/second ment, the local remote devices can communicate on private lines; Telpak A, C & D ranging from with the NCIC computer through the terminal 48 kc to 1 mc rates and from 40,800 to 500,000 computer... bits/second; "Half video" at 2mc and video at "There are presently over 260,000 active records 4.5 mc. (p. 31), the T-1 carrier with 1.5 million infile. These can be roughly broken down into bits/second for short distances suchas the local 115,000 stolen vehicle and license plate records, campus and the T 2 carrier at 6 million bits/ 95,000 stolen gun records, 30,000 stolen article second. (Ibid, p. 32). 3.26"Thedigital records, and 20,000 wanted person records. communicationschannel .. must transmit many kinds of messages including "The system is operational 7 daysa week and 22 alphabetic and alphanumeric data (teleprinter), hours a day with limited down time. Transactions computer-to-computer data and digitizedforms with the NCIC computer, entries and inquiries,are of speech, television and facsimile." (Francoet al., approaching 10,000 a day... Responses are 1965, p. 130). averaging less than 15 seconds from the time the last 3.26a"Since its inception, facsimile communica- character of the incomingmessage is received until tion has not been exploited to any great degree. the first character of the replygoes back to the Early development of facsimilewas hindered by transmitting terminal...Actual incidents have more successful means for communication and by demonstrated that an inquiry from thestreet by the lack of techniques that were not available until radio or phone to a dispatcher atan NCIC terminal after World War II. Facsimile communication today can be answered back to the street in 90 seconds." isstill secondary todigital communication, but (NCIC Progress Report, 1967,pp. 2-5). some knowledgeable people in the communication 73 field refer to facsimile as a sleeping giant. Facsimile manufactured under licenseof Faximile, Inc., originated in 1842 when Alexander Bain, a Scottish the patent holding subsidiary of Hogan Laboratories physicist, developed an electrochemical .recording "Time and Life use this 18 inch Mufax recorder telegraph. This rudimentary system functioned by to transmit last minute page proofsand corrections swinging a pendulum across metallic type; contact between theeditors and printers. with the tyre caused an electric current to flow "The most exacting use to which facsimile is over wires to a remote swingingpendulum synchro- now being put is the transmissionof pictures by nized with the first. Chemically treated paper,posi- the three U.S. news picture agenciesAssociated tioned beneath the pendulum, produced abrown Press,UnitedPressandInternational News stain when an electric current passed frompendu- Service. These services offer facsimile in place lum to paper. Bain's facsimile was widely used in ofWirephoto,TelephotoandSoundphototo the United States, but wasabandoned because of televisionstationsforthereceptionof news the rising demand for electromechanical printers, pictures... which provided a faster means of communication "The AP now used Mufax recorders, UP and and were simpler to operate. The widespread use INP use Hogan recorders manufactured by General of telegraph equipment also decreased the popu- Electric.All three use Hogan Fax paper, and larity of facsimile." (Axner, 1968, p. 42). electrolytic recording paper... 3.26b"The Westrex development program for "For its own use Western Union also has devel- this type of equipment was begun in 1956 when oped a high speed facsimile system which transmits The New York Times asked for machines to trans- telegraphmessagesbetween New York and mit a 10-page special edition from New York,N.Y., Washington, and perhaps now between other to San Francisco, Calif., every morningduring the points, at the approximate rate of one standard Republican National Convention in August of that letter-size page in 45 seconds. year. One hour a day wasallowed for transmission. "The recording medium used by Western Union Video facilities were obtained from the American is Teledeltos paper ... Telephone and Telegraph Company to test stability "One present use for facsimile has implications and continuity through 300 microwave repeaters that may interest you. Jiji Press is a Japanese for extended periods. The facilities wereeminently news agency specializing infinancial and business satisfactory and work was started on the equipment news. This organization hasbeen licensed under to generate and record thesignals." (Shonnard, Hogan patents andin cooperation with Kyodo, 1962, p. 176). the leading general news agencyis producing "Japanese newspapers are now using this high- facsimile for the transmission of news. Now all fidelity facsimile system. It is ideally suited to their the 3,000 or more ideographs that make up the needs because of the thousands of intricate fineJapanese alphabet can be utilized and users will detailed characters or ideographs that make up be freed of the stringent limitations placed on their written language." (Shonnard, 1962, p. 176). by the Japanese typewriter and the teletype ... "The equipment is now transmitting newspaper "The Federal Reserve Bank transacts millions pages from Tokyo to Sapporo,Japan, a distance of of dollars in business over fax circuits with member about 600 miles, and on May 27, 1962, The Wall banks; a large utility transmits charts and tabular Street Journal began use of this system to transmit information from plant to control headquarters; pages between SanFrancisco and Riverside, Calif. a large suburban bank with manybranches connects According to an article appearing in the newspaper's these branches for the exchange of credit infor- May 29 Pacific Coast Edition, The Journals printed mation, bank balances and for signature veri- at Riverside were exactduplicates of the Pacific fication; a department store is connected to its Coast Edition of The Wall Street Journal prepared warehouse... by conventional methods at the Journal's San Fran- "Publishers (for example, McCalls with editorial cisco plant. The two printings cannot be told apart, offices in New York and printing plant in Dayton, even on close inspection'."(Shonnard, 1962, p. 178). Ohio) will send raw copy from editor to printer, get galley proofs back by fax, correct thefaxcopies 3.27"The United States Air Weather Network and re-transmit these back to the printer, get transmits weather maps over land lines and radio page proofs, correct and returnthem all over to approximately 600locations, including some 1956, pp. 41-42). foreign installations. These maps are 18 X 12 inches a telephone line." (Crooks, and are used by the Air Corps, the Navy, the 3.28"Handwriting transmission terminals trans- Army, the Weather Bureaus, the commercial mit written messages or sketches over communica- airlines and by several miscellaneous commercial tion lines... users. This equipment is furnishedby the Times "The message originating at the transmitter is Facsimile Corporation, and the recording medium written with a ballpoint pen on regular paper. As is Times-Fax paper. the pen moves on the paper, varying tones are gen- "A similar network of about 100 recorders is erated and transmitted by means of the data set over operated in Canada. The equipment isMufax, a communication line to thereceiving data set and manufactured by Muirhead & Company, Ltd. in its associated receiving business machine. There Great Britain. The paper used is electrolytic paper the tones are interpreted and the receiving pen 74 reproduces the handwritten copy." (Gentle, 1965, a bandwidth-saving coding system which cuts p. 129.) down the transmission time. Recognition is then 3.29"Sylvania'sEducational Communication performed centrally where such problemsas multi- System Model 100 is a good example of the applica- font and hand-printed digit recognition can be tion of transmission methods to specialcases. With overcome. the ECS-100, informationmay be distributed over "Cognitronics'methodofon-linecorrection telephone lines to any terminal-equippedremote permits unidentifiable characters fromone scanner location, using graphic illustration andtwo-way to be held at the converter until corrected, while audio conversation. As the user handwrites withan processing from other readers continues. When electric "pen" on a 6" X 8" writing frame, the pen's a character is unrecognizable, it is automatically position is encoded into electrical signals and then displayed on the screen of a video unit at Cogni- transmitted over the telephone line. The graphical tronicsservicecenter." (Commun. ACM 11, information is received at the remote locations, No. 7, 532 (July 1968).) decoded, and displayed (in real-timeas it is being 3.29b"Most facsimile units presently marketed written) on a direct-view storage tube." (Van Dam arecapable of transmitting colorsasvarying and Michener, 1967, p. 201.) shades of black and,white." (Axner, 1968,p. 46). In addition, we note that: "Bolt Beranek and "A new portable telephoto transmitter has been Newman (BBN) has introduced and demonstrated developed [Rudolf Hell] with which color photo- a new Tcleputer System, a man-machine communi- graphs can be transmitted to the editorial offices cation system capable of transmitting graphical of press agencies. The dimensions of the picture. information over ordinary telegraph and telephone drum are adapted to the picture size of the Polaroid lines, utilizing novel coding and logical techniques. Land Camera. The main advantage of thisnew Bandwidth compression ofan order of magnitude transmitter is the separation of the three basic has been realized in display generation, permitting colors, and the transmission inan uninterrupted normal telephone communication lines to handle 3-stage process over telephone lines. This equipment 16 or more independent displays simultaneously. also permits the transmission of black and white Both alphanumeric and graphical information, in photographsinone scanningprocess, or the the form of lines, drawings, functions and arbitrary scanning of color photographs which have to be forms, can be communicated simultaneously in two published as black and white pictures." (Bliss, directions between theuser and the computer." 1966, p. 3). (Information Display 3, 64, Mar./Apr. 1966). 3.30"...Xerox Corporation's work in 3.29a"Cognitronics Corporation has announced facsimile...LDX, whichstandsfor'long development of a remote optical character recogni- distancexerography'...Theequipment,at tion system that solves data processing problems the output end, is much like our Xerox 914or 813 with accurate and efficient conversion of informa- copier. The input, however, may besome distance tion into computer language. The system is called away, up to thousands of miles away, if neces- Remote Optical Character Recognition (ROCR). sary ...We propose, in order to produce high ROCR uses a small desk top scanner reachedat quality, that we scan the original documentat remote locations and connected by ordinary tele- approximately 200 lines per inch. phone lines to a powerful central recognition unit "The LDX system ispresently designed for capable of converting multifont data to punched full-sizeoriginalsat the transmitting end, but it is a straightforward design job to providemeans paper tape, magnetic tape, or punched cards. to accept the input in the form of microfilm." "The system visually signals recognizable char- (Mayo, 1964, pp. 77-78). acters and permits their manual insertion while "One of the projects at the University of Cali- data from other readers continues to flow. Computer fornia'sinstitute of Library Research involved program validity checks of the document being read the use of Xerox LDX facsimile transmission eliminates the search through the input fcr incorrect data. equipment in library applications. Morehouse & "The entire system operates under computer Shoffner report that'the LDX system appears control in a timesharing mode. This enables all tobe capable of providing rapid,high-quality operations such as facsimile transmission, recog- transmission...from onelibrarytoanother, nition,errorcorrections, with elaspsed time for each transaction averaging validity checks, data 21 hours instead of the weekor more now typically formation, as well as output creation to flow without required' although they find cost to be prohibitive time delay. at current service levels. The authors find great "Images of characters on a documentare trans- potential for LDX to expandaccess to existing mittedtothecentral recognition system over large central collections of serials, by establishing regular telephone lines. Theuse of a laser light an LDX network that would serve several smaller source in a high resolution scanning system assures libraries, each of which would haveto retain small quality document facsimile in reliable character collections of only the most frequently used titles." recognition. (Markuson, 1967, pp. 277-278.) "The remote unit sends a line by line analog .the adaptation of Long Distance Xerog- signal to the converter of what is scanned, using raphy .(LDX) to computer input/output. LDX takes 366-107 0 - 70 - 6 75 bit-by-bit raster scan information (131 lines per handset...claimed that an 81. x 11 inch letter, inch) and converts it to graphical information on on any kind of paper, can be transmitted and standard paper, at a maximum rate .'f 768 lines ofreceived in six minutes." (Electronics 38, No. 24, characters a minute. The main advantages of this 26 (1965).) output system are that no limit exists on the char- "Plans call for the use of Magnafax machines, acter set and that graphical information can be manufactured by Magnavox, with transmission printed directly. However, the low speed slower by telephone circuit between the University of than a high-speed impact printer and high cost Nevada library at Reno and the University of ($550 per month for the printer and $1050 per month California library at Davis. In a second phase for the computer adapter) make it economically of the experiment, transmission between Reno feasible only in special cases." (Van Dam and and the University of Nevada's branch at Las Michener, 1967, p. 191.) Vegas is planned. "In order to explore advanced facsimile systems "The experiment is expected to yield information operation and to provide the corporation with a very on techniques and on the quality of reproduction rapid communications system, Xerox has installed of material transmitted and its adequacy as a a telephone-dial-controlled, 10-station LDX network substitutefortheinter-libraryloanof books, in its own plant. The network spans a third of the periodicals, and other material." (Lib. Res. & Tech. continent, tying together New York and Chicago." Serv. 9, 461 (1965).) (Bliss, 1966, p. 9.) 3.32a"Stewart-Warner was the first company 3.31"High-Speed Facsimile. The high-speed todevelop and market a facsimile system to transmission of graphics is now practical with the operate over switched voice facilities using the use of wideband communications service. Con- Bell System 602A Data Set. Since it is now pos- siderable work is also under way to enable the trans- sible to 'dial a picture' by either long distance missionofhigh-speedfacsimileoverregular or local telephone, at regular voice rates, this voice-grade channels. Although a technological system has been widely accepted and is especially breakthrough is first required it is conceivable that attractive to subscribers of the Bell System WATS documents and pictures will someday be trans- service." (Bliss, 1966, p. 6). mitted over such channels in less than one minute "The Stewart-Warner Dial Datafax equipment per page." (Gentle, 1965, p. 92.) includes the Dial Datafax transmitter and receiver; "One promising development is Western Union's both devices are specifically designed for use on Broadband Switching Network, which providesa the public telephone network. Both units employ dial-up broad-band system on a toll basis that vacuum-tube construction and require a minimum services 39 major cities." (Berul, 1968, p. 31.) warmup period of 30 seconds." (Axner, 1968, p. 52). 3.32"Xerox's Magnafax Telecopier, another facsimile device using normal telephones for trans- 3.32b "A specialversionoftheMuirhead mission, is portable, since it contains an acoustic K-300 automatic photofacsimile receiver has been coupling mechanism like those described previously. developed for use by the National Aeronautics Because it is only slightly larger thanan office type- and Space Administration (NASA) in the reception writer and weighs less than 50 pounds, itcan be of APT pictures from Nimbus and TIROS weather carried easily. 8y,11" documents can be trans- satellites." (Bliss, 1966, p. 4). mitted in six minutes. The copier isa continuous- "Muirhead and Company, Ltd., hasa ... scanning facsimile transceiver; light reflected from device, called the Mufax Courier 500, that can the document is picked up by photocells andcon- transmit an 81"x 11" page in approximately two verted into frequency-modulated audio that is trans- minutes." (Berul, 1968, p. 31). mitted through ordinary telephones. At the receiving "The prime conclusion to be drawn from the end, the document isreproduced by pressure large number of line tests carried out is that page changes of two mechanical Styli that make contact facsimiletransmissionisaworkablesystem with a special carbon-backed paper. Thispaper- over all the common forms of 'group' telephone pecking method of reproduction, while notas circuits...The performance of exchange reliable as xerographic copying, is quite inexpen-lines as local ends for facsimile working in the sive." (Van Dam and Michener, 1967, p. 200.) relatively high frequency band employed is also 4. ..A Magnavox-developed system is about to encouraging. Most of the prospective users of be marketed by Xerox for remote facsimile repro- page facsimile equipment are situated within a duction via acoustic coupling to ordinary telephones. short distance from the local carrier terminal." You place your telephone receiver in a cradle next (Phillips, 1962, p. 13, 16). to this Magnafax machine; it operates on audible 3.33"In the first place, the experimental use of beeps ata reproduction rate of 6 min/page." a telephone combined with a picture service is being (Herbert, 1966, p. 37.) explored in Picturephone. Here, using a 500 KC "A portablefacsimile machine thatdelivers band, a 275-/line TV picture is being sent between the message over any distance by telephone.. selected locations. Such a unit could be useful for the Magnafax 840....The paper is electronically the retrieval of printed documents, provided a scanned and its contents converted intoa series restricted field of view is employed to limit the of tones that are pickedup by the telephone amount of information sent in any frame, thus 76 making itpossible to obtain suitable resolution. "The Alden ALSPEED 18-inch systemcan be The realization ofa completely satisfactory service operated over dial telephone networks, 3-kc voice of th:q type, however, awaits the arrivalof a switch- conditionednetworks,and 16-kcbandwidth ing system to provide serviceon a common user networks such as the Western Union Wide-band basis." (Wier, 1965, p. 107.) or the AT & T Telpak A-3. At the high speed 3.34"Facsimile equipment normally requiresa of 960 lines per minute, itcan be operated over Schedule 2 telephotograph channel. Fora small a 32-kc band similar to the Western Union Wide- additional terminal charge, thesame line can be band circuits or the AT & T Telpak Aor A-2 ... used alternatively for voice messages." (Alden,This equipment has been operated both land-line 1964, p. 11.) and microwave circuits. A special feature of the "In recent years, facsimile hasbeen directed system's flat-bed scanner is the expandable head toward business applicationsas an important alter- that accepts drawings up to 60 inches wide and native to data communications. A keyto the future scans 18 inches at each pass.... success of facsimile in the business world is the "The 35-mmmicrofilm ALSPEEDscanners Bell System 602 Series Data-Phone Data Setswhich transmit directly from 35-mm microfilmaperture enable the public telephone networkto be used for cards to ALSPEED 18-inch recorders whenboth low-speed facsimile communication.These units are operating at the same system speed. The were introduced by AT & T in 1962." (Axner, 1968, line advance on this recorder is 96 linesper inch, p. 43.) and the comparableor corresponding advance "Western Union has also developed itsown fac- of the microfilm is 1360 linesper inch.... simile equipment employinga special dry recording "The Alden 16-mm microfilmscanner has the paper called Teledeltos. Teledeltos paper resulted capability of looking atany one of 2,000 frames from Western Union investigationsto produce a on a 100-foot reel of 16-mm film. The imageon recording paper that didnot require processing the selected framecan be recorded on a facsimile after recording. In additionto AT & T's facsimile scanning head for transmissionto any distant channels, Western Union inauguratedfacsimile point for printout onan Alden 11-inch facsimile service for photographs and telegramsover its own recorder." (Bliss, 1966, p. 2). circuits. Both common carriers,at present, offer wirephoto service for the "Further complicating the picture is the virtually press in addition to voice- limitless choice of input/output equipment (terminal band and broad-band facsimile channels forbusi- ness use." (Axner, 1968, p. 43.) devices) available to theuser. Devices most com- 3.35 monly used are those which handle perforated "An experimental pulse code modulation tape, punched cards or magnetic tape. But, com- system that transmits 224. million bitsper second puters are 'talking to each other' in increasing over coaxial cable has been developed at Bell Tele- numbers nearlyallthird phone Laboratories. Thesystem converts television, generation machines voice, and data signals into have this capability. Display/keyboard terminals, a stream of digital pulses in which data appears on a TV-likescreen, are capable of being transmittedover transcontinental becoming more common. Facsimile transmission distances. Signalscan be taken from the digital is winning acceptance, although devices for this stream and new signals can be addedas desired along the route. purpose are not numerous. Xerox, Stewart-Warner "Designed to handle various and Alden Electronicsare in the market now. combinations of BruningDiv.of Addressograph Multigraphis signals, the experimentalsystem can handle the poised to enter it, as is Electronic Transmission transmission of 144 T1 signals, equivalentto 3456 Systems Inc. Transmission of microfilm images voice channels. (T1 is the Bell System'scommercial also is possible. Alden and Eastman Kodak have .pulse code modulation system, which has a capacity jointly developed a system for thatpurpose, and of 24 voice channelson a 1.5 million-bit-per-second 3M has one under development." (Menkhaus, pulse stream.) Alternately, 46 T1 linesignals, one 1967, pp. 30-31). television signal, anda "master-group" signal, com- prising 600 frequency multiplexedvoice channels, 3.36a"Alden Electronic and Impulse Recording may be simultaneously transmittedon the 224 - Equipment has developeda number of facsimile million- hit -per- second line." (ComputerDesign 5, units that can be linked directly toan automated 20, Feb. 1966.) microfilm retrieval system. The transmission speed 3.36"Microfilm Transmission. A requirement of the system ranges from 4 minutes downto -.3s has developed during the last fewyears for trans- minute for an 8,y,11" page, depending upon the mitting images thatare recorded on microfilm type of communications channel used." (Berul, records. Work has already progressedinthis 1968, p. 31.). area, and a few terminals of this type are available 3.37"Using TV, microfilm may be viewed at dis- today." (Gentle, 1965,p. 92). tances of several feet or several miles, for use in "Facsimile scannersare available which operate both airborne and fixed station environments. The with microfilm chipsor slides; hence can be used Microteleviser acceptsany film image, magnifies it to transmit microfilm records and reproduce them to the desired size, and when coupled with an in enlarged form at remote locations."(Schatz, automaticmicrofilmretrieval system allowsa 1967, p. 3.) remote operator to select a microfilmed document 77 and view any area at the magnification desired." "In the field of education, educational television (GPL Division, General Precision Inc., press release is leading the way in the development of nationwide of August 24, 1965.) systems." (Johnson, 1967, p. 14). "General Precision, Inc., Pleasantville, N.Y., has 3.39a "CCTV can be provided with two-way been working for a number of years on equipment audio-video multiagency use of long lines between designed to transmit microfilm images by closed points would then provide TV circuitryto be circuit television. Earlier equipment allowed a user usedforhighwaysurveillanceand education to view remotely any portion of an aperture card between insitutions or schools, crime detection, image with a wide choice of magnifications.. The law enforcement training, disaster damage assess- latest version shown at the 1965 Western Electronic ment, etc. Only organization of agency "rnel on Show included the GPL TV Printer which allows line is required. The night hours might well provide the user to take a hard copy of whatever he is view- Freeway Rest Stop surveillance or on highway ing on the screen." (Veaner, 1966, p. 208.) point watch. Criminal suspect viewing by picture 3.38"Transmission devices can now be linked could bring the far near. Live, direct or taped directly to microfilm retrieval systems, and thus picture for delay send or file requires only line provide a hardcopy alternative to closed-circuit time scheduling. Inexpensive camera and video television.The Alden/Miracode system, which tape records might be in every law enforcement integrates Alden's Alpu:-Fax facsimile system with vehicle. Ready pictures of damage or accident Kodak's Miracode automated microfilm retrieval could be directed to central office at any time. system, scans documents in the microfilm viewer The actions of a law officer at the point of a stopped, and transmits the information via telephone lines on highway, suspect could be video recorded from (at a rate of three minutes per page) to make hard- patrol car, on every case if desired. Unused closed copy at remote locations. In this manner the re- circuit TV time on long line could transfer any trieval and transmission of documents is made action for file or staff viewing. The door is open almost completely automatic. The Department of to many possibilities." (Penterman and Casey, Defense's Engineering Data Systems Project uses 1967, p. 88). Alden/Miracode to make engineering drawings and 3.40" `Telereferenceisa name, coined by component descriptions in the central library avail- the Council, to designate a system for consulting able at remote locations. The Army Materiel Com- card catalogs by television...[but] less expensive mand's helicopter repair ships also use this system equipment is necessary..." (CLR 2nd Annual to ensure realtime distribution of parts lists and 1958, pp. 13-14). drawings from a central microfilm storage library "On October 2, 1957, a proposal was submitted to the repair shops within the ship. An entire instal- to the Council on Library Resources by a manu- lation of both the Alden Facsimile transmission facturerofclosed-circuittelevisionequipment system and the Miracode retrieval system costs for design and construction of a prototype, remotely under $100,000; a modification of an existing Mira- controlled,catalogcard viewingsystem. The code system costs less than $65,000. Receiving units proposed`Telereference' system wouldpermit can be furnished for $90 permonth, plus 25c for an a researcher to view uatalog cards in a central 8P x 11" page. The Alden system can also be catalog from a remote location by the use of closed- easilyinterfaced with other automatic retrieval circuit television and a remotely controlled card systems." (Van Dam and Michener, 1967, p. 200.) manipulator." (Bacon et al., 1958, p. 1). 3.38a"Fifteen...universities with television "The University of Nevada Library has received productioncapabilitiesproducemedical video a grant ...from the Council on Library Resources, tapes. Video tapes are transmitted to scrambled Inc., for an experiment in the library application TV receivers at scheduled times for the remote of telefacsimile..." (LC Inf. Bull. 24, 389 (1965).) physician." (Mayeda, 1967, p. 9.) "The Council has made a small grant to the 3.39"Smart etal., have reported on the ex-Institute of Library Research, University of Cali- perience at UCLA School of Medicine with encoded fornia, to enable it to plan an experiment with two-way medical television used for continuing telefacsimileincomparisonwithinterlibrary education of the physicians. in 15 hospitals. By loan." (CLR 10th, 1966, p. 49). questionnaire, the participating physicians agreed 3.41"...The Houston Research Institute... that there was a good-to-excellent future for tele-has begun a study to determine the usefulness visioninpost-graduate medical education. The ofexistingfacsimiletransmission systemsfor authors claim for their method 'most of the advan- communicatingtechnicalinformation... Use tages of closed-circuit television [privacy of medical will be made of the extensive microwave trans- communication and an audience which can be missionnetwork employedbythepetroleum measured,evaluated,andcontrolled],aswell industry for the conduct of its business..." as the cheap, potentially wide coverage of open- (Scient. Inf. Notes 6, 4, 11 (1964)) circuit television.' They note that the development 3.42"As a prelude to these networks, the New ofinexpensivevideotaperecordspermitsthe York Times announced an experiment using fac- hospital to replay the programs whenever they simile transmitters that link the New York Public wish." (Spring, 1967, p. 328). Library with the State Library in Albany. Before the 78 end of the test in July 1967, 25 libraries will have author) was implemented by personnel of the NBS been included. The New York Legislature allotted in 1953-1954, wherea cable link was made between S700,000 to the test. Library patrons could request the SEAC and the DYSEAC." (Nisenoff, 1966,p. through their local library up to 12 pages of material 1824.) at 25c per page transmitted from any of the partici- 3.47"Anexperimentconducted between pating libraries. The experiment was part of New Buenos Aires, Argentina, and Camblidge. Mass., York's reference and research resources (the "three has shown that a large, complex computer can be R's") program. This plan, as described by Nyquist, operated by a distant user via a radio teletype includes the development of a coordinated state- link.Civilengineersfromthe Massachusetts wide reference program that will include the aca- Institute of Technology and from the University demic and research libraries in the state. It also of Buenos Aires carried out the unique experiment, includes various phases of library automation, and using commercial radio teletypefacilities from a study of the research libraries which have subject BuenosAiresto RCA Communications, Inc., strengths that could be incorporated into a state- in New York City. There, the channel was patched wide network." (Hammer, 1967, p. 408.) into the Western Union Co.'s commercial Telex 3.43"Each 'on-system' group was connected system to put the researchers in Buenos Aires with the Project Staff at the John Crerar Library into direct contact with a time-shared IBM 7094 by a facsimile system. The facsimile system was computer at M.I.T. in Cambridge. At the same leased from the Dictaphone Corporation." (Rath time, some 20 other personsalso were using and Werner, 1967, p. 59.) the computer from consoles located at remote "The facsimile system transmits a letter size stations around Cambridge and linkedtothe page in six minutes over lines leased from Illinois central machine through telephone lines. Bell Telephone Company. The original must be a "Previous experiments have been conducted flat copy, and therefore material from books and inwhichthe time-shared computer has been journals must be reproduced before transmission. used from as far away as Edinburgh, Scotland, The output from the system is on a continuous roll and Oslo, Norway. But in those experiments, of light-weight paper. Although the output paper is connections were through commercial wire and moist before transmission, it is dried automatically cable systems. The Buenos Aires experiment was during transmission and can be trimmed for perma- the first in which a radio link was used." (Com- nent -storage." (Rath and Werner, 1967, p. 59.) puter Design 4, 8, June, 1965.) 3.44"What is required here is a machine with a "Internationaldatacommunicationsprovides reliable automatic page turner, which would trans- a plethora of opportunities. One dramatic example mit hundreds of pages, perhaps, in a few minutes." of what already is being done is the link-up of stock (Fussier, 1953, p. 226.) brokerage officesin Hong Kong with a master "The essential feature that is missing from cur- computer center operated by Ultronic Systems rently available machines is the ability to scan Corp. in Pennsauken, N.J. Information on American directly from the pages of a bound book. Fortunately, stock and commodity exchanges istransmitted several manufacturers are making progress in this to Hong Kong from New Jersey via the Ultronic area and new developments are forthcoming." satellite computer in Montreal, Canada, through (Schatz, 1967, p. 19.) transpacific cable. If there were limitations to the CC .The inadequacy of available page-turning amount of international traffic that could be handled equipment has already been found to be limiting by carriers, these will undoubtedly vanish with in the case of the interlibrary televising of book the use of earth satellites in microwave channels materials in the University of Virginia project..." and the addition of new transoceanic cables." (Council on Library Resources, 2nd Annual Report, (Menkhaus, 1967, p. 37). 1958, p. 24.) 3.48". .An inter-agencyrapidcommuni- 3.45"In 1940, George Stibitz demonstrated for cations test link [teletype-dictaphone] has been the first time the remote operation of a relay-type established for transmission of data from survey computer from a console at Dartmouth College in andresearchshipsoperatinginthetropical Hanover, New Hampshire, over telegraph lines to Atlantic...to theNational Oceanographic a computer in New York." (Commun. ACM 9, 49 Data Center..." (Scient. Inf. Notes 8, 2, 13 (1966). See also, Remote Control of the Computer, (1966).) letter to the editor by T. C. Fry, Elect. Eng. 73, 3.49"Informatics Inc....will implement an 297-298 (Mar. 1954).) AutomatedInquirySystemunderacontract "The Bell Telephone Laboratories operated some awarded by the Department of State. This system of their early relay computers with remote consoles, willoperateinanonline,multiprogramming and so deserve the credit for the first 'time-sharing' environment and will enable Department officials system," (Samuel, 1965, pp. 4, 8.) stationed in various cities in the United States, 3.46Stevens, 1955. Figure 13 illustrates the Mexico and Canadatointerrogatespecialized general scheme of what was probably the first informationfilesmaintainedonlarge,direct multiple-access, time-shared program to be demon- accessstoragedevices."(Commun. ACM 9, strated. "The first computer-tie-up (known to this 401-402 (1966).)

79 DATA DYSEAC STARTS PROCESSING SHARED BY TWO COMPUTERS - L, RI SECTION 15r 4 FILE .., 14,...... o, SIGNALS L1 TO START SEAC L _, _1 PROGRAM 0{; CONTINUATIONr ; 01I 01111 1111111111 I SEAC STARTS PRESET II r-e-411 INTERRUPT I1 I1 REVERSE DISTRIBUTION HALT 1 v- I t I RESET RI,ROALL THIS SET? t YES NO SET RO RO ALL SIGNAL //gni ofitivb=,, RO FILE S/C4Ntilf- trifearibl FILE j TALLY YES 1 ' COARSESORT RI ITEM eNNENNINAST ITEM? NO 1 YES I ALL FILE READ, 1.611.111LE riff NEAT 1 READY FOR DYSEAC NO am am men 1.0 me mem well RO' FINIS 1 Writ I YES TOCONVERT DY$EAC orA TO DYSEAC r iimmum40 TALLY SOS MIN MI MI MI UM 1ER NMI Inn 11111 3 3.50"On December 15, 1965, United Air Lines foreach savingsaccount.Proper information contracted for what is believed will be the largest on- isvisually displayed to the teller and updated line computerized information system in the busi- by the teller within seconds, thus providing new ness world. It will bebuilt around a centralized and improved customer service. Branch office complex of three Univac 1108 computers that will tellers are served from the home office by a data- be accessible by some 2000 cathode ray tube input/ phone link." (Bowers et al., 1962, p. 101). output devices on a nation-wide basis with one 3.53"Metropolitan Life is operating what could second response times. The system will be designed ultimately become the world's largest commercial and built by the Univac Division of the Sperry Rand informationprocessing network, a $15million Corporation, and represents an overall investment system ...willlink900 fieldoffices witha of $56 million for United." (Porter and Johnson, Honeywell H-1800 in New York and one in 1966, p. 74.) Canada..." (Data Proc. Mag. 9, No. 6, 3 (June "The airline industry was among the first to use 1967).) real time processing in conjunction with a national "One of the largest installations now serving communications network. The distribution, volume, business isin use by Hardware Mutual-Sentry and pace of their business left them little choice. Life Insurance group. The company is speeding The main function of the system is to process reser- transmission of information from 34 branch offices vations. The result is that any reservation clerk can to its data processing center inStevens Point, provide accurate information on the availability of Wise.,providing faster and better serviceto a particular flight and confirm a reservationfor that customers. Hardware Mutual reports that Data- flight within seconds. The hub of their system is a Phone service is saving the firm more than $1 centrally located digital computer. The computer million annually in operating costs.... is accessed by one or more multipoint voiceband- "An even larger system of data transmission is width lines. Each point on the line has a high-speed now being developed forMetropolitan Life Insur- data set and an interchange terminal. The inter- ance Co. This titan has contractedwith Honeywell change terminal interfaces the high-speed line to Electronic Data Processing for a setup to link its many telegraph lines. A teleprinterdevice termi- WellesleyHills,Mass.,headquarterswithits nates the telegraph lines at each reservation counter more than 800 district offices. and serves as the input/output device." (Balkovic, "Byprovidingon-linecommunicationsfrom 1967, p. 155.) each district office directly to its central electronic 3.51"In the aerospace industry, load sharing office each day, the system is planned to permit among a company's computer centers can generate MetropolitanLifetomaintain up-to-the-minute veryrarge daily transmission volumes, insome cases premium, loan and insurance records of its 29 on the order of 50 reels ofIBM low-density mag- million policyholders." (Marks, 1964, p. 16). netic tape, or more." (Perlman, 1961, p. 209.) "Met Life's 800 district offices are being equipped "NAA (North American Aviation). Used for load- with desk-size communications consoles consisting balancing via broad band microwave of multiple of:a Honeywell optical codereader, a control computers (7094's). Transmission is at the rate of and adapter unit,a Teletype teleprinter (with 62 thousand ch/sec." (Brown et al., 1967, p. 21.) page printer, keyboard, paper tapereader and "Organizations that have two or more computers punch), a Tally high speed paper tape reader, in geographically separated locations may find it and a data set. District offices in the Boston area advantageous to connect them by means of com- have been phased into the system and new offices munications links. This permits more effective are being added at the rateof three a week." utilization of each of the interconnected computers (Menkhaus, 1967, p. 34). because the slack time in one computer's schedule 3.54"Oneofthemostambitiousshared can be used to help smooth out thepeaks in EDP systems was established by a group of hospi- another's. Reliability is greatly enhanced because tals in Minnesota in cooperation with the Minnesota the communications links make it easy for one or Hospital Service Association (BlueCross).. more computers to take over anothercomputer's private telephone lines....model 35...tele- workload when a breakdown occurs. The data flow typewriter...Data Phone sets...Honeywell pattern in this class of application is bidirectional; H-200 communicationscomputer."("Hospitals input data and results are transmitted between each Share...", 1967, p. 13). pair of interconnected computers, and the volume 3.54a"U.S. industry is recognizing, at an ever- of data flow depends upon their relative workloads increasing rate, the advantages of company-wide at any given time." (Reagan, 1966, pp. 22-23.) data communications networks and of the closely 3.52"Systems operated by different companies relatedconceptsofreal-timedataprocessing could be interconnected in such a way as to appear and integrated management information systems. totheir customers as an integrated computer Although only about 1 per cent of the computers network. Commercial banks form at present a sold in 1965 were linked to a data communications network of this kind." (Fano, 1967, p. 35). system, Western Union has predicted that 60 "A central 40-million bit memory stores unposted per cent of the computers likely to besold in 1975 interest, customer balances, and other information will be so linked. A. T. & T. expects that the 81 volume of information transmitted in the form form may be relayed throughout the world. When of digital data will eventually equal the volume the message switching systemisused, many transmitted by voice." (Reagan, 1966, p. 20). advantages may be obtained, some of which are: 3.55"ITT Worldcom has won FCC approval for Economic utilization of communication facilities a computer-based switching service thatwill sort, Automatic transfer of messages through the store, and route messages and data between the network U.S. and other countries for subscribers to the Effective utilization of communication facilities company's privately leased international circuits ... operating at different data rates, with different Rate is one cent per 220-character message. Center codes, and with different coordination proof the system, called Automatic Retransmission cedures Exchange, is a dual ADX 7300 computer at World- High degree of message protection all with com headquarters in New YorkCity. Like Western Building block expansibilityto provide in- Union, the ITT subsidiary has long-range plans for creased service as required. offering computer services through the network." (Datamation 12, No. 11, 88, Nov. 1966.) "The design features of the system are presented 3.56"Although graphic transmission may be by first describing the configuration of a com- sent on a wide range of frequency bandwidths, with munication network, then explaining the basic the lowest and slowest a 3-kilocycle voice band- switching center functions, outlining them in system width, Mr. Stafford reported that a technological design considerations, and finally reviewing the breakthrough in the near future might increase the flow of messages throughatypicalswitching rate of graphic transmission on voice communica- center." (Genetta et al., 1960, p. 365). tion lines by two, four, or even eight times its 3.58a. "The Dalcode is a multiplex system in present rate." (LC Inf Bull. 25, App. 288 (1966).) operation for over a year. It is a Western Union 3.57"When wide band communications become acronym that stems from the equipment use as a available on a practical basis, the concept of many data line combiner and demultiplexer. Similar interconnected computers as a single multiprocessor types of equipment have been produced by other will be even more appropriate." (Clippinger, 1965, manufacturers.Basicallythesystemaccepts p. 209.) asynchronouscharactersfrom TTY units and "With modern communication linksintheir higher speedterminalsandmultiplexesthem present accelerating stage of development, it cer- onto one synchronous data stream.It may be tainly appears that remote multicomputer networks regarded as a wired program type store and forward will be common by the end of the decade." (Wagner system that operates on a single character per and Granholm, 1965, p. 288.) input line rather than on a message basis." (Kaplan, "Over-all, the outlook for computer communica- 1968, p. 128). tions complexes is one of increasing importance "The characteristics of the communication for to the realization of national objectives and to the interconnected computer systems are such that the effectiveness of government at Federal, State and information will be transmitted in real time with a local levels." (Johnson, 1967, p. 21.) minimum of delay and it shall have characteristics 3.57a"The distinguishing characteristic of a of store and forward systems since it deals with message-switching network is that messages are message flow. Ideally, the system should have the stored and forwarded. The originator transmits a best features of both message and circuit switching message into the switching center athis own speed, systems: negligible waiting time that favors the high with no waiting for a direct connection. As soon as priced human, and an efficient message interface this input transmission is completed, the sender is tofavor the highpriced computer." (Kaplan, free to send other messages, thus realizing maxi- 1968, p. 128). mum utilization of his input channel.Meanwhile, 3.58b"The ability to handle several kinds of the network takes on the responsibility of relaying terminalequipment makes message switching the message to its destination when channels be- particularly valuable in the integration of existing come available." (Shafritz, 1964, p.N2.3-1). systems into major networks. Subscribers are not "Present-day 'hard copy' written-text military limited intheir communications, nor is a high- communications networks are slow-speed store-and- speed device ever forced to keep step with a slower forward systems. Long-time intermediate storage one. Finally, the techniques of store-and-forward is used at the switching nodes to improve high-cost service insure against interruption of service or loss long-line-circuit usage. When the traffic volume of messages, and facilitate monitoring of the switch- arriving at the intermediate switching center from ingservice andevaluationofitsoperation." the many feed points is greater than the output Shafritz, 1964, p. N2.3-4). circuit can handle, messages must be backlogged." 3.59"Television offers some very interesting (Baran, 1964, p. 5.) possibilitiesforspecialized communication... 3.58"AutoData is an automatic, fully transistor- Most specialized use of television has been in the ized store and forward message switching system. form of 'closedcircuit'transmission,utilizing It is designed for use in common user communica- coaxial cables. Because of its cost, use has been tion networks in which message traffic in digital limited to very selective audiences located in a 82 limited number of broadcast centers. However, with the new equipment can be handled in 172 future developments in cable design may make it hours." (Electronics Weekly, May 26, 1965, p. 8). possible to utilize standard telephone cables for 3.62a"Noise is the great enemy of bandwidth closed circuit transmission.... compression schemes. Any device which measures "A special type of television set...is in limited picture detail must do so in such a way as to dis- use at present. It turns itself on at a pre-set time criminate between true signal and noise, perhaps on and records the program on tape...Recently, some statistical inference basis." (Cherry et al., the president of one of the major television net- 1963, p. 1508). works told me that this approach istechnically "In an automated information storage, retrieval possible right now." (Maass, 1962, p. 47). and transmission system, encoding of images by 3.59a"For example, leased lines may currently linear scanning, as in ordinary facsimile, is not carry little or no traffic at night; these lines could generally satisfactory. It may take too much time be used for the transfer of data during the slack on the communication channel or (depending on the hours at no increase in the present cost of com- trade-off of time and bandwidth) may requirea munications facilitiesother thanfor switching, very large memory for storage. Since most document termination, and terminal devices. Use of present images have considerable redundancy, code com- facilities requires close coordination between the pression can pro5tably be introduced to minimize company personnel responsible for general com- the time-bandwidth product. Although savings of the munications and the personnel responsible for data order of ten to one hundred are possible, the factor communications. The local common-carrier com- in practical coding systems is much lower (roughly munications consultants should be contacted for three to five) because signal distribution properties help in determining the ways in which existing fa- vary so much from image to image. The practica- cilities can best be used or modified for data com- bilityof code compression depends upon the munications." (Reagan, 1966, p. 24). structure of the total system, particularly upon 3.60"In another area, the trend toward higher the availability of processing logic and processing speeds will also affect the operation of some busi- buffer memory. It should be possible to develop an ness information systems. With a higher rate of optimum code compression system for each class of transmission, less circuit time per call will be re- documentstypescripts,drawings,photographs, quired. When a large volume of information can be etc. If a prescanning device could determine the transmitted over a line in a very short length of kind of document being processed, the logic could time, the minimum time period for the call is fre- then switch to the appropriate compression coding quently not fully used. In such instances users may system and much greater economy would be find potential savings in stering and batching mes- possible. Before we can design such an adaptive sages until enough data exists to fill the minimum system, we must know the types of documents to be calling time." (Goettel, 1966, p. 192). classified, develop means of automatically identi- 3.61"In any event, even with private lines there fying each class, and select the best code-compres- is still some advantage to storing or batching the sion scheme for each class." (Dodd and Wood, data at the remote terminal and keeping the final 1968, p. 60). transmission of that data under control of the central "It has been shown that for two-dimensional band- terminal. One such advantage is the possibility of limited functions, such as those imaged by an optical retransmissionof thedata when receptionis system, there are in general an infinite number of garbled. Another is the advantage of permanent different optimum scan patterns that yield data record or audit trail which the magnetic tape storage sufficient for the reconstruction of the original of the A-1200 affords at the remote site.Still picture. The scan patterns involve the sampling another advantage is that by storing the data at of the intensity and its directional derivative at remote sites and allowing the central terminal to equally spaced points along the direction of scan initiate and control the transmission, we eliminate with the scan lines themselves parallel and equally any scheduling or user query problems." (Hickey, spaced. The word optimum here means that the 1966, p. 178). area density of points sampled is a minimum." . 3.62"Instead of operating a continuous tele- (M -ntgomery, 1965, p. 206). printer circuit, traffic will be allowed to accumulate 3.63"It has been estimated that the informa- to, say, 15,000 to 20,000 words and then a normal tion content of the digital voice signal is as low as telephone call put through, probably four or five 50 bits per second. Direct digital encoding of an times a day. The messages will be sent at up to analog speech signal by sampling, quantizing, or 2,850 wpm, using phase-modulation, in sessions pulse cod.: modulation techniques requiresa band- lasting about 10 to 15 minutes. This 'burst' tech- width of 56,000 bps. With fixed channel vocoding nique is confidently expected to prove considerably techniques, the bit rate required to transmit... cheaper than the use of leased teleprinter circuits [speech] in digital form with adequate scoreson working at 66 or 100 wpm. About 25 million words of intelligibility, articulation, and speaker recognition routine plain-langb..1ges will be transmitted by Esso tests, has been reduced to 2,400 bps. There are other across the Atlantic during 1965. This would have vocoding techniques which will eventually reduce required 4,300 hours of teleprinter operation, but this bandwidth still further to 1,200 bps, and even 83 to 300 bps but with markedly reduced quality." "A conference on compressed speech, sponsored (See Franco et al., 1965, p. 130). by the Division for the Blind...The method "The use of vocoder techniques enhances coding may be used by students as a review technique efficiency and permits reduction, by a factor of 10 and by blind persons to increase their reading rate or more, of channel capacity necessary for speech appreciably when using recorded books." (LC Inf. transmission...The reduction of the required Bull. 24, 674, (1965).) channel capacity facilitates computer input/output 3.66a"The encoder operates by looking up operations and permits a larger quantity of speech segments of the input text in a stored dictionary signalsto be stored and processed in a given thatcontains commonly occurring words and computer." (Rothauser, 1966, p. 455). phrases as well as letters and letter combinations. 4 4 ...Many speech-bandwidth compressionsys- The encoder obtains the longest possible character tems have been developed, such as vocoders, match between the input text and a dictionary amplitudeor frequencylimiters, and formant entry and sends out a code designation for the coders. These machines do not recognize speech; selected entry. On the average the code designation what they do is transmit sufficient verbal cluesso can be specified with fewer binary digits than the that a human listener can piece together the lin- entry itself, and thus a compression of the input guistic content of the utterance." (Lindgren, 1965, text is obtained ...An important feature of this p. 116). type of encoding is that no information is lost, 3.64"A process has been developed in IBM to but the time required to transmit a page of text compress words or segments of speech into 500 - will vary with the amount of compression achieved.' millisecond time-slots. Words havinga time .dura- (White, 1967, p. 390). tion greater than 500 milliseconds are placed in 2or "For the present encoder simulation, the initial more time-slots or tracks." (Urquhart, 1965, p. 860). dictionary contained 400 of the most frequently used 3.64a"Results of the present simulation demon- English words and 200 commonly occurring digrams strate the practicability of a speech transmission and trigrams were taken from lists that were tabu- system based upon analytic-signal rooting. In the lated for cryptographic purposes." (White, 1967, present instance the system has been implemented p. 394). for a square rooting of the input signal, and a conse- "The simulation of a dictionary encoder indicates quent 2-to-1saving intransmission bandwidth. that a printed English compression of 50 percent is Roots higher than the second appear feasible with realizable for a broad type of English language concommitantbandsaving.7(Schroederetal., text when using a 1000-entry dictionary." (White, 1967, p. 401). 1967, p. 396). 3.65"Tests to determine the loss of intelligi- 3.67"To get a digital picture with quality com- bility of digitized voice transmissions show thatan parable to that of commercial television pictures, error rate of 2 X 10-2 is tolerable. In fact, over-all one needs about 500x500 samples and 6 bits (64 speech quality does not degrade significantly until levels) for each sample: hence 1.5x106 bits per the error rate exceeds 5 X 10-2." (Francoet al., picture. Since the bandwidth required of a channel 1965, p. 131). . increases with an increase in the number of bits 3.65a"Certain economies of storage can be one has to send through it, one would like to reduce achieved by the use of the channel vocoder mainly the number of bits needed to transmit a picture. because of the redundancy of human speech... "Schreiber's synthetic highs system took ad- By the use of this technique it is possible to store vantage of the fact that the human eye tends to reconstructable digitized speech so that approxi- emphasize edges (abrupt changes in brightness) in mately twenty-four hundred bits representone sec- a picture but is relatively insensitive to the amount ond of speech. This is a considerable saving in space of changes in the brightness over edges: on the over the minimal 24,000 bits and the desirable other hand, in areas where the brightness changes 48,000 bits." (McDonald, 1966, p. 53). slowly, quantization noiseiseasily discernible. See also note 3.63. Therefore, edges and the slowly varying parts of a 3.66"Compressed speech, a method of increas- pictureweretreateddifferently." (Huang and ing the rate of speech, and expanded speech, a Tretiak, 1965, p. 48). means of lengthening normal speaking rates were "Schreiber implemented a real-time system of demonstrated by Leo Levins of the American edge detection in one dimension along a scanning Foundation for the Blind..." (LC Inf. Bull. 25, line,run-length coding the edge Iota pionand App. I, 425, (1966).) magnitude,andthengeneratinga'synthetic "It will be possible, for example, to make the highs'signal from the edge information. The computer pronounce aloud any word upon whichsynthetic highs signal was added to a low-pass, the subject maintains fixation for more than one signal to give a good-looking output picture witha second. It would be interesting to find out whether reduction in bandwidth by a factor of four." (Graham quick feedback of this general sort can accelerate 1967, p. 337). children's learning." (Quarterly Progress Report, "In Schreiber's system...the edge informa- No. 80, Research Lab. for Electronics, M.I.T., tion was sent by run-length coding (essentially, the 425, 1966). magnitude and the position of each edge pointwere 84 transmitted). A reduction of 4:1 was achieved with high frequency part- of the picture, which is then rather good received pictures. added to the low-frequency part to form the recon- "A picture coding scheme investigated by Huang structed picture. Graham transmitted the gradient could be considered as an extension of Schreiber's information by following the contours. For each system to two dimensions...A set of basic points continuingcontourpoint,hetransmittedthe (e.g., one out of every sixteen samples...) were changes in contour direction, gradient direction, transmitted for all picture frames. These points and gradient magnitude using a Huffman code. essentially constituted the low-frequency part of the From the probability distributions of these quan- picture. In addition, extra edge points... were tities, he estimated that compression ratios of 4 sent for each frame. Whether any given point was to 23 (depending on picture complexity) could be an edge point or not was determined by a threshold achieved on 256x256-point,6-bit(64-brightness function which depended only on the basic points. level) pictures. Graham's reconstructed pictures Therefore, if the transmitter and receiver agreed on were of good quality except that some of the texture the threshold beforehand, the positions of the edge information in the original were lost. Work isnow in points need not be sent. At the receiver the blanks progress at M.I.T. to try on the one hand to increase were filled in by linear interpolation." (Huang and the compression ratio of this scheme by curve- Tretiak, 1965, pp. 49-50). [Illustration shows good fitting the contours, and on the other hand to quality at 7:1 reduction]. improve the quality of the reconstruction by the "Huang studied the subjective effect of pictorial additional transmission of texture information." noise. The main goal was in finding out how the (Schreiber et al., 1968, pp. 1, 5). objectionability of a two-dimensional low-pass noise 3.68"Other experiments aimed at reducing the depends on its bandwidth. redundancy in motion pictures, and at the optimum "Cunningham investigated several systems of coding of digital color pictures. 4 transmittingmonochrome motionpictures... "The research program of this group [Picture Besides introducing temporal filtering as a band- Processing Research, RLE, MIT] has been to in- limiting process, the systems transmitted picture vestigate schemes for reducing bandwidth required sequences by correcting a fixed fraction of the pic- for transmission. It was realized that in order to ture sample points during each frame transmission solve this problem, it is necessary to better under- time. The results indicate that a saving of 6:1 in stand how the human observe' perceives the image, bandwidth is possible using such methods." (Huang what features of the picture are important, so that a and Tretiak, 1965, p. 52). collateral interest of the group has been to investi- "ln the search to reduce the required channel gate ways of mathematically representing pictorial capacity an obvious procedure is to attempt to data, of finding out details of what features of the reduce the number of picture samples. If this picture are important." (Huang and Tretiak, 1965, reductionis accomplished by selecting samples pp; 45-46). either at random or according to some regular "If the received picture is to be viewed by hu- system which takes no account of the properties of mans, then one can take advantage of the properties the picture, then all that has been done is to ex- of human vision. Here, the purpose is to distort the change channel capacity for picture quality. While picture in such a way that the distorted picture can this is a perfectly acceptable procedure a more be described by a smaller number of bits, yet the interesting effort is to attempt to reduce the number change is not noticeable to human viewers." (Huang of transmitted samples without reducing the quality and Tretiak, 1965, p. 48). to the same extent as the procedures mentioned previously. When this is the goal the problem divides 3.69"Briefly, a run-length may be defined as a itself into three parts: the selection of the points continuous segment of the signal over which some to be transmitted, the coding procedure for trans- stated property may be considered as constant, mitting the required information about the points, within stated criteria." (Kubba, 1963, p. 1518). and the decoding procedure by which an image "This 'reduced' signal...consists essentially may be produced from the received information. of a continuous sequence of run-lengths over which In this paper, we describe an approach to this the video signal is assessed as being of 'constant' problem in which the points along the outlines or brightness (or other parameter, if desired) by the Detail Detector according to criteria built into that contoursinthepicturearetransmitted.... "The gradient approach was first suggested by detector." (Cherry et al., 1963, p. 1508). Schreiber and later simulated on computer by "Visual communication signals, generated by Graham. Here a point is considered a contour conventional scanning techniques as in facsimile point if the magnitude of the gradient at that point and television systems, make very poor use of the exceeds a certain threshold. In addition to gradient bandwidthsallocatedfortheirtransmission." information at the contour points, a two-dimen- (Kubba, 1963, p. 1518). sionally low-passed version of the picture is also "Variable velocity scanning provides an excellent, transmitted. At the receiving end, the gradient accurate means for reconstructing pictures at the informationispassedthroughanappropriate receiver. It also leads to a relatively cheap receiver." two-dimensional linear filter to give rise to the (Cherry et al., 1963, p. 1508). 85 3.70"The design and the characteristics of an a)BBBBB experimental encoder for digital transmission of video signals [using a unity bit coding methodj W W y. are described ...The experimental results show A white element would usually occur in the y that considerably good reproduction of video pic- position, following this neighborhood. tures is obtained with sampling frequency as low as 30 Mc ..." (Inose and Yasada, 1963, p. 1524 W WBWW (abstract).) b)WWBWW 3.70a"The human eye objects much more to noise with strong structure, such as quantization W W y. noise, than to random noise. Therefore, a smaller number of quantization levels can be tolerated, if A black element would usually occur in the y means can be found to transform quantization noise positionfollowing this neighborhood." (Wholey, to random noise. An example is the method of 1961, p. 99). Roberts, who added pseudo-random noiseto a "For each of the different neighborhoods, a unique picture before quantization, and later at the receiver prediction y is determined as the color which is most subtracted the same noise from the quantized likely to follow the neighborhood (on the basis of the picture. It can be shown that by this maneuver the statistical survey). We thus obtain a 'prediction quantizationnoiseistransformed into random function', a table in which each of the possible neigh- noise with the same rms value. This method gives borhoods 'predicts' (i.e., is paired with) the color of acceptable pictures with only four bits per sample." y which is most likely to follow it." (Wholey, 1961, (Huang et al., 1967, pp. 332-333). p. 100). 3.70b"Since the noise contained in the pictures "Now each matrix representing a picture is paired from the Roberts method is random and independ- (in a one-to-one fashion) with an 'error matrix' which ent of the signal, it can be reduced by some averag- indicates the elements y of the picture matrix fr.',r ing process, such as the snow-removal technique which the prediction function gives an ince' ,-ect value. In order to make possible a uniform predic- of R. E. Graham...Although Graham's technique by itself does smear the edges in a three-bit Roberts tion procedure, the picture is treated as if it were picture, with the additional information (about 0.1 surrounded by a white border two elements wide. bit per sample on the average) on edge points, a For each element y in the picture matrix, the actual picture almost as good as the continuous original color of y is compared with the color predicted by can be obtained." (Huang et al., 1967, p. 333). the function above, on the basis of the neighbor- 3.70c1"Pseudo-random scanning can also be hood of y. If the colors are the same, then a 0 is used for cryptographic purposes." (Huang et al., stored in the corresponding position in the error 1967, p. 334). matrix; otherwise a 1 is stored there... "The error matrix (which under ideal conditions "The so-called pseudorandom scanning... can be used for secrecy transmission. In pseudorandom should contain a small percentage of Fs) is now scanning, the scanning beam hops from point to coded by some process like run-length coding (which point in a seemingly random fashion. However, gives the number of O's between successive pairs the transmitter and receiver scanners are synchro- of l's).... nous, so the receiver scanner can reconstruct the "To reverse the process and obtain a picture from picture. The coordinates of the successive scanning its code number, we first obtain the er-or matrix points are specified by a sequence of pseudorandom from the run-length code. Then, the outp -t of the prediction function for each successive element of a number .Anyone who doesnot know the particular pseudorandom sequence willnot be new picture matrix is compared with the corre- able to reconstruct the correct picture even if he sponding element of the error matrix. The color pre- shouldintercepttheone-dimensionalsignal." dictedis entered in the new matrix if the corres- (Huang, 1965, p. 60). ponding element in the error matrix is a 0; if the 3.70c2"A system is described for digital en- corresponding element is a 1, the other color is coding of continuous information sources based on entered." (Wholey, 1961, p. 100). quantizing thedifference between theoriginal "Our experiment was made on weather maps, in continuous signal and a predicted version thereof, view of military requirements for efficient storage as opposed to quantizing the original signal itself." and transmission of information of this type.... (Graham, 1958, p. 147, abstract). {Average compression coefficient for 10 maps= 3.70d"Statistics were gathered on the number 0.38]. of times each of the 2'2 different neighborhoods of "Greater compression (i.e., a smaller average black and white elements was followed by y=B compression coefficient) is expected for the high- and the number of times by y= W. resolution data which might be required in a military or commercial display device, since thehigh-resolu- Examples: tion data would tend to be smoother and the proportion of errors, therefore, to be smaller." (Wholey, BBBBB 1961, p. 100). 86 3.70e"The aim of efficient coding methods is to bandwidth) are required for transmitting the error reduce the channel capacity required to transmit a samples than for transmitting the original samples signal with specified fidelity. To achieve this objec- for a givenfidelity of reconstruction." (Lucky, tive, it is often essestial to reduce the redundancy 1968, p. 550).. of the transmitted signal. One well known procedure "When noise is added in the transmission channel for reducing signal redundancy is predictive coding. there is some probability of the received digits being Inpredictive coding, redundancy is reduced by incorrectly detected by the slicer. Even though the subtracting from the signal that part which can be transmitted power might have been substantially predicted from its past. For many signals, the first- reduced by the redundancy removal, the probability order entropy of the difference signal is much of an initial error is identical to that of a full power smaller than the first-order entropy of the original system. Once an error has been made, however, signal; thus, the difference signal is better suited to the probability of making subsequent errorsis memoryless encoding than the original signal. Pre- increased because of the incorrect symbol being dictive coding offers a practical way of coding signals usedinredundancyrestoration.Thus,errors efficiently without requiring large codebook mem- tend to bunch together in the reveived data. Be- ories." (Atal and Schroeder, 1968, p. 1). sides increasing the average probability of error "Previous studies of predicitive coding systems thiserror propagation considerably complicates for speech signals have been limited to linear pre- the problems of error control in the entire system." dictors with fixed coefficients. However, due to (Lucky, 1968, pp. 561-562). the nonstationary nature of the speech signals, a 3.70g"It is intuitively evident that there is much fixed predictor cannot predict efficiently at all times. n.ore correlation between TV picture elements in For example, the speech waveform is approximately the frame-to-frame time dimension than there is periodic during voiced portions: thus, a good pre- between adjacent elements in a single frame. By diction of the present value of the signal can be using delay lines, sufficiently long enough to contain based on the value of the signal exactly one period a complete frame, and high-speed digital logic to earlier. However, the preiod varies with time, and compare and operate on elements from frame-to- the predictor must change with the changing period frame, it is possible to take advantage of this fact in of the input speech signal. In the predictive coding picture coding. system described below, the linear predictor is "Early experiments showed that simple frame adaptive; it is readjusted periodically to match the repeating could be used to reduce the rate of picture time-varying characteristics of the input speech transmission, avoiding flicker by displaying each signal. The parameters of the linear predictor are frame several times before replacing it with a new optimized to obtain an efficient prediction in the frame. Still better was replenishment of a fraction sense that the mean-square error between the pre- of the picture elements in each frame, thus avoiding dicted value and the true value of the signal is even the jerky motion characteristic of the cinema, minimum... but leaving some peculiar small area patterns "The study reported here shows that predictive around rapidly moving elements of the picture." coding is a potent approach to digital encoding of (Mounts, 1968, p. 28). speech signalsfor high-qualtiy transmissionat 3.70h"The on-line Fourier Transform is used to substantial reductions in bit rate. Unlike past speech eliminate periodic noise from long distance televi- coding methods, the predictive coding scheme de- sion transmissions by noting that this noise gives scribed here attempts to accurately reproduce the rise to a striation of .each frame and that this stria- speech waveform rather than its spectrum. Listening tion corresponds to a constant Fourier Transform, tests show that there is only slight, often impercep- independently of translationin the input plane. tible, degradation in the quality of the reproduced Sending a sequence of frames containing respec- speech. Although no detailed investigation of the tively 'noise only' and 'video plus noise', it becomes optimumencodingmethodsofthepredictor possible to subtract out the noise portion of the parameters was made, preliminarycalculations transform and to display a cleaned up version of suggest that the binary difference signal and the the incoming signal.... predictor parameters together can be transmitted "A spatial filter is located in the back focal plane at bit rates of less than ten kilobits/second, or sev- of the first lens where the Fourier Transform of the eral times less than the bit rate required for PCM picture on the crystal is formed. Filtering of the encoding with comparable speech quality." (Atal videoisaccomplished asfollows: The remote and Schroeder, 1968, pp. 1, 7). camera transmits alternate frames with its shutter 3.70f"The philosophy of predictive systems has closed and open, so that at the receiver the incoming been widely studied for its application in band- information is of the form 'noise only' followed by width compression of telemetry data and of tele- `picture plus noise.' The noise frame is inscribed on vision; for example, sec Kortman, Davisson, and the crystal which is briefly illuminated (during ver- O'Neal.Inthese examples theerror samples tical retrace) by a burst of laser light, thereby form- ek are quantized and transmitted by pcm. Because ing the noise spectrum in the Fourier Transform of redundancy, that is, predictability, in the source plane where the spatial filter is located. The filter is data, fewer digits per sample (and consequently less adaptive in the sense that it can store this informa- 87 tion (or some approximation to it) for one frame. At vices varies and so does the price. Most systems the end of the following (picture plus noise) frame correct errors through retransmission, but there are the crystal is illuminated again, and from its Fourier some which make possible error correction without Transform the filter 'subtracts' the spectrum of the retransmission; generally, the latter are more ex- previous frame. The difference is reconstructed by pensive. The potential trouble an error could create the second lens and the resulting image, picked up must be weighed against the cost involved in pre- by the vidicon, is displayed on a m..:titor." (Poppel- venting the terror in the first place." (NIenkhaus, baum and Faiman, 1968, pp. 1-2). 1967, p. 35). 3.71For example, "any attempt to increase re- 3.73"Two considerations affect the selection dundancy by coding implies an increase in channel of the most appropriate error-control code. The bit rate for a fixe.). user information rate. But, as first involves the requirements of the user, the toler- the bit rate is increased, the basic error rate of the able delays, error rates, data rates, channel effi- channel usually increases. This means that any ciency, and other factors: the second deals with the improvement in output error rate gained from the characteristicsofthetransmissionmedium." additional code redundancy, could be partially or (Franco et al.,1965, p. 125). completely negated by the deterioration of the 3.74"The method chosen to avoid dilution of channel error rate." (Quarterly Progress Report single-error-detection capability isthat of repre- No.80,Research LaboratoryforElectronics, senting each of the additional (transparent-mode) M.I.T., 195 (1966).) controls by a sequence of two characters from the 3.72"...Decoder complexity increases drasti- `wrong' parity subset. This method provides the cally as one strives to approach channel capacity required security, since a single-bit error within a (100% efficiency) for a fixed error rate. Furthermore, string of ASCII characters could falsely generate ...for a fixed efficiency, operation at low output at most only one character of such a sequence .. error rates also demands high decoder complexity. (Transparent-Mode Control Procedures..., The search for ways to overcome this equipment 1965, pp. 204-205). barrier has been a motivating force in the develop- 3.75"The gross modification of the facsimile ment of techniques such as sequential and threshold signal would very likely not be available as an off- decoding, low density parity check codes and error the-shelf feature of existing facsimile equipment detection with automatic request to retransmit in- and therefore would require some research and formation received in error." (Franco et al., 1965, development to produce the required equipment p. 130). modification. In spite of the need for research and 3.72a"Many communications equipment manu- development, itis very likely that this approach facturers can provide error detection and correction would be considerably cheaper than the use of (EDC) equipment. The sophistication of these de- electronic scramblers." (Geddes et al., 1963. p. 133).

4. Audio and Graphic Inputs 4.1For example, "the main ingredient of a at the TOUCH-TONE phone calls the computer touch-tone data gathei-ing system is the touch-tone as he would call another telephone station. The card-dialer telephone made by the Bell Telephone data set at the computer automatically, answers and Systems...Depression of a key causes tones to puts the computer on line. The attendant is now be generated withinthesetbyitssolid-state capable of keying signals into the computer. Replies circuitry. The only power required is obtained from generated by the computer are received audibly the telephone line. Tones are also generated by the by the attendant through the telephone handset. insertion of a plastic pre-punched card into the card One technique used at the computer is to generate reader, which is an integral part of the telephone voice reply with the aid of speech segments preset... recorded on magnetic tape. The computer as- "The function of the data subset [Data-Phone sembles these segments in the proper sequence to 401, 403A or 401J] is to receive the tones generated produce the required replies. Another technique by the touch-tone telephone and convert them to that can be used is computer generated synthetic relay contact closures. The relay contacts are cable- speech. This approach has been the subject of connected to the code translator through a multi- considerableresearch. Signals corresponding to pin interface." (Davenport, 1965, pp. 36-37). voice pitch, loudness, tongue position and other "There are systems utilizing this capability today. speech variables are combined to produce the con- The majority of these are of the Digital Inquiry- sonant and vowel sounds. These are generated in Voice Answerback(DIVA)type.The system the proper sequence to produce sounds that are like consists of a computer which is configured as a human speech." (Balkovic. 1967, p. 156). station on the Switched Telephone Network and 4,2"Voice inputis quite another matter. In the various TOUCH-TONE telephones that can spite of the existence of some interesting Uork. this access the computer via the network. Anattendant reviewer feels that it will be five to ten years before

88 any significant amount of computer input will be the "yes" or "no" answers to questions like these, received in the form of human speech. The intensity the computer selects the next logical question and of the work in this field is interesting and gratifying; finally chooses a pre-selected vacation package for Hogan includes 21 books, articles, and documents the customer based upon the entire range of re- in the bibliography of his survey of the voice input/ sponses. Simple as it may seem, a basic telephone output literature. Almost every entry deals with serves as the inquiry device for this system. All some aspect of speech recognition. It is Hogan's that is required is the ability to talk to and listen to opinion that we are approaching the point of having the computer." (Gentle, 1965, p. 87). a limited amount of spoken input to a computer 4.4"At the remote laboratory the scaler-counter system and that the telephone will become an im- which accepts the raw data from the scintillation portant computer terminal. If this latter prediction detector...provides an output pulse for every is realized any time soon, which is not unlikely, it 100 (or whatever scaling factor is desired) detector will be because of Touch-Tone 'dial' input and pulses...[reducing] the pulse rate to an ac- voice response, not voice input. The reason is simply ceptable rate for the telephone data-link. It has that, unlike voice-recognition equipment, devices proved convenient to use the telephone dial to trans- for 'recognizing' Touch-Tone signals are fully de- mit simple order codes and timing marks to the veloped, inexpensive, and commercially available." computer. For example, in the application being (Mills. 1967, pp. 233-234). described we dial '0' to indicate to the computer the "The use of voice input to computers is still in moment of injection of the radioactive tracer into the research and development stage. One reviewer the venous systems... feels that the state-of-the-art permits reliable rec- "At the t:omputer the pulses are picked off the ognition of digits spoken by a speaker for whom the telephone by an inductive pickup coil which clips to recognition program has prior information. To date, the phone. The pickup signals are amplified and no operational speech recognition systems have been those exceeding a threshold level (set to discrim- designed." (Van Dam and Michener, 1967, p. 194; inate against line noise) are fed into the computer. see also Gentle, 1965, p. 87 and Goettel, 1966, p. The intervals between pulses are measured by the 191). computer itself by comparison with itsinternal "For many years man has been recehing mes- clock. Actually a program loop is used as a time base sages from machines in printed form. Teletypes, and the number of times the computer executes the computer console typewriters, high-speed printers loop between pulses is counted..." and. more recently, character display oscilloscopes "If the computer is not available at the moment have become familiar in the role that they play in when data are being generated, the information machine-to-man communication. Since most com- pulses can be stored on an ordinary inexpensive puters are now capable of receiving instructions `home-type' magnetic tape recorded for transmis- from remote locations through ordinary telephone sion as soon as the telephone line and/or the com- lines, it is natural that ive ask whether with all of the puter becomesavailable."(Neilsen,1965,pp. sophistication that we have acquired in computer 634-635.) usage, we can communicate with the computer in 4.4a1Information obtained from Moore, Dr. normal speech. On the input of the computer, there Rolf, private communication, Aug. 11, 1966. is the automatic speech recognition problem, and at 4.4a2"There has been progress in research into the output, the problem of speech synthesis from the problem of optical recognition of printed text messages in text form. The problem of automatic characters (Potter, 1964) but there is real difficulty speech recognition is substantially more difficult in developing the type of flexibility needed to handle than the speech synthesis problem. Whilean auto- the various idiosyncrasies of an individual's hand- matic speech recognizer capable of recognizing written print or script. It has been shown that high connected speech from many indiviudal speakers reliability of recognition of handwritten words can with essentially no restrictionon the vocabulary is be obtained when the machine is programmed to many years away, the generation of connected deal with the handwriting of four subjects; reliability speech from text with similar restrictions on vocabu- decreases when the program is based on a larger lary is now well within our reach." (Lee, 1968,p. group, and deteriorates further when attempts are 333). made to deal with the writing of subjects, samples 4.3"Cognitronics, however, has developed a of whose script has not been included in the conversation machine that can ask questions, using programming(Mermelsteinand Eden,1964)." a Cognitronics Speechmaker, and can understand (Spolsky, 1966, p. 493). the spoken words "yes" and "no." This machine 4.4b"Whirlwind (at MIT) had a cathode ray could be used to access a well-defined and tightly tube and light pen in the early 1950's. A prototype structured data base both quickly and cheaply.... of the APT system (computer controlled machine "The computer may ask a question such as "do tool) was programmed on Whirlwind in1955." you enjoy golf?" If the response is "yes", the next (Wigington, 1966, p. 86). query might then be "Do you like the mountains?" "The light pen, which was originally used in If the answer is "no" the subsequent question may military air-surveillance systems for target identifi- by "Do you like the seashore?" Depending upon cation...has been adapted by means of tracking 89 programs to permit direct entry of hand-drawn in- operations where one looks for certain descriptor formation." (Ward, 1967, p. 48). words." 4.5"The Elastic Diaphram Switch Technology 4.8cThis includes thepossibilitythat "the (EDST) is being used in an experimental device coded tape produced in the keyboarding of manu- created by...[IBM's] Advanced Systems De- scripts may be used in all subsequent phases of velopment Division which can enter handwritten disseminating and handlingscientificinforma- information directly into a computer's memory." tion." (Kuney, 1967, p. 136). However, "untilwe (Commun. ACM 8, 718 (1965).) have progressed first of all to the graphic insertion 4.6"With Data Trend's MIMO Hand -Print promised by Photon, IBM, and others tothe Model input-output terminal for real-time on-line actual computer storage of both line diagrams and operation with a data processor, the user writes his half-tone images and their recall and production message on a platen with a special stylus. All the along with the text, when called for by the editor alpha-numerics and several symbols may be used. or makeup man, we cannot pretend that we have a Little or no operator training is required. Informa- satisfactory comprehensive automated publishing tion concerning the character written is transmitted system." (Duncan, 1967,p. xi). via a teletype code to the computer which identifies 4.9"Another data coding scheme might be it and takes whatever action is appropriate." (Data explored for application beyond its originalpurpose: Processing Mag. 7,50 (Feb. 1965.) Galli (1962) has describeda method for transcribing "Data Trend's Solid State Keyboard MIMO text into machine form using the stenotyping tech- (Man In-MachineOut) systems-orientedinput/ nique,whereindividualphoneticsyllablesor output terminal provides real-time on-line operation common words and phrases are manually tran- with a data processor." (Data Proc. Mag. 7, No. 2, scribed into a 23-bit code. The principal applica- 50 (1965).) tion is rapid data transcription using the stenograph 4.7"If a speaker's previously recorded 'voice- machine, with decoding performed by photostore print' . could be fed into the computer as the dictionary lookup technique. Butone can envision object to be found, the computer could monitor other applicationsthatrelateto storage com- telephone conversationsor search masses of tapes, pression (not necessarily text storage), where the punting out only where the conversation included steno symbols are stored and only decoded at out- the voice-printed subject.... put time." (Climenson, 1966, pp. 117-118). "It may also be possible to use phoneme sounds "Stenotype devices will bean attractive alterna- ...to enable computers to spot particular words in tive to the keypunch or typewriter for preparing oral communications, such as 'atomic', 'Pentagon', inputs for the optical readers." (Berul, 1968,p. 32). `Rockefeller', or 'Costa Nostra', and to take off 4.9a1Another type of combination isrepre- onlyconversationscontainingthesesignals." sented by the following: "Dependingon the specific (Westin, 1967, p. 87). applicationat hand, two separate, modern mechanical inputtechniques have beenusedto 4.8"A solid state, optical scanning device, the replace keypunching. IBM Scanistor, converts images into analog voltage output representing amount and position of light 1. For alphanumeric data,we are using NCR detected, with high resolution and fastresponse, 731 key-operated magnetic tape recorders. and has been suggested for developmentas "a Applications include master file conversions, hand-held 'reading' device which could be passed fileupdating, program entry andprogram over a line of printed text to enter data into a com- editing. puter." (Commun. ACM 8, 82-83 (Jan. 1965).) 2. For entry of numeric transaction data,we are utilizing optical font sales registers and adding "The scanistor (an integrated diode resistor array) machines." (Lee, 1968, p. 51). shows promise for future application in the charac- 4.9a2"The use of highly specialized fonts as ter recognition field." (Henle and Hill, 1966, p. exemplified by the Farrington scanner has been 1860). limited primarily to systems that are self-contained 4.8a"For some years the Institute for Scientific within an agency. The United States Department of Information has been working on the Copywriter, Agriculture is an example of use of the Farrington a portable device for selectively reproducing a word, process. This process is limited to the recognition line, sentence or paragraph from a printed page. of unique characters produced by specially equipped Its developers estimate that J 0% of the total copy- typewriters. It costs approximately $40 per instru- ing market would find a selective copiei appealing." ment to convert to Farrington characters. In essence (Veaner, 1966, p. 208). this process is restricted to internal processing of 4.8bThus Vander Lugt et al. (1965, p. 139) documents which originate and end in the using propose "the use of spatial filtering in coherent organization. The psychological factor of 'difference' optical systems to construct filters for the detection has been a deterrent to broader usage for alphabetic of key words in printed material. Such a system data." (Lannon, 1967, p. 48). might be used in the Post Office Departmtnt for 4.9b"input to a computer isstill one of the routing mail to given areas by using the Zip Code. slowest elements in automatic data processing. If It might also be useful in abstracting and cataloging we compare the internal speed of the computer 90 system scheduled to be installed at the National "Retyping after making cleanerasures has been Institutes of Health of my Department we find that found to be an acceptable correction methodfor the central processor is 8,000 times faster than the OCR. The use ofan erasing shield is -recommended card reader of the system. The card reader is byno toavoiddisfiguration of adjacentcharacters." means slow and processes 1,000 80-column cards (Greenly, 1966, p. 47.) per minute. ... "A simple formula for determiningthis cost "Dependent upon a variety of factors- it takes in [keypunch vs. OCR] is to evaluate: the range of 16 manhours of key punch and key verifier time to keep the card reader busy for one minute, which in turn keeps the central processor F=b+ a c busy for .075 seconds. It becomes quite apparent Where: that the input process is quite expensive and still quite slow." (Lannon, 1967, p. 48). F= number of characters processedper dollar. a = total characters processed per month. 4.10 Some examples are as follows: b= monthly equipment rental andoverhead "The monthly Bibliography of Agriculture now costs. is issued by using an optical scanner-computer c = monthly employee salary costs, including combination." (Brown et al., 1967, p. 33). supervision and fringe rates." "The Council is assisting the [Los Angeles County Public] Library to evaluate a procedure by which (Feidelman and Katz, 1967,p. 0210:28.) bibliographic records would be prepared from the "By dividing the cost of the key punchoperator beginning in a form suitable for optical scanning." into the cost of the reader the break-evenpoint is (Council on Library Resources, 10th Annual Report, to the advantage of character reading when the load 1966, p. 42.) exceeds the capacity of 15 key punchoperators." "It might be noted parenthetically that, from the (The British Computer Society, 1967,p. 155). typist's viewpoint, the simplest solution would be The British Computer Societysurvey also shows to use optical scanning. The Olivetti-Underwood that a comparison of costs of key punching cards Praxis 48 typewriter, for example, is used by Docu- and typing and optical reading (witha typical cost mentation Incorporated to provide input to an REI reduction of close to 50%) indicates "the potential scanner. The font is a standard gothic, only slightly of typing and reading to reduce considerably the modified as to descenders; the scanner does not costs of data conversion, particularly for large sense red, making forms typing possible. This volumes of mainly alphabetical content." (1967, approach entails the least disruption to the typist's p. 156). routine." (North, ]967, p. 5.) "One may readily ask: -If the input data.must "One other agency reporting a large use ofbe rekeyed, what is the advantage of rigid font optical readers in the place of other input conversion scanning?' The answer lies in the fact thatmany methods was the Department of the Apiculture typewriters equipped with normal font can be readily which uses 2 CDC 915's, fed by 25 typewriters in changed to rigid optical font by mere selection. The the home office, and numerous, a thousand or more, ordinary typewriter can then becomea substitute in the field." (Auerbach Corp., Source Data Auto- key punch device...What are the economics of mation, 1967, pp. 2-13.) the rigid font optical scanners? The typing in Fig. 3 "For large conversion tasks the OCR exceeds all could be read by a device costing $120,000or renting other methods in economic excellence,even when for $3,500 per month. Such an optical scanning including typing cost." (Moore, 1967, device is made by the Control Data Corporation. p. 95.) This device could maintain a rate of input reading "Probably the greatest interestisin optical that could remain well ahead of 40 operators keying readers, and studies have shown its high-speed difficult textual material." (Wishner, 1965, p. 240). potential (since it is not dependent on direct human "On the other hand, the typewriters used by the operation), and flexibility to formats to be advanta- catalogers could be obtained with the Farrington geous to cost reduction in large-scale operations. Selfchek font and the preparation process could be This has been demonstrated even when an initial continued just as it is now. Then, once a catalog card conversion step from source data to typewriter had been completely prepared, the card itself could input must, be made." (Auerbach Corp., Source be optically scanned and a 'magic' typesetter used Data Automation, 1967, p. 2-12.) to produce the cards. This process would also offer "The results of a comparative cost analysis economy, speed, and efficiency. The machines exist showed that keypunching would cost approximately now; nothing is stopping us except ourselves and, 3.5 times as much per yearas optical Scanning; perhaps, the funds." (Patrick and Black, 1964, p. 39.) therefore, the committee publisheda formal justi- 4.11 fication recommending the purchase of "One [of the problems that cause particular at least difficulty]...is the presence in a large proportion three optical scanners, one for BuPers in Washing- of all input documents of non-alphabetic material ton and two for Navy Finance Center..." (Auer- which cannot be recognizedor dealt with as text. bach Corp., Source Data Automation, 1967,p. 2-13.) This includes graphical material, pictures, special

366-107 0 - 70 - 7 91 )

field defining lines or formatting marks, characters portions of the text to facilitate subsequent retrieval. from strange alphabets or bizarre and exotic fonts. Some preliminary experiments, designed to expose ...Perhaps the simplest example of this kind of the problems likely to be encountered in imple- problem is that caused by underlining a word or menting these ideas, are described...(Nagy, phrase with a solid line. This causes a major problem 1968, p. 481). with every simple character separation system we 4.13In the Army Chemical Typewriter develop- have conceived.... ments of Jacobus and Feldman at Walter Reed "For certainapplications, such as the input Army Medical Center, a set of characters and system toa translator, some portions Jf non- conventions of usage have been established. "Then, alphabetic material may have to be made available as the structure is typed, a tape is perforated with the for insertion in the translated text. An indication codes of the characters typed, and the codes of the of where this material is located, and how extensive coordinates of the position of the characters on the it is, would need to be recorded by the recognition typewritten page." (Feldman, 1964, p. 206. See also system." (Potter, 1964, pp. 321-322). Feldman et al., 1963). 4.12For example, inrelativelyearly Baird 4.14"The rules established for coding struc- Atomicdevelopments,"theifistripisauto- tures are integrated in the program so that the matically placed along the left hand margin of computer is able to take a fairly sophisticated look each page by the camera when photographing it onto at the chemist's coding and the keypunch operator's the 70 mm film. This serves both as a reference for work. It will not allow any atom to have too manyor grey scale calibration during scanning and as a too few bonds, nor is a '7' bond code permissible means for instructing the text reader to start with atoms for which ionic bonds are not 'legal'. scanning a line. The 11."strip is manually placed Improper atom and bond codes and misplaced along both sides of all graphics prior to scanning. characters are recognized by the computer, as are The reader stops scanning when it encounters the various other types of errors." (Waldo and DeBacker, first strip and resumes scanning when it crosses the 1958, p. 58.) second strip. The space between the strips is sensed 4.15"Provisions have been made to accept and recorded by the reader which prGtiuces ap- input data on magnetic tape in BCD or binary for- propriate control signals for the translator." (Buck mat. The input is coded into machine media from et al., 1961, pp. IV-4.) line-a-timedevicessuchasflexowriters,card 4.12a"With the current research on pattern punching equipment, computer programs, and the recognition devices, we may soon expect automatic Army Chemical Typewriter." (Burger, 1964, p. 8.4.) processing of at least simple graphic diagrams. "The Chemical Typewriter produces X and Y With an oscilloscope input into a computer, one coordinates to locate the characters on a grid which might draw graphs with a light pen and let the enumerates, respectively, the horizontal spacing computer itself, in whatever ways it is capable, positions and relative to ztro vertical positions of reorganize this data for storage and manipulation." its platen." (Burger, 1964, p. 8.) (Barnes, 1962, p. 29.) 4.16"...The molecular formula, which is "Most character recognition machines now in use printed adjacent toits corresponding structural require either a fixed format document, where the diagram and read with it, is an inventory of the material to be read appears in well-defined fields, atoms which appear in the diagram, and so the or manual editing with colored pencil or magnetic computer identification of each alphabetic character ink to outline the areas to be processed. More in the diagram can be checked against this inve a- versatile readers would find application in massive tory ... file conversion and information retrieval projects, "The complete prototype system...consists of andinautomatictranslation,abstracting, and a general purpose binary computer and an optical indexing.... scanner. On a signal from the computer, the scanner "In an interactive system it should be possible transfers the entire image to computer core storage for the operator to direct the scanner only to areas by a series of subprograms to search for black area 1 of the page of interest by virtue of their meaning, ...,to recognize each character and supply a and set in a typeface compatible with the recognition machinelanguagerepresentation...andto logic of the machine. 'Graphics', including photo- construct and check the connection table." (Cossum graphs, line drawings, graphs, charts, and esoteric et al., 1964, pp. 271-272). symbols, could be scanned without any attempt at 4.17"The typing and language rules are quite recognition, and stored in binary arrays for eventual flexible, unrestrictive, and easy to learn. Typing redisplay by some type of facsimile device. Impor- errors are easily corrected by back spacing and over- tant headings, equations, footnotes, and other non- typing or pressing a special "erase" button. Super- machine readable textual material would be typed in scriptsand subscriptedarithmeticexpressions by the operator on an alphanumeric keyboard and can be typed conveniently and summation, product, stored in code. Alternatively, the operator would integral symbols, etc., of arbitrary size can becon- have the option of keying in a few lines of a new font structed from elementary characters or formed by in order to provide the machine with an identified selecting the desired symbolor subprogram from training set to adjust its decision parameters. A an accessory console keyboard. Since the system function keyboard would allow labeling the various has been designed for human interaction, it is not 92 necessary to type neatly formed or symmetric programming language up to the level of ad- symbols. The system will recognize even badly vanced calculus,"(Reinfcldsetal.,1966, formed symbols. Usage may be off-line or in an p. 469). on-line conversational mode." (Amann and Klerer, 4.17c"In the improved terminal area I men- 1966, abstract). tioned above, I'm aware of only one outstanding 4.17a"Our inputdeviceis(similar to the attempt the work done by Columbia University's MADCAP device (Wells, 1961)) a Friden Flexowriter Oceanographic Laboratory. They've built a terminal that has been modified so that subscript and super- that permits keyboard entry of a wide range of script positioning can be done automatically under commonly used mathematical symbols. Although keyboard, paper tape reader, or direct computer the transcription task is thus more difficult, the control. In addition, the 88 available typable symbols have been chosen to give flexibility both in the programming task is simplified since the oceano- typing of equations and the construction of mathe- graphers use their familiar mathematical notation. matical operators of arbitrary size. These may be This approach has not been used widely because it composed manually by typing stroke by stroke or is difficult to build an inexpensive, easy-to-operate by pushing an appropriate key in an optional con- machine that directly transcribes many special sole keyboard." (Klerer and May, 1965, p. 103). symbols. But this experiment shows clearly the "The purpose of this paper is to describe some importance and value of being able to use, in a new applications in the realm of two-dimensional programming language, the necessary symbol." (Reynolds, 1967, p. 29). input-outputbytypewriterterminal....The principal elements of this approach consist of a 4.18"The Army Chemical Typewriter (ACT) versatile programming language and a reverse- ...,was developed by the Walter Reed Army indexing typewriter equipped with a special char- Institute for Research as a means for inserting acter set. This set permits not only normal alpha- chemical structural diagrams into a digital computer numerics but also the construction of arbitrarily- for storage and subsequent searching. This is sized symbols by the use of a few interlockingaccomplished by typing the 'chemical symbols primitive strokes. The reverse-indexing and index- and bond configurations into a matrix from which ing, together with spacing and backspacing, allow isthen reduced to a linear coded representa- full two-dimensional keyboard control of the typed tion on a paper tape. The ACT will be used in the document. Typing errors are corrected by moving IDEEA network as the man-machine interface. All to the desired point and overtyping, or by pressing communications with the systems computers will an 'erase' button." (Klerer and Grossman, 1967, be accomplished through the keyboard of the ACT. p. 675). Thus both alphanumeric and symbolic data can be 4.17b"Our objective has been to devise a inputted and outputted." (Hoffman, 1965, p. 95.) remote computer terminalthrough which the "By interfacing the ACT with the FADAC, we scientist or engineer can converse with the computer create an input/output device that has both memory in the natural language of mathematics. For ex- and arithmetic capability. Thus we can perform error ample, he can enter an equation exactly as it appears checks on the input data as it is typed. The marriage in a textbook and receive an immediate graphical of these two devices further allows for a good deal of and alphanumerical display of the results." (Clem, preprocessing of the data at the input end, such as 1966, pp. 119-120). screen derivation and data classification." (Hoffman, "For the last 18 months we have endeavored to 1965, p. 96.) develop such a system called AMTRAN (for Auto- 4.19"Three basic methods of input are used to matic Mathematical TRANslation) and this paper convert into machine language the structural dia- will review our progress so far. AMTRAN was grams prepared by a chemist. Two of these, the inspired by a similar online system developed by keypunch and thetypewriter,requireclerical Culler and Fried and it shares many basic principles processing steps. The third method involves direct with the Klerer May system. The development of scanning of the diagrams into the computer record. AMTRAN was initiated by Seitz and it is at present Waldo & DeBacker, and Horowitz & Crane have carried on by several of us in Huntsville. The basic used a 48-character keypunch to record structural goals of AMTRAN are: diagrams directly. This method in input provides 1. To use the natural language of mathematics as no direct hardcopy of the keyboard data for editing, a programming language without any arbitrary thoughkey-verifyingisapplicable.Moreover, restrictions whatsoever. because of the restricted character set, keypunching 2. To obtainimmediate graphical output of requires a great many conventions both in the form intermediate and final results. of character substitutions and in the representation 3. To retain a hard copy of useful results and of angular bonds required in the diagrams. Newer programs. keypunches have as many as 64 characters, but this 4. To retain copies of programs in an easily limited extension of the character range does not reusable form. provide for inclusion of an adequate number of 5. To retain utmost flexibility in the system so as special structure-composing characters to simplify to allow its use from the level of existing diagrammatic input." (Tate, 1967, p. 296).

93 .

4.20 Some details are as follows: "Besides auto- to the normal mode is automatically punched." matically producing punched paper tape, the modi- (Bozman, 1967, pp. 2-3.) CC fied typewriter has the capability to provide coordi- . 3 Nixie tubes...are used to display the nate positioning of each type character on the page. numerical code of the "Typit", when reading back In addition, the machine can automatically encode a punched paper tape. When the tape reaches a any one of approximately 90 selected (commercially place which calls for one of the special symbols, available) auxiliary type faces inserted at the platen. the coded number for that particular symbol is This device has been named the `taxywriter' by the shown on the Nixies. The operator can then place NBS staff to indicate a type face augmented XY the correct symbol in the typewriter, type that recording typewriter. character and restart the automatic typing process "The machine can operate either as a standard again." (Bozman, 1967, p. 3.) tape typewriter or in an augmented mode in which "Another feature of this typewriter is that the itproduces four punched codes for each typed rotation of the platen is also punched into the paper character. The four codes are those for the char- tape." (Bozman, 1967, p. 3.) acter, its X-coordinate, its Y-coordinate, and an item "The typist can make up the formula in whatever separator. Insertion of a modified type face element order seems most convenient, rotating the platen inactivates the standard coding for the typewriter up or down, as needed. Since the information is keys; striking any key then produces on paper tape all punched into the paper tape, the computer can five codes, which are a type-face-element prece- reconstruct the formula line-by-line, if necessary, for dence followed by the auxiliary character, X- and the final typegetting." (Bozman, 1967, p. 3.) See also Y-coordinates, and item separator. The modified "Modified Tape-Recording Typewriter", 1966. type face element has four pin holes on each side, 4.20a"Part of the program of the NBS Office into which metal pins may be inserted to signify an. of Standard Reference Data is the development of 8-bit,odd-parity, binary code. The pins make a General Purpose Scientific Document Image electrical contact with a guide at the platen of the Code (GPSDIC) system...The design and devel- typewriter to activate the coding... opment of this system is being carried out by "The Office of Standard Reference Data intends Blanton C. Duncan and David Garvin...As to use the new machine to prepare scientific text designed, the system permits a scientific typescript, and tabular data. Any application in which super- with all its complex symbolism and highly structured scripts, subscripts, special symbols, or coordinate- page format, to be transferred to digital machine oriented material must be prepared in machine- form with virtually no limitations on the notations sensible form could benefit from its use. Its princi- employed.... pal advantages are: "The system was designed to assure broad applicability by emphasizing: (a) exchange of infor- (1) Immediate visual verification by the typist mation via telecommunication devices compatible of the copy being produced. with the USA Standard Code for Information Inter- (2) Rapidly prepared proof copy available on a change (USASCII), (b) design of hardware to permit day-to-day basis (as output from an extended the use of proven skills of ordinary scientific typists character printer) for technical review. in the record capture process, (c) exploitation of the (3)Daily editing on a line correction basis in capabilitiesof commercially available extended conjunction with technical proofing. character high-speed line printers for direct com- (4)Availabilityoftypewriter-qualityscientific puter output, and (d) publication using this type of text with adequate symbols for preparation machine record as the 'typescript' input to com- of manuscripts. puterized typesetting." (NBS Tech. News Bull 52, (5)Preparationof machine-sensible copy for 86, Apr. 1968). use in producing publications by computer- 4.21"The third method of recording structural generated photocomposition. diagrams is by scanner. Meyer has developed a (6) Rapid, low-cost preparation of selected por- photocell scanner for structure input. This scanned tions of recorded scientific text on cards or input, supplemented by a limited amount of key- paper stock for manual filing." (NBS Tech. boarded data that identify noncarbon, nonhydrogen News Bull. 50, 118 (Apr. 1966).) atoms, bond multiplicates, and special structural "The thing thatisspecial about thistype- features, provides a combination scanner-keyboard writer is that these "Typits", as they are called,input record that is transferred by the computer are coded by pins placed in the sides so that the into a single structural record. Ledley, and Cossum information can be read by small switches and can et al., have Developed direct input of structural be punched into a paper tape at the same time that diagrams drawn freehand. Though these systems the symbol is being typed onto the hard copy. are claimed to offer economic advantage and in- "Up to 4 pins can be placed in each side of a creased reliability, only Meyer's system is presently "Typit", giving an 8-level code. When a key is scheduled for operation." (Tate, 1967, p. 297). struck with a "Typit" in place, a precedence code 4.22"What a chemist would most like to do is to is automatically punched, the code for the particular simply draw a structural formula on a sheet of paper. s3 mbol is read and punched, and a code returning We therefore built a machine which by means of 94 photocells can scan structures writtenon paper and to its tip. This serial pulse pattern (in Gray code to punch the information obtained into punched cards eliminate errors) is converted intoa parallel binary or paper tape." (Meyer, 1963, p. 131.) address with appropriate peripheral logic, which 4.22a"The importance of the sound spectro- includes a shift register anda code converter." graph lay in the fact that it provideda visual image (Lewin, 1965, p. 831). of the spectra of speech sound. Itwas in effect the "The RAND tablet devicegenerates 1O -bit x automation of Fourier analysis of speech spectra. and 10-bity styluspositioninformation.Itis It immediately made evident acoustic factors of connected to an input channel ofa general-purpose speech that had not been suspected, and helped to computer and also to an oscilloscope display. The consolidate or eliminate various aspects of the displaycontrolmultiplexesthestylusposition theories that analytical methods had only gradually information with computer-generated information been yielding." (Lindgren, 1965, p. 127). in such a way that the oscilloscope display contains 4.23Theseauthorssuggest,further,that: a composite of the current pen position (represented "Computer measurement of ECG's isnow suffi- as a dot) and the computer output. In addition, the ciently accurate to provide a basis foran automated computer may regenerate _meaningful track history diagnostic classification...In a computer diagnos- on the CRT, so that while the user is writing, it tic program, 'abnormal' measurements are given appears that the pen has 'ink.' The displayed diagnostic significance only when present in certain `ink'isvisualized from the oscilloscope display combinations of leads. This computer diagnostic while hand-directing the stylus positionon the tablet program isbeing written and tested". (Bobrow ....Users normally adjust withina few min- et al., 1965, p. 126.) utes to the conceptual superposition of the displayed 4.24"The program employs a multiple adaptive ink and the actual off-screen pen movement. There matched filter system with a variety of normaliza- is no apparent loss ofease or speed in writing, tion, weighting, comparison, decision, modification, printing, constructing arbitrary figures,or even in and adapting operations. The flexibility of the method penning one's signature." (Davis and Ellis, 1964,p. has permitted study of the effects of experimental 325). variations of these operations on the pattern classi- 4.27"The Grafacon 1010 is basedon the Rand fication process to simulate human interpretation of Tablet, has a 'writing' surface and control elec- electrocardiograms more closely." (Okajima et al., tronics. It does not requirea computer-controlled 1963, abstract.) scanning system to locate and track the stylus. 4.25For example: "...(a) a series of typical The 10 x 10-inch writing surfacecan accommodate filter patterns for P, QRS and ST-t segments of the 106 input locations with 100 lines/inch resolution EKG has been developed, (b) a diagnosticprogram in both x and y." (Datamation 11, No. 5, 99, 101 for rhythm analysis has been prepared, (c)an inter- (1965). [Data Equip. Co., Santa Ana, Calif.]. pretation matrix and a diagnostic matrix have been "Two-dimensional digital graphic inputsystem, devisedthatincorporatepatternrecognition, based on the Rand Tablet, consists ofa "writing" rhythm and extracted parameter data withsome surface, stylus, and associated control electronics. stratifying clinical information for the determination It is said to permit the plotting ofa wide variety of of an EKG interpretation, (d) the loop of the diag- input information directlyon an integral screen. nostic system has been closed by the frequency- Pulses sensed by a high-impedance, pen-like stylus modulated transmission of EKG data by telephone are encoded as serial XY Gray-code information. lines from the Electrocardiography Laboratory of This data is then strobed, convertedto binary code,. the Massachusetts Memorial Hospitals to a GE 225 assembled in a shift register, and gated in parallel digital computer at the. ..M.I.T....and the to computer interface circuitry. It does not require subsequent return of a diagnosis to the laboratory a computer-controlled scanning system to locate and for teletype display..." (Whipple et al., 1965, track the stylus. The printed-circuitscreen, with p. 49). capacitive-coupled encoding, is the hub of thesys- 4.26"The system is entirely digital and the tem. The 10" x 10" writing surface can accommodate tablet is relatively inexpensive. In addition, thin 106 input locations, with excellent linearityand 100 paper sheets can be inserted between the tablet lines/inch resolution in both X and Y.The stylus surface and the pen for tracing maps and curves." contains a pick-up probe, amplifier, anda control- (Lewin, 1965, p. 831). signal switch which is actuated by tippressure. "The RAND tablet consists ofa thin Mylar sheet Data Equipment Co., Div. of BBN Corp., Santa containing on one side, an array of etched copper Ana, Cal." (Computer Design 4, 74, Nov. 1965). lines in the X direction and,on the other side, a "A new graphical input system for off -line digi- similar array of fine lines in the Y direction. By tizing of graphic drawings, strip charts,maps, etc., means of capacitor encoding networks, also etched or from projected slide or film images onto computer on the same sheet, a unique voltage pulse train is magnetic tape has been introduced recently by applied to each X and Y line from acommon pulse Bolt Beranek and Newman's Data Equipment Div., pattern generator. The pen in this case is merely a Santa Ana, Calif. The Grafacon 205-1 Magnetic metallic electrostatic pickup connected toa high Tape Digitizing System is claimedto be several input-impedance amplifier. The pulse train picked times faster to use than other graphic digitizing up by the pen depends on the X and Y lines nearest systems currently available. 95 "The new system consists of the Grafacon 1010A new concept or at least onewhich was not developed Digital Tablet a 10k" square production version of fully in the past. Its greatest advantage in this the RAND Tablet Graphic Input Device with a application (over other conductive film techniques stylus, alphanumeric keyboard, incremental mag- dependent upon voltage gradient) is that capacitive netic tape recorder, and associated electronics coupling to the conductive film by the pen can be assembled in a 30" H X 8' Lx 20" D cabinet. used even through dielectric layers such as a pro- "Optional off-line digitizing equipment includes tective glass covering or sheets of paper. The latter either the Model SS Storage-Oscilloscope for visual items and the fact that the pen used in the Data display of the data as it is digitized, or the Model Tablet contains a standard ball-point pen stylus PL X-Y Recorder Monitoring System for tracking offers the ability to produce a hard copy simultane- the Grafacon stylusposition." (Computers and ously with the data entry.... Automation 16, No. 12, 50 (Dec. 1967).) "A drive network excited by the electronics 4.28"To beat the language barrier between package drives the film at discrete points along its man and machine, Itek has, in effect,hitched the circumference in such a manner that a traveling digital computer to the draftsman's stylus. With a wave (ina mathematical sense)is established photoelectric light pen, the operator of an EDM parallel to each orthogonal axis. This wave has the can formulate engineering problemsgraphically property that its phase is a linear functionof position (instead of reducing them to equations) on a console as in the relationship: that looks like a flat, unflickering television screen. The operator's designs pass through the console V= K sin (cotaX) into an inexpensive computer, which solves the problems and stores the answers in its memory were a, coare constants and X is theposition banks in both digitalized form and on microfilm. coordinate.(It should be mentioned that a true By simply pressing buttons and sketching with the propagating wave does not exist on the writing panel light pen, the engineer may enter into a running since a frequency in the hundreds of megahertz dialogue with an EDM, recall any of his earlier would be required to give a significant phase shift drawings to the screen in a millisecond and alter along on eleven inch path. Actually, only one kilo- its lines and curves at will. The whole system, Itek hertz is used as the phase shifted frequency.)" engineers claim, can be hooked up to permit long- (Teixeira and Sallen, 1968, p. 316). distance design conferences between field sites 4.29a"The actual writing is done on ordinary such as a missile launching pad and the home paper (with any desired printing format)placed on office." (Time, Mar. 2, 1962, p. 76). top of the array. The pressure ofnormal printing 4.29"Sylvania is offering an electronic ballpoint forces the paper and the top membrane down onto pen which translates graphic material to computer the bottom set of conducting lines, producing a language as it writes, and simultaneously transmits contact point between the two orthogonal sets of data to computers for storage and analysis. Called lines in the location directly under the printing the Data Tablet..." Data Proc. Mag. 9, No. 6, stylus. As this stylus moves from point to point, 72 (1967). successive momentary contacts are made, and a "Sylvania ElectronicSystems, a divisionof raster-type image of the character, as it is traced Sylvania Electric Products Inc., Waltham, Mass. out, is signaled to the electronic system. Thestylus The Sylvania Data Tablet is transparent has a can be any natural implement, such as aballpoint protective coating on the surface and does not pen or a leadpencil." (Simek and Tunis,1967, p. 77). require contact by the stylus. Any nonconductive material such as paper or film up tol inch in thick- 4.29b"A necessary requirement for the system, ness, can be interposed between stylus and surface. and a novel contribution of this present work, is the An operator can write on a pad of paper placed over development of an extremely low-cost hand-printing the tablet, and his writing will register. An inking transducer. Functionally, it is similar to the Rand capability on the stylus provides the option for Tablet, but conventional paper and pencil may be creating hard copy simultaneously with electronic used." (Simek and Tunis, 1967, p. 72). entry. The DT-1 can be used as a desk, console, "Basically, the system makes use of a square glass or as a transparent overlay for a CRT orother dis- tablet as a drawing medium and as an ultra-sonic play, such as the map on the cathode ray tube at the delay line, through which mechanical vibrations are left. The operator can post new information graphi- transmitted periodically and alternately by station- cally on the tablet which the computer will accept ary transducers installed along two adjacent edges for up-dating the map. This graphic input device of the glass tablet. A separate transducer is built is easily interfaced with almost any computer. It is into the pen used to make the drawings; when the suitable for many diverse applications such as pen comes into contact with the glass top, the military command andcontrol,machine-aided mechanical vibrations can be detected by the pen design, training devices and as a general research transducer. Since mechanical vibration can be tool." (Computers and Automation 16, 54 (Dec. propagated with a constant velocity through an 1967).) isotropic and homogeneous solid such as glass, the "Use of phase measurements to obtain position time delay between transmission of a wave and its as in the Sylvania DataTablet is apparently a fairly reception by the pen transducer is an accurate 96 measure of the distance between the edge of the easy to study mechanical linkages, observing the drawing board and the pen position.... path of some parts when others are moved. "When mechanical pulses are rapidly and alter- "As a Topological Input Device for Circuit nately transmitted in the X and Y directions of the Simulators, etc.: Since the storage structure of glass tablet, the coordinates of the pen can be Sketchpad reflects the topology of any circuit or determined. This informationisconverted into diagram, it can serve as an input for many network digital form and stored in a computer memory. or circuit simulating programs. The additional effort When recycled out from the memory and recon- required to draw a circuit completely from scratch verted into the graphical form, the information is with the Sketchpad system may well be recom- projected through a CRT system to the back of the pensed if the properties of the circuit are obtainable glass tablet." (Woo, 1964, p. 609). through simulation of the circuit drawn." (Suther- 4.30"The III COMPUTER EYE is an optical land, 1963, p. 332). information sensor and processor capable of measur- "For Highly Repetitive Drawings: The ability of ing and interpreting real world scenes. The sensor of the computer to reproduce any drawn symbol any- the EYE...differs from the tv camera in that, where at the press of a button, and to recursively under control of the Image Processor (a stored include subpictures within subpictures makes it program general purpose computer), it selectively easy to produce drawings which are composed of examines points in its field of view. The EYE may huge numbers of partsalisimilar in shape." seek out a significant part of the image, concentrate (Sutherland, 1963, p. 332). on it, and then move to other areas under guidance 4.30b"A more serious drawback of keyed of a pattern recognition program." (Information numerical input is that it cannot be used to achieve International, Inc.) continuous variation of parameters in the manner of 4.30a1"Canada's Geographic Information Sys- a shaft encoder or a tracker-ball. Many display tem as now set up makes use of an IBM System/360 installations do not include any input devices of Model 65 computer along with equipment to create this type, and a graphical technique has been a central data bankbringing together all existing developed [Aiken Computation Laboratory, Harvard geographic and geodetic data on capabilities and University] to achieve a similar effect using a light uses of land in Canada. pen. This technique has been called the Light "Of necessity maps comprise the main input. Handle, as it simulates the effect of winding a Hence the call for the now installed IBM Carto- handle or rotating a knob. Any coordinate-input graphic Scanner, as it is named, specially designed device such as the RAND Tablet (Davis and Ellis, and built by IBM's Systems Development division 1964) or the SRI Mouse (English, Engelbart and in Kingston, N.Y., under contract with the Canadian Berman, 1967) may be used to control the Light Government's Agricultural Rehabilitation & De- Handle." (Newman, 1968, p. 63). velopment Administration. Geols in fact has become 4.30c"Most of the required components already the key to the entire Canadian Land Inventory existand have been put together at System (CLI) program. Development Corporation, as a prototype system., "Maps are stored in the system by two devices. For the hardware, the user terminal is a RAND ThefirstistheCartographic Scanner which Tablet (Grafacon 1010A), for input to the computer records information on magnetic tape and produces and a CRT display for output. The CRT image is a map in line form. rear-projected onto the Grafacon, so that input and "The second is a manually operated XY digitizer output images are coincident. The terminal is which translates geodetic data on maps into punched connected through a peripheral processor (a PDP-1 cards.Classification data alsois punched into computer) to the AN/FSQ-32 computer, which is cards." (Boggiss, 1967, p. 64). the central processor for the system. Details of the 4.30a2"As more and more applications have terminal and computer interface were described been made, it has become clear that the properties by Gallenson at the 1967 Fall Joint Computer of Sketchpad drawings make them most useful in Conference." (Bernstein and Williams, 1968, p. 28). four broad areas: 4.31"Control Data Corporation announced the "For Storing and Updating Drawings: Each time development of the Digigraphic 270 series... a drawing is made, a description of that drawing Combining free hand drawing using a light pen and is stored in the computer a form that is readily automateddrawing(activatedthroughcontrol transferred to magnetic tape. A library of drawings registers and buttons on a keyboard) for production will thus develop, parts of wh :ch may be used in of perfect circles, angles and lines, the operator can other drawings at only a fraction of the investment create any type of graphic representation on the of time that was put into the original drawing. screen of the Digigraphic console. Drawings are "For Gaining Scientific or Engineering Under- transmittedto the CDC 3300 as thousands of standing of Operations That Can Be Described indidic;ual vectors, the position of each given as Graphically: A drawing in the Sketchpad system relative to a built-in x- and y-axis." (Commun. may contain explicit statements about the relations ACM 9, 468 (1966).) between its parts so that as one part is changed the "Experimental work being performed atthe implications of this change become evident through- Control Data Digigraphic Laboratories in Burlington, out the drawing. For instance, Sketchpad makes it Massachusetts, uses electronic drafting boards and 97 light pens supplemented by keyboards to permit the "The DACI System for design augmented by man to provide input to the computer comfortably." computer at the General Motors Research Lab- (Macaulay, 1966, p. 579). oratories...facilities sketching and designing on "Control Data Corporation, Minneapolis, Minn., the cathode-ray screen. It provides on-lineaccess to is marketing a Digigraphics System designed for *Large,computer-processiblefilesof'blueprint' use with a small -scale Control Data 1700eomputer information.Automobilecomponentsdesigned System. Digigraphics refers to a CDC system in through DACI are now on the street." (Licklider, which an operator seated at a special console equip- 1967, p. 9). ped with a large TV-like display screen is able to "Of considerable historical interest, a program enter data in graphic form using a light pen and which roughly paralleled the M. I. T. work has been keyboarddirectly into a computer system for carried on independently at General Motors since processing, storage and retrieval. about 1959. It was kept in completesecrecy until "The controller used in the 274 Digigraphics its disclosure in the fall of 1964." (Prince, 1966,p. System has a 4096 word Luffer memoryexpand- 1698). able to 8192 words for operating a single Digi- 4.33"It is possible, in conjunction with the light graphicsconsole. Avector-oriented,22-inch pen and a suitable program, to read to progressively diameter, flatface visual display screen on thecon greater levels of detail in an area of interest selected sole is capable of displaying the equivalent of 2000 by the light pen. Looped playbacks of the successive inches of curves, or up to 1800 characters of variable reads on a visual scope can be enlarged to reveal size or font.... discrete points. It is possible to apply the differential "The software package (Function Control Pack- principle in reading aroundcurves in order to age) provided with the system offers a large number produce a center line of h'igh quality. It is alsopos- of universal design features. This means that the sibleto recognize patterns of objects, and to user, rather than structuring a software system from combine code-generated pictures, computer- scratch, need only interface his application pro- controlled visual input and lightpen visual input in gram(s) with existing routines to achieve full opera- a single operation. The recorded output of the com- tional capability. Integration of user oriented pro- puter can be stored on film and read back into the grams with the FCP software is accomplished via computer visually at any time. Overlays can be FORTRAN IV CALLS. quickly accomplished on standardmaps by reading "Upgrading any general purpose Control Data the maps with the computer-controlled CRT reader 1700 Computer System to Digigraphics capability and combining them inmemory with elements to requires only the addition of the 274 Digigraphics be displayed. The resultcan be put out on film.... Controller, the 274 Digigraphics Console, and 24,000 "A computer-controlled televisioncamera is a words of core and disk pack memory." (Computers general-purpose visual input device capable, when and Automation 16, 60 (Nov. 1967).) properly programmed, of readingany gauge or 4.32"The [RCA] system...utilizes the pen as instrument panel or observingan experimental the signal generator and the writing surface as the situation and, through its computer, controlling it address detector. The pen contains in its tip a rapisdel small magnetic head which periodically generates a yvt:aleComputer programs have been written localized magnetic field pulse. (Since the coupling to permit an operator to perform drafting using the is magnetic, it is not shielded by most materials visual scope and the lightpen. The computer- placed between the pen and the tablet). The writing controlled CRT reader enhances this operation by surface contains a number of thin winding layers in permitting the introduction of photographs and a laminated structure ...There are as many layers drawings." (Fulton, 1963,p. 38, 40). as there are address bits, each developing a positive 4.34As examples of current hardware capabili- or negative induced voltage as a function of the ties for this purpose we note the following: "A new pen position." (Lewin, 1965, p. 832.) photographic data processing system which will 4.32a"The initial goal of the Design Augmented speed post-flight analysis of photographic data... by Computer project was the development ofa was developed by David W. Mann Co. ...The new combination of computer hardware and software Microdensitometer system. . measuresthe posi- which(a)would permit'conversational'man- tion (in two coordinates) and the relative lightness machine graphical communication and (b) would or darkness of micron-sized images while scanning provide a maximum programming flexibility and photographic plates at rates up to 625 millimeters ease of use for experimentation. This goal was per minute." (Data Processing Magazine 7, No. 2, achieved in early 1963." (Jacks, 1964, p. 344). 40 (1965).) "The IBM 7960Special Image Processing "A new, low cost method of input to computers System was designed and built by IBM to specifica- of diagrams, drawings, and photographs has been tions provided by the General Motors Research developed by D-Mac Ltd. The device, the Pencil Laboratories. The system is the man-machine and Follower, allows outlines to be traced and converted image processing hardware for the GM Research to computer input, or entered on cards or tape for DACI System." (Hargreavesetal.,1964, pp. subsequent processing on a computer." (Data & 384-385). Control 3, 32 (1965).)

98 "The imagery...was processed in IMITAC not noise as long as nontargets do not also generate (Image Input to Automatic Computer), a specially silhouettessimilarincharacteristictothose constructed scanner which converts images to generated by any target." (Holmes, 1966, p. 1681). computer language, and can also convert computer "The CAL [Cornell AeronauticsLaboratory] output to images. The Phi leo IMITAC can scan a flying Spot Scanner ...can sample a 3" X 3" 3-by-3 inch image with a 1024 line raster, each line photographic transparency witha spatial resolution being sampled at 1024 points. An analog-to-digital of 1024 X 1024 sample points with better than 95% converter encodes the photographic density to 64 independence and with a density resolution of 64 levels. This 6-bit code is then transferred to the levels of gray. Of particular value to this research UBC (Universal Buffer Controller) of the Phi leo 2000 is the fact that the scanner can sample the pattern computer system, and then to magnetic computer scene transparency in a random-access fashion, tape." (Kanal and Randall, 1964, p. D2. 5-5). under control of our IBM-7044,so that it can sense 4.35"The input data for a given compilation is directly those portions of the scene required by the in the form of a pair of aerial photographic trans- object extraction process.... parencies together with pertinent camera data; "In research reported earlier by Trabka and i.e., position and attitude of camera for each trans- Roetting objects in a complex sceneare located by parency, focal length of camera and distortion performing anopticalcross-correlationof the characteristics of the lens." (Bertram, 1963,p. 105). pattern scene and an aperture filter of the shape 4.36"We have constructeda scanner system of the object to be located. This technique locates that will look at a resolution element 1/200 inchon a the desired object in the objectscene, but it does side, that will assignone of eight gray levels of light specifically define its boundariesor extract it from density to each element, and that willput the in- the scene. However, the technique is very fast,as formationina systematic way on a standard it can process a whole scene in parallel, and is magnetic tape. We Lave computerprograms that very tolerant of changes in shading in the interior will then manipulate the information containedon of the object, since it detects objects by the shape the tape. We can display it bymeans of a pictorial or their edges." (Muerle and Allen, 1968, p. 5, 3-4). printout; we can list the actual density values for 4.36c"A prototype Map Analysis System any area of interest." (Moore et al., 1964, p. 927). (MANS) has beer developedas an experimental 4.36a1"Aeronutronics Div. of Ford Motor Co., first approximation to... ageneral picture de- has ordered a Digital Equipment Corp. PDP-7to scription system." (Pfaltz et al., 1968,p. 361). use in an experimental film-scanning operations. "Within the geographical profession, themap is They plan to mechanize and automatemany steps thetraditionaldata-storageandnomographic- in detecting and analyzing the data contained in analysis device. These two functions of themap are aerial photographs." (Data Proc. Mag. 7, No. 7, being 'computerized' as rapidlyas is economically p. 12 (July 1965).) feasible. The theoreticallymore important use of 4.36a2"The system for automatic photointer- the map as a 'hypothesis-generation' device would pretation in Karlsruhe, which is still under develop- require a computer that can read that is interpret ment, will be used for scanning and preprocessing and not just scan maps. In spite of intensive efforts negative films such as aerial photographs, bubble on the part of the military to develop an automatic chamber photographs or X-ray photographs without photointerpretationcapability,theresultsfall any restrictionasto special structures of the largely into the category of data reduction rather pictorial information. Aerial photographs will be the than pattern recognition. For thisreason maps chief subject of investigation, since they have the remain an oft-desired output format. The art of most general structure." (Kazmierczak and Holder- computer mapping has been developed fairly ex- mann, 1968, p. 45). tensively using offline printers, graphic plotters, "Considerable work has been doneon the simpli- and cathode ray tubes to produce statistical atlases, fication of the information content of aerial photo- contour-like maps, andso on." (Tobler, 1967, graphs by, e.g., successive processes of averaging pp. 57-58). and differentiation." (Rosenfeld, 1965,p. H4). 4.36d"The bubble chamber is one of the most 4.36b"The digital filters are designed to produce versatile tools for high energy physics research. silhouettes on the basis of broad target and/or It is used to detect the passage of particles and to background characteristics but not to dependon measure their trajectories, momenta and velocities. specific target shape or configuration. Every sil- Nuclear events are produced by the interaction of houette produced representsapossibletarget beam particles from the accelerator with nuclei and is individually tested as such in a final stage of the chamber medium." (White, 1968, p. 175). recognition system....The generated silhouettes "Charged parti^.1es whichpass through the cham- not only include a large number of nontarget shapes ber during the sensitive period of its expansion cycle which must be reliably rejected to achieve a satis- cause formation of strings of bubbles along their factory false alarm rate, but may also include trajectories, with bubble density determined by the imperfect silhouettes of targets which must be particle velocity. A magnetic field is presentto reliably recognized as such. Some noise effects allow particle momentum to be measured by the are inevitable, but consistent imperfections are radius of its trajectory. Generally threecameras 99 photograph the bubble chamber during the sensi- 4.39a"It is possible to obtain a spatial derivative period of its expansion cycle, so that by inter- tive of the scanned image which simultaneously comparison of views a space reconstruction of rejects those parts of the original image whichare particle tracks can be made." (White, 1967, p. 175). out of focus and preserves the boundaries where 4.37"Dr. Bruce McCormick has proposed a the value of the derivative exceeds some threshold." scanning technique which allows rapid recognition, (Lipkin et al., 1966, p. 1006.) separation and measurement of the photographic 4.40"Specialequipmentdeveloped by the records of star type nuclear events. A device known Electronic Systems Laboratory at M.I.T. makes it as the Spiral Reader measures background and possible to draw with the light pen a three-dimen- star type event features impartially, discriminating sional object and display its projection on the face against non-radial patterns by the geometry of its of the cathode-ray tube just as if the objects were rotating scanning element. The event measurements continuously rotating in three dimensions under are separated from the background measurements control of the viewer." (Fano, 1967, p. 31). by an IBM 704 computer under the direction ofa "Another device that is in many ways superior to program called FILTER. The separated nuclear the light pen and RAND tablet is the Lincoln event measurements are subsequently reconstructed WAND*. It uses 4 ultrasonic transmitters and one in space for physics analysis." (Inner, 1960,p. 25.) receiver to obtain digital delay measurements that 4.37a"Visual inputtothe computer enters determine the stylus position, and can operate over through flying spot CRT scanners: two for 70-mm a working area four feet square with a resolution of film, two for 46-mm film, two for 35-mm film and 0.02 inches." (Coggan, 1967, pp. 73-74) [Ref. to two for microscope slides. Two modes of raster Roberts, 1966] scanning are provided: raster mode, where cell-by- "The Wand is an interesting development pres- cell the image is encoded uniformly in 1, 2, 4, or 8 ently being tested at Lincoln Laboratory for inputt- bits of gray scale; and coordinate mode, where each ing three-dimensional data to the display. Four black/white transition triggers the readout of 16 acoustic transducers located around the periphery bit X(Y) coordinates. Raster mode is the normal of the display transmit pulses which are detected recognition scan input; coordinate mode is primarily by a microphone in the wand. Counting circuitry reserved for precision measurement purposes." measures the four distances which are converted (McCormick et al., 1966, p. 359.) into xyz display coordinates by the computer." 4.37b"Because the part [i.e., design component} (Prince, 1966, p. 1699). can be rotated to any position in space, lines can be "Research continues actively in the areas of drawn directly in three dimensions by drawing ina new input devices and display media. The Lincoln plane. The part is rotated until the area in which WAND, described by Roberts at the FJCC, allows the line is to be drawn is parallel to a viewing quad- the operator to position a hand-held ultrasonic rant. The line is then drawn true length; the depth receiver in three-dimensional space. The sensor coordinate remains constant as the pen moves receives signals from four transmitters aligned in across the plane of the scope screen. Specification in a plane near the scope face and allows the com- of the single depth coordinate is done byprogram puter to resolve its position in three dimensions. interpretation...This program, called the Pen Thus, the 'menu selection' area used for pointing Space Location program, is the backboneprogram at program alternatives can be significantly in- of three-dimensional sketching. Sophisticated draw- creased from that of the two-dimensional scope ing is made possible by this program." (Johnson, face used with the light-pen. In addition, three- 1963, p. 350.) dimensional 'drawing' can be simulated." (Van It is noted, however, that "general three-dimen- Dam and Michener, 1967, pp. 197-198.) sional graphical communication, which deals with "The well-known 'light pen' or 'light gun' has arbitrary surfaces and space curve intersections, been used since the early 1950's as a mechanism presents many difficult problems; the beginning has for drawing or pointing on the face of a CRT. been modest and much work remains before the Roberts describes the so-called 'Lincoln WAND,' a completegraphicalcommunication problemis related device for pointing in three-dimensional solved." (Johnson, 1963, p. 347.) space. The pointer contains a microphone, which 4.38"Sketchpad has been extended to three detects pulses from several ultrasonic transmitters dimensions by Johnson. In Sketchpad III, theuser in the vicinity. The result is a sort of small-scale can add a line to a plan and have it appear simul- ultrasonic LORAN (Long Range Radio Navigation) taneously in the front field, the side view, and the which is already good enough to give a positioning oblique representation. When he rotates the oblique accuracy of about two-tenths of an inch. A few representation, the orthogonal views change appro- suggestions are given as to the importance of adding priately, etc." (Teitelman, 1966, p. 11.) a third dimension and considerable spacial freedom 4.39"The only work I know of on machine depth to conventional pointing schemes. A final item of perception is that on binocular images. Julesz has display hardware is the Stratton description of a reported a procedure which shifts the binocular novel replacement for the conventional light pen. pictures to find the areas at different depths. This The general idea involves direct detection by the procedure uses only texture, not edges, to develop `pen' of the CRT electron beam rather than of the depth information..." (Roberts, 1965, p. 161.) the light emitted when it strikes phosphor. Stratton's 1011 paper describes the hardware in intensive detail and formation technique reduces the information con- will be of greatest interest to other hardware de- tent of a contour or graphic representation to a signers and implementers." (Mills, 1967, p. 232.) compact form, it is ideal for information transmis- 4.41"I would estimate that within the next two sion. Two otherpotentialapplications include decades a three-dimensional scanner will be avail- analyses of handwriting and recognition of signa- able that will describe the shape of an object in tures. The representation of a signature's boundary (elms of a grid ...which can be as fine or as by Fourier harmonics presents especially interest- coarse as the investigator desires. The object would ing possibilities. The areas of filtering and smoothing be placed between two such grids, and where the also offer many possibilities for applying these beam could pass between two parallel points of techniques." (Brill, 1968, p. 9). course the object would not be in between these Most of the applications of optical systems to points, and where the beam was reflected back the qualitativepattern recognition deal with trans- distance on either side would be calculated, and the formations that directly involve spatial frequencies. difference would be the thickness of the object at Many other simple transformations that are useful that point." (McDonough, 1964, pp. 34-35.) inqualitative pattern recognition are notwell 4.42"Three dimensional x-ray photographs are understood, and cannot be realized by simple optical being produced...by thnGianniniControls systems. A few of such transformations are noted Corp.,...to see the distance between component here. leads in a potted module...also has medical 1. Interpolation and extrapolation, elimination or applications. For example, it has been used to deter- retainment, rotation and translations of speci- mine the depth of a bone infection." (Electronics fied lines and curves in a pattern. 37, 26, 1964.) 2. Increasing or reduction of thickness (width) of 4.43"Ford is reported to be using a light-scan- a pattern in a specified direction or vicinity. ning measuring machine, built by the Ex-Cell-0- 3. Separation of superimposed patterns with over- Corp., to inspect clay mock-ups for X (length), Y lapping spatial frequencies. (width) and Z (height) coordinates... 4. Nonlinear transformation such as clipping, "Ranging circuits in the scanning machine meas- thresholding and clamping the pattern ampli- ure the light's travel time, convert it intodistance tude for which simple analog realizations do from the source, and feed the data into a card- not yet exist. punch." (Gomolak, 1964, p. 66). 5. Transformation from defined shapes to other 4.44"The method records the light's electro- shapes such as simultaneous mapping of magnetic field exactly asit existed, so that the triangles into circles and squares into ellipses images it regenerates are three-dimensional images in a pattern consisting of a number of tri- indistinguishable from the original scene... on angles and squares. the fringes, three-dimensional television." (Stroke, 6. Conformal transformations." (Hemami, 1968, 1965, p. 53). PP.6-7). 4.44a"This paper is involved with one aspect of 4.45For example, "the program may be com- the pattern recognition problem, namely, that of manded from the typewriter to analyze a complex qualitative pattern recognition. A pattern recogni- scene. Such a scene may consist of known items, i.e., tion problem is qualitative if certain transformations shapes which have been defined as being significant of the pattern are either the end result in themselves (the alphanumeric characters); and of other items, or are necessitated for ease of further processing or some of which may be significant to the viewer but recognition. One can extract partial information not to the system, and some of which may be merely (both qualitative and quantitative) about the pat- background 'noise'. An arbitrary number of known tern through such transformations. ... items may be present simultaneously; they may be. "The transformations of interest cannot, in gen- of different sizes and orientations; they may overlap, eral, be expressed in terms of known mathematical or be inside each other; they may be super-imposed transformations. Even if such transformations were on an arbitrary background."(Marill et al., 1963, found, it may not be presently possible to obtain the p. 28). correspondingphysicalrealizations.Forthese 4.45a"In dealing with... aphotograph, con- reasons, qualitative pattern recognition will be a ventional techniques of shape and pattern recogni- challenging area of research in future, both in terms tion arnot directly applicable, since there is no of new mathematical tools and new physical compo- naturadistinction between figures (= shapes or nents." (Hemami, 1968, pp. 1-2). patter s) and background in the photograph. ... "Although qualitative pattern recognition has not "Statistical analysis of the input image...is been emphasized, there are numerous applications used to determine the boundaries of the conspicuous of these techniques to qualitative problems. For figures which the image contains al.d to generate a instance, the techniques could be used to accentu- simplified description of the'visual texture' of ate specific features on contour maps, such as various parts of the image.... regions with steep slopes or regions which are nearly "An arbitrary point pattern (or figure made up level. Another application is compact information of straight lines) is determined if certain lengths of transmission of graphical forms. Since the trans- line segments or angles between them are specified.

101 More generally, an arbitrary shape is determined in Perceptrons and other 'trainable' networks." by specifying the curvature of its boundary as a (Kanal and Randall. 1964. p. D2.5-7). function of arc length." (Rosenfeld, 1962, p. 114). 4.46"The first step taken by ANALOGY is to "What makes picture processing a subject in its decompose each problem figure into 'object' (sub- own right is that it deals with pictures which are figures)...While a decomposition program of the not merely arbitrary functions or matrices, but fullgenerality desirable has not yet been con- which are pictures of something which purport to structed, the most recent version of the program is representarealscene(terrain,microscope capable, in particular, of finding all occurrences of slide,...) or an ideal symbol (such as an alpha- an arbitrary simple closed figure x in an arbitrary numeric character)." (Rosenfeld, 1968, p. 1-10.) connected figure y; for each such occurrence the "Experimental work... suggeststhat...de- program can, if required, separate y into two objects: tailed contrast frequency analysis provides enough that occurrence of x and the rest of y.... information about the visual texture of the image "Next, the 'objects' generated from the decom- to make possible the automatic identification of position process are given to a routine which calcu- many basic terrain types." (Rosenfeld, 1962, p. 115). lates a specified set of properties of these objects "The extraction of conspicuous figures from the and relations among them...[e.g.,] that the ob- photographic image also involves the concept ofject labeled P2 lies inside that labeled P3 and visual texture, since the boundaries between adja- generates a corresponding expression ..." (Evans, cent areas having significantly different textures 1964, pp. 329-330.) will define such figures." (Rosenfeld, 1962, p. 115). 4.47"...We are now able to begin serious 4.45b"A two-level statistical classification pro- study of the most difficult problem facing the proj- cedure has been applied to the problem of detecting ect: The analysis of real-world three-dimensional complex targets in aerial photography. At the first scenes." (Minsky, 1966, p. 12). level, a set of classification functions designed on "Programs have been developed to read selected the basis of samples from the target class and from parts of the visual scene, analyse them for parts other images is used to make subdecisions on local- of polygonal objects, and then transform them to area statistically-designed featuresassociated with real-world coordinates. The present programs are the target class. At the second level these sub- still rudimentary, and their extension is vital to the decisions are combined into a single decision as to project." (Minsky, 1966, p. 16). the presence or absence of the target. The nature "Our goal is to develop techniques of machine of the data does not allow for the direct application perception, motor control, and coordination that of classical methods of multivariate discriminant are applicable to performing real-worldtasks of analysis; rather, modifications of classical methods object-recognitionandmanipulation."(Minsky, are used. This procedure has beensimulated on a 1966, p. 11). digital computer with the aid of a special input- "The sensory equipment includes two visual- output device which converts imagery to computer input devices. TVA, a vidicon television camera, language." (Kanal and Randall, 1964, p. D2.5-1). and the more precise TVB, an image-dissector "Thus the imagery screening system consists of deviceforcontrolled-scananalysis."(Minsky, (a) a high-speed flying spot scanner, (b) a Laplacian 1966, p. 13). preprocessing stage that converts the video to 4.48"...The problem is to make measure- binary data,(c)a shift register correlator with ments on a curve ...Since the lines are thick, statistically designed coefficients, and (d) the final the measurements must be made from points in the decision logic. The result is a machine that can middle of the lines." (Ledley et al., 1966, p. 79). rapidly search large amounts of photography and "The transform of a pattern consists of a locus, reliably detect a variety of tactical targets in any called the medial axis of the pattern, together with position.... a function defined along this locus, which describes "The work described here and other related the exact shape of the pattern...The function efforts lead to a (preliminary) conclusion concerning ...is the distance from the curve to the pattern therelativemeritof some competing design boundary. The transform is generated by causing approaches. For the identification of small targets the pattern to shrink down in size, or equivalently in aerial survelliance photography, the two-layer by allowing the area outside of the pattern to propa- statisticalclassification method based on local- gate with uniform velocity into the pattern. The areastatisticallydesignedfeatures andusing medial axis of the pattern is described as the locus Laplacian pre-processing isfar superior to the ofself-intersectionofthepropagatingarea." `random-mask' methods used (at least in the past) (Philbrick, 1968, pp. 395-396).

5. Preprocessing Operations and Pattern Recognition

5.1"...Filtering for irrelevancy should be 5.2For example, in standardized OCR font performed at the sensor site in most cases in order readers being developed for European postal check toreduce the demands on informationtrans- handlingandsimilarapplications,Standard mission facilities." (Edwards, 1965, p. 148). Elecktric Lorenz and Telefunken use infrared- 102 range scanning, while Olivetti has a prototype of electrical signals but performed in two dimen- readerunderdevelopmentusingultraviolet. sions. The concepts of bandwidth, frequency and "In general, use of the infrared range makes rejection or pass bands are directly applicable. interferences from fountain and ball-point pens, Since, in the Fourier plane frequency increases .colored pencils, office stamps, colored paper, and with distance from the optical axis, a circular axial finger smudges tolerable, while materials contain- stop will reject the d.c. and low frequency compo- ing carbon or graphite have a reflectance similar to nents of the pattern. A circular axial aperture will printing inks a:id thus cause trouble." (Dietrich, reject high frequency components. Clearly an annu- 1965, p. 319.) lar aperture will pass only middle frequencies. 5.3The technique of carrying analog gray-scale "The collimated output from a laser, after passing values rather than quantizing early in the recogni- tnrough a phototransparency, is brought to a focus tion process is associated first with Taylor (1958, by lens LI. This focal plane contains the Fourier 1959, 1960). transform of any pattern in the transparency. The is itself imaged, 5.3a"...We find almost universally that the plane containing the transparen very first operation after linearamplification of the again by LI, onto the photocathode of a television video signal obtained from the scanning process is camera. The video signal from the camera could a 'black-white' clipping. From thedetection point then be inputtoa pattern processor/analyser of view, this nonlinear process irretrievably rivets proper."(NationalPhysicalLaboratory,n.d., the noise to the signal, and now all kinds of logical pp. 2-3). acrobatics are required to save the situation." 5.4aSee Moore (1968), as follows: "STRIP-3 (Nadler, 1963, p. 814). See also Hart (1966) p. 17. provides for the use of a number of different raster 5.3b"If there is a contrast boundary at the sizes appropriate to pictures in most net print sizes scanned point, the preprocessing of the scanned from 30 millimeters up to 24 X 24 centimeters. All information consists in computing the gradient of images used in any one run of the program must the contrast. If there is no contrast, only the grey however be of the same size. In the anticipated value in this location will be measured. While the machine configuration, a minimum of two '8 X 10 grey value can be measured on ascale of 64 levels, inch' (24 X 19 cm) binary phase images can be held the contrast gradient will be measured on a scale of in the active memory, while up to 8 binary images in 40 steps in a 3600 range with an accuracy of -±4.5°." `4 X 5 inch' or smaller size can be retained and (Kazmierczak and Holdermann, 1968, p. 49). processedincombination.Logicallymodified 5.4"It is well known that, due to diffraction images can be stored without destroying the orig- effects, the image of a point source produced by a inals. Where operations directly upon the multiple convex lens is surrounded by the Fouriertransform level density scale may be necessary, these are of any aperture placed in the plane of the lens. if accomplished by logical combinations of binary this aperture is a photo-transparency then the image imagesrepresentingthe8-4-2-1or 4-2-1 plane contains the Fourier transform of any pattern bits of the density characters. While this method in the transparency. The spatial frequency compo- appears indirect, it is in fact several times faster nents of the pattern are displaced radiallyabout in operation than the arithmetical methods which the optical axis, displacement increasing with spatial would of necessity be performed one gray character frequency. Opaque stops can be introduced into at a time ... the image or Fourier plane to suppress selected "A...powerful series of image transformations frequency components, an image of the aperture are accomplished through the useof 'BIT OPera- pattern is then transformed according to the trans- tions'. These are processes in which new image fer function defined by the stops. The Fourier bits are determined for each raster point as a func- transform, being, in effect, a modified image of a tion of both their original value and the value of fixed point source, is independent of the position each of the eight neighboring bits. Thus individual of the pattern in transparency, and hence also of points of the image can be classified as being within any part of the pattern so that patterns maybe the body of particles, in specific edge positions, or processed without precise location, alternatively in isolated situations..." (Moore, 1968, pp. 287, several patterns may be processed simultaneously. 296). If the low frequency components of the Fourier transform are suppressed areas of sensibily constant 5.4b"In the case of two-level pictures, such as density in the pattern are eliminated and an outline printed matter, the entire image can be recon- effect is obtained; the suppression of high frequency structed from the contours. For pictures with a components reduces fine granular noise and finecontinuous tone scale the contour information can structure. By the use of appropriately shaped stops be used to reconstruct the high frequency com- a large range of transfer functions canbe obtained ponent of the complete picture. In either case and this technique may conveniently and economi- efficient coding of the contour data is essential. cally be used to preprocess patterns before more Since the contour is by nature a highly connected elaborate analytical processes are applied. The set of points advantage may be taken of theinherent operation of spatial filtering of optical signals (pat- correlation to reduce the total required channel terns) is directly analogous to electrical filtering capacity." (Schreiber et a1.,1968, p. 1). 103 5.5"Most approaches to preprocessing can be for more detailed analysis of the input pattern. classified as either attempts to reduce the dimen- Feedback functions... mayalso arise where any sionality of the sample space or attempts to build input pattern, or input pattern element, determines up a description of the samples by isolating features. the selection of succeeding elements of the source The former approach, which is usually analytical, pattern. In reader developments at Philco, several suffers from the use of narrow assumptions which different scan modes and techniques for focussing severely restrict the applicability of the results. and defocussing serve to enlarge or reduce the The latter approach is usually based upon intuitive area of pick-up." (Stevens, 1961, pp. 56-57). considerations." (Hart, 1966, p. 29.) 5.9"...The probability with which the various He suggests further that "one might first try to elements may be black or white can be determined isolate a number of features and then use a dimen- by examining them in relation to their neighbors.... sionality-reducing technique to condense the de- "The analog image signal...consists of about scription of the pattern further. This combined 500 picture elements per character and is converted approach is, in fact, not often seen, and perhaps by a trigger amplifier into a digital black and white deserves closer attention." (Hart, 1966, p. 4). image. Noise due to imperfect printing is reduced 5.5a"Many measurements which seem other- by comparing each picture element with its neigh- wise very useful in discriminating printed char- bors and subsequently the appropriate form ele- acters are sensitive to line widthcertain first order ments are detected by line following, whereby autocorrelation measurements, for example. One character height serves as a relative dimension. technique for filtering this out has been designed In parallel with this operation the automatic reader by H. B. Baskin at our Laboratory. It utilizes a com- determines the size of each character along with its bination of local analog and binary operations on position in the scanning field, and compares it the video signal from a scanner to reduce line width with limit values." (Gattner and Jurk, 1963, pp. variations." (Andrews, 1962, p. 291). 392-393). 5.5bSee also the following: "The commonly used spatial filtering techniques "Although large character size variations might are differential or edge enhancing, filtering, and be considered as font variations, it is worth con- lowpassfiltering...Lowpass filtering smooths sidering them separately since it is possible to data, for example, by replacing each matrix element filter them at an early point in a recognition sys- by a weighted average of itself and the surrounding tem by automatically varying magnification. NV lk.,t eight elements." (Hankley and Tou, 1968, p. 421). seems to be the simplest way to accomplish this 5.10"A 'noisy' picture can often be effectively is to vary the scan resolution of a CRT type opti- `smoothed', or an unsharp picture 'enhanced', cal system in response to previously sensed aver- by a single neighborhood operation which takes a age character height." (Andrews, 1962, p. 290). localaverageorcomputesafinite-difference 5.6See note 2.7a. Laplacian.Similarly,a picture which contains 5.7"In an early Post Office address-reader de- thick 'roads' (lines or curves of points having given signed by Farrington-IMR, the input process con- values) can be 'thinned' by iterating a 'border ele- sists first of finding the lowest line of the typed ment deletion' operation, perhaps alternated with a address on an envelope, with scan of each succes- smoothing operation, where the number of iterations sive one of four areas 1+ inches wide to determine required is relatively small since the roads are where this last address is located, and with sub- narrow compared to the picture size." (Rosenfeld sequent scanning such as to follow the apparent and Pfaltz, 1966, p. 473). lowest line. The shadow cast by the window of a "Edge differentiation may be used to obtain window envelope has, for example, caused diffi- more accurate area values, and the tone scale culties in adjustment for this lowest line position." may be enhanced or corrected if the original tran- (Stevens, 1961, p. 56). scriptionisinaccurate." (Moore, 1968, p. 285). "The work done by E. Milbradt and J. Bauldreay "The procedure we have implemented (`STRIP') in solving registration problems for the recognition ...results in an object which consists of of mail addresses utilizes the edge of the character arcs: connected sets in which all points have as a datum line for decision making. A preliminary exactly two neighbours in the object reading is made to obtain a registration error. This error is fed back to correct the optics such that the nodes or nodal regions: 'tips' consisting of character placement is correct at the second read- single points with one neighbour only, and ing station. Pre-biasing is also established to take `nodes' in which each point has more than advantage of a priori knowledge of character drift two neighbours." (Rutovitz, 1968, p.127). along the line." (Stein, 1965, misc. notes). 5.8"Feedback functions may affect this opera- "Two-dimensionalfiltering with nonlinearities tion as in cases where the first input pattern is used introduced can produce interesting and useful to adjust repositioning of the source image, to trigger results. The necessity for nonlinearity precludes the beginning of specific recognitionsteps,to optical processing for the present and makes the standardize the dimensions of succeeding input digital computer a unique tool for research, and pattern elements or to set the frame of reference probably an essential element of a final system. 104

t The scope available for further research on two- which have consistent values over variations of a dimensional, nonlinear, digital filters favors con- particular symbol, yet which have different values tinued work in this area. The work should con- for different symbols." (Groner, 1966, p. 592). centrate in three fundamental regions. "The total number of features is not necessarily (1) Filters to define and measure the orientation the most critical consideration. The reliability with oflow-curvaturecontours.Initially,non- which these shape features can be detected in all noisy contours should be studied but, ulti- significant type fonts is much more important, and mately, thestraightbut locally irregular will probably be the subject of much additional contours, characteristic of the boundaries work in the development of multifont and hand- of wooded areas adjacent to cultivated land written character readers." (Greanias, 1962, p. 145). should be included. 5.11b"The main purpose of feature-ordering is (2) Filterstoincorporate texture aswell as to provide, at successive observations, the feature amplitude distribution continuity in an ex- which is most 'informative' among all possible tension of the KS filter approach. choices of features, for the next observation." (Chen, (3)Processes to assemble silhouettes fragmented 1966, p. 552). by shadow effects and close the gaps created. "The importance of utilizing measurements of the The processes may take advantage of prox- order of encountering points as one traverses the imity, edge alignment effects, or target size branches of a pattern is especially pointed up by requirements (that is, assembling proximate the comparison of the results obtained with the silhouettes to make silhouettes of approxi- Highleyman data....Closely associated with the mately the correctsize)." (Holmes, 1966, point-ordering feature is the use of topological data." pp. 1685-1686). (Minneman, 1966, p. 95). 5.11"Classification or the arrival at a category 5.12Soine examples of stroke direction analysis assumes a matching or near-matching of something techniques are provided in Shepard's 1953 patent with something, or the failure to achieve such a for an "apparatus for reading". Bomba (1959) matching...data-reductionprocessesreduce used local operations to extract horizontal and things that are non-matching in detail to things vertical lines, slanting lines at various angular dis- which are matching in some overall aspect. In some placements from the vertical, and selected orienta- sophisticated types of pattern recognition, matching tions of F-, L-, and V-shaped line intersections. or near-matching within tolerance is done only after Selfridge (1955) and Dineen (1955) were early considerable data reduction has taken place." exponents of property filtering techniques to select (Minot, 1959, p. 11.) .significant features such as edges and corners. "Reliance is normally placed on differences in In 1966, at the University of Naples, a pattern color or tone density, or on general logical process- perception study involves use of a curve analyzer ing, to eliminate features for which measurements to detect maximum and minimum inflections and are not desired. The numerous parameters measured the investigation of coding into a discrete number for individual particles or objects may, when appro- of symbols of scanned line connections, corners, 1 priate, be used in suitable combinations to classify and other characteristic features. (Stevens, 1968, 1 the measurements into groups pertinent to single classes of objects. Such edited data may then be p. 17). used to compute summary information on individual Rochester et al. (1959) claim with respect to the classes of objects." (Moore, 1968, pp. 324-325). "lakes" and "inlets" method that "the use of "The effectiveness of a recognition system will mathematical topology is getting at the crux of 1 depend upon how well the significant differences distinctive features...little attention is paid to between pattern classes are characterized by the the lines of the character as such, but the lines are prescribed set of measurements, and the accuracy only important insofar as they bound regions." with which the decision criterion categorizes the Further extensions of this technique are exemplified pattern measurements." (Greaniasetal.,1963, by Kamentsky (1962), Schultz (1963), and Sublette p. 14). and Tults (1962). "Thus, the problem of feature selection consists 5.13"The main concept of the system is the of three parts as follows: use of a collection of programmable pointers, which 1. To find criteria for measuring the separability arevisualizedasafamily of 'bugs'...the of pattern classes. PLACE statement initiatesor sets up a bug by 2. To evaluate the effectiveness of each feature assigning a name and initial coordinates to it... or each mapped feature to the criteria. The MOVE statements move the bug a specified 3. To find the mappings which allow us to use distance (i.e., number of spots) in either the x or smaller dimensions, keeping the criteria at type y directions...A series of statements... certain values." (Fukunaga, 1968, p. 1). TEST the grey-level value of the picture at the location of the bug...A bug's grey-level may be 5.11a"The purpose of data analysis is to extract changed in statements that are particularly useful those features most valuable for discrimination for 'thick line' analysis are the STICKs. If a bug among the allowable characters; i.e., those features wishes to walk along the middle of a 'thick line', 105

1 these statements will readjust the bug location to screensto provide mappings that discriminate the 'middle'... criterial features. In the second level of operations, "Bugsys can be used for any types of applications coincidences are detected for relative dispositions ...(including the analysis of photomicrographs of the features in the characters making up the ofneurondendritesand...processingof alphabet. For example, "h", "u", and "n" will have Schlieren photographs taken for molecular weight. the same parallel-verticals feature, the possibility determination when using ultracentrifuges."' of "u" will be eliminated by the mapping for curves, (Led ley et al., 1966, pp. 79, 81-82.) and a long-ascender feature will show "h" rather "The x- and y-coordinates of the curve are than "n", or, conversely, the negative of the long- measured by taking the distance between bugs AA ascender will detect "n". and origin for x and the distance between bugs AA It is to be noted that the input image may be a and BB for y." (Led ley et al., 1966, p. 83). sequence of characters, such as a complete word of 5.13a"The processing of pictorial information text, with or without spacing between its characters. contained in aerial photos has been considered inThus, prior segmentation is not required since in detail." ("Technische Hochscule,Karlsruhe, effect, the technique scans the text through a moving ResearchonPatternRecognition...,1966, aperture, and checks continuously for specific abstract, p. 1). groupings of local features that will identify the "A system development consists of special scan- characters, one by one. In this connection, simula- ning and preprocessing devices for pictorial data tion experiments have been run on computer for all is going on. The scanner and preprocessor combines possible pairs of 10 alphabetic characters with no analog and digital techniques. The scanner is space between them, detecting 12 different local realized as flying spot scanner. The scanning modes features in the first level of operations and 4 dif- are controlled by a computer Control Data CDferent possible definitions for each character at 3300. The processor generates directions of contour the second level. While false recognitions occur gradients, grey shades, contour coordinates and (e.g., two quantized "o" 's detected at one stage as otherlineelements."("Technische Hochscule an "x"), the results appear promising. Further re- Karlsruhe, Annual Report 1966", 1967, p.52). search will apply threshold requirements with re- "The Karlsruhe system for automatic photo-spect to clumps of the most probable character interpretationisa scanning and preprocessing values. Statistical data from tests to date is available system under computer control." (Kazmierczak as to the extent by which recognition improves as and Holdermann, 1968, p. 47). spacing or segmentation is introduced. (See Stevens, 5.13b"There are other methods currently in 1968, pp. 6-7 and Clayden et al., 1966). vogue in the field of character recognition for elimi- 5.15a"Characters may be recognized in terms of nating the effects of such distortions as stretching, properties more abstract than geometrical features. skewing, magnification, etc. Many of these may be For example, Giuliano, Jones, Kimball, Meyer and referred to as 'feature detection' methods, although Stein (1961) and Alt (1962) have obtained the higher they variously go by the names: n-type detection, moments of patterns, blackness being analogous zoned n-tuples, stroke detection, lakes and bays, to mass; and Horwitz and Shelton (1961) and Clowes etc." (King and Tunis, 1966, p. 70). and Parks (1961) have used autocorrelation as a 5.14The phraseology,"autocorrelation tech- firststage of automatic recognition.Generally niques for pattern recognition", is in a sense am- speaking, the output of the first (preprocessor) biguous and misleading. Target pattern identifica- stage is a set of numerical values. It is often found tion operations in these techniques (whether applied expedient to multiply these values by weights in toprintedcharactercharactersortospoken the decision process, and several methods for deter- numerals or words) are as much dependent upon mining suitable weightsare now well known. cross-comparisons and cross-correlationsof the Statistical weighting was proposed by Selfridge input pattern with the set of available reference (or (1955) and implemented for example by Doyle master) patterns as in other techniques. The point (1960) on results of fairly complicated geometrical is rather that, in the autocorrelation techniques, the tests. Following Roberts (1960), Duda and Fossum . transformations performed upon the source pattern (1966) have experimented with a perceptron-type to generate the input pattern or further to process systematictrialand error method forfinding the input pattern depend upon autocorrelation func- weights, and also for finding more than one set of tions of the source pattern as it is variously manipu- weights per character." (Ullmann, 1967, p. 256). lated in order to extract those criteria! features or "Misregistration arises not only from the char- properties that are to be used in the recognition- acters out of line on the document, but also from decision process. variations or instability in the doucument handling 5.15In particular, a multi-layer hierarchical sys- and optical equipment. A generel method of elimina- tem is currently under investigation that applies ting such effects is to make use of measurements some of the original ideas of auto-correlation but no or transformations on the image which are invariant longer integrates. Instead, both positive and nega- with respect to displacement within an allowed tive copies of the input image are superimposed field. One such transformation which has been de- and shifted in various directions and with special veloped by several workers is first order autocorrela- 106 tion. Its use with variations has been reported by optical systems to achieve the analog of the above G. L. Shelton and L. P. Horwitz of IBM, andby mathematicaltransformationposeschallenging Clowes and Parks. Higher order autocorrelations problems." (Brill, 1968,p. 1). involving products of threeor more displaced 5.16"The number of optical readers presently images have also been studied." (Andrews, 1962, in use is estimated at 1,200 (200 being character p. 288). readers), with a steady increase foreseen." (Auer- 5.15b"Since the [Laplacian] function is deter- bach Corp., Source Data Automation, 1967,p. ined by brightness derivatives rather than by bright- 2-106). ness itself, it produces no output on large uniform "There are currently nineteen different scanning areas. In the presence of line boundaries or cor- machines manufactured in the United States." ners, the output is high. Thus the function tends to (Bus. Forms Int'l., Inc., 1967,p. 1). outline objects and emphasize their contours." "There are five major competitors fora basic (Kanal and Randall, 1964, p. D2.5-4). scanner business that is estimated at $100 million. 5.15c"Another application in which Hilbert The companies are the Burroughs Corp..Farrington transform filtering can aid the sensitivity ofa Electronics,Inc., National Cash Register Co., processing operation that follows is in signal detect- Philco Corp., and the Rabinow Engineering Co." tion by optical matched filtering. By performingan (Kornberg, 1964, p. 116). operation on the input that is related toa differen- 5.16a"At present, thereare 14 manufacturers tiation, viz, a Hilbert transform, and the.sametrans- in the business segment of the field, anda 15th, formation is made of the reference imageor filter, Scan Data in Norristown, Pa., whichis believed then the matched filter performs, in effect,a deriva- to be staffed by ex-Philco-Ford personnel, isex- tive correlation. In this manner it is possibleto pected to have a hardwareannouncement later achieve a high degree of correlation fora signal this year. In addition, Western Union is workingon and a reference filter,even though significant dif- character readers, but these developmentsare still ferences in the low frequency noisecontent of the in the laboratory stage." (Drattell, 1968, two may exist. There is the added advantage that, p. 38). 5.17"Farrington Electronics, Inc....Model at least in the ideal case, no loss in signalenergy 3030 Page Reader reads multi-linedpages as is incurred as a result of the two transformation many as seventy lines per page." (Business Forms operations." (Haagen, 1968,p. 9). International, Inc., 1967,p. 8.) 5.15d"Of late, considerable interest hascen- "The Farrington Model 3030 isa machine that tered on the generation of the Fourier Transform of can read ledger or bond paper documents, orpages pictorial information, largely for thepurpose of up to legal size. Each line of information to be read performing contrast enhancement and othertypes may contain up to 75 characters and spaces. Char- of selective filtering. To this enda number of algor- acters are read at the rate of 400 per second." (Busi- rithmshave emergedin particular,the'fast ness Forms International, Inc., 1967, p. 17.) Fourier Transform' for producingdigitalap- "The Rome Air Development Center, Information proximations to the FoUrier Transformon a com- Processing Laboratory has the onlypage optical puter. All of these are, however, basically ofan scanning equipment which reads both off -line nature, since upper and none of the algorithms is fast lower case alphabetical letters, the numeralszero enough for real time applications,at least where to 9, and punctuation marks for a total of 70 char- standard video frame ratesare concerned." (Pop- acters. This Optical Character Recognition equip- pelbaum and Faiman, 1968,p. 1). ment [Farrington] has been in operation in testing "This paper describesan approach for trans- auxiliary research on lexical data handling forover forming the exterior boundary of a character into two years." (Shiner, 1962, p. 335). a unique one-dimensional waveform. This waveform "And at Time, Inc., in Chicago, wheretwo Far- is neither sensitive to character orientationnor rington page readers andone Farrington document affected by scale, and it offers exceptional possi-scanner are being used by the publishingcompany bilities for implementation into recognition schemes. as basic input on all subscriptions (includingnew, The waveform is expanded into Fourier series and change of address, payments), OCR isestimated to the coefficients of the series are usedas descriptors be 15 percent cheaper than other input sources." of the character. It is shown that certain recognition (Drattell, 1968, p. 36). techniques, employing these Fourier descriptors, 5.18"The Electronic Retina Computing Reader can be developed which eliminate or significantly presently is reading witha substitution or error rate reduce all the shortcomings mentioned above. of less than 1 error per 100,000 characters. To "These recognition techniqueswere simulatedachieve this degree of reliability with keypunching by discrete methods on a digital computer,pro- would require 100% verification." (Perry,1966, ducing excellent results in recognition experiments p. 132). with both machine-generated and hand-written "The system consists of: A rapid indexpage characters. However, in practice, real time analog carrier, an Electronic Retina and recognition unit, systems are to be preferred over digital computa- a Scientific Data system 910 computer, and two tions because, in general, they are faster and less Ampex magnetic tape stations whichare com- costly. The design of real-time opticalor electro- patible with RCA. Hardcopy is fed into the system,

366-107 0 - 70 - 8 107 and the characters are read by the Electronic tion directly from the document it has read onto Retina. This hard copy, prepared with special magnetic tapeunitsforlater data processing typewriters developed for us by Olivetti, can be functions." (Janning, 1966, p. 114). editorialtextmatter,classifiedadvertisingset "We haveano-compromisemachinethat solid,semi-display,display,legals,accounting takes advantage of the matrix technique.... data,or any other alphanumericinformation. Our character display...is a 680-element dis- The Electronic Retina converts the typewritten play which we believeiscurrently the highest characters and spaces into predetermined elec- resolution being offered in mat-ix type recognition trical impulses which are stored on magnetic tape. systems.Itcombinesfeaturerecognitionand The magnetic tape, in turn, produces justified template recognition. The machine has five common paper tape to drive the linecastingmachines type faces ...five [more] come at slightly extra- either hot or cold type. Printed or typewritten cost options." (Gibbs and MacPhail, 1964, p. 96.) copy is read up to the rate of 2,400 cps(approxi- "The General Purpose Reader recognizes all of mately 28,000 words per minute)." (Perry, 1967, the common business font styles in upper- and p. 86). lower-casealphanumeric, andsymbols. Up to Text corrections can be accommodated as fol- 10 fonts may be included in one machine." (Com- lows: "The typist then retypes the lines, or seg- mun. ACM 8, 255 (1965).) ments of lines affected by the changes, using a "Cards or documents on paper up to legal size new piece of copypaper as a correctionsheet. (71 x 13) may be read." (Commun. ACM 8, 255 The correction sheet then- is fed into the Elec- (1965).) tronic Retina ahead of the original copy which Further, a New ERA Data Systems advertisement is to be corrected. The reader stores these cor- in the October 1966 issue of Business Automation rections anditscoded instructions concerning said: "If you're an average reader you'll read this them in its memory system. Then, as the regular ad in just 34 seconds. Our Philco Reader read it text comes in behind the correction sheet, the and put every word directly onto magnetic tape in Electronic Retina's computer determines whether 0.6 sec." there is a correction in each line. If there is one, 5.20"Another possible application...of a the Retina wipes out the original copy and accepts page reader would be the ability to reduce proof- the corrected line, or lines, inserting the revised reading. To take the original manuscript, read it material until the correction.is completed." (Perry, into a type-composition system, through a page 1966, pp. 130-131). reader, set the type, make galley proofs, and com- "Another Electronic Retina Computing Reader, pare these through a page reader with the original recently delivered to the Swedish Postal Bank manuscript which was stored on discs or tape ..." inStockholm, automatically processespartof (Merz, 1964, p. 85). thedaily banking transactionsfor more than "Additional areas of applicability of automatic 500,000 account holders. Daily banking volume character recognition techniques include the check- averages one to one and one-half million transac- ing, verification, and proofreading of machine out- tions, but sometimes reaches two and one-half put, such as tables of numerical values, ballistic million.The OCR systemreadsinformation tables, code-book data, printed indexes, and biblio- from original turnaround documents, formats it, graphical lists. The use of high-speed on-line or convertsitdirectly to magnetic or paper tape, off-line printing of the results produced by com- and prints out a list of specified information and puter generation and processing of such data totals to each account holder. The system processes emphasizes the need for high speed verification 900 to 1,200 documents per minute, depending processes to keep pace with output." (Stevens, on their size and the number of linesread from 1961, p. 15). each. Previously, daily transactions were handled 5.21"When a scanner for ordinary typing manually for sorting, preparing totals, checking, becomes available at a reasonable price, the pages and booking on ledger cards." (Philipson, 1966, of an author's manuscript can be fed directly into p. 130). the scanner, without retyping for conversion to Recognition Equipment, Inc.'sdeliveries,be- magnetic tape. Colored pencil marks can be made ginning in1964, now include use by a service atpoints where editingisdesired." (Markus, bureau agency, the Data Corporationof Los 1965, p. 3). Angeles (Business Automation 12, No. 6, 86 (1965).) "The potential advantages of a page scanner are: 5.19"ThePhilcoGeneralPurposePrint 1. The input keying of the library surrogate can Reader, which employs the flying spot principle, become decentralized. The elements of the is a multi-font reader which can be programmed surrogatecan bctypedon a document to read a variety of different type styles throughout `traveler' and added by one stationafter a fixed customer installation. A goodpotential another, thefinalstationperforming the application would be an installation where there's a finalediting on the surrogate." (Wishner, varietyofdifferenttypewritersinvolved,with 1965, p. 240). different fonts. This scanner will read up to 2000 "One method of cost reduction that looks promis- characters per second and transmit this informa-ing is the use of multiple remote scanning equip- 108 ment to feed a single recognition logic assembly. reference vocabulary, the recognition criteria being Thus, if a user has only a small number of docu- the least value of the time integrals of the total ments to be read, a simple device can scan the number of absolute differences between the incom- documents, transmit this video information to a ing video and stored descriptions of each character central processor or to a magnetic tape recorder." in the vocabulary." (Greenly, 1966, p. 3.) (Rabinow, 1966, p. 24). "The addition of two fonts...to the Link Page 5.22"...The IBM 1975 Optical Page Reader Reader required minor modifications...The (which is installed at the Social Security Adminis- system had previously been designed to read two tration). The 1975, with a contour pattern recogni- type fonts one of which consisted of 55 characters tion process controlled by an IBM/360 model 30, while the second included 19 characters. The can recognize characters in more than 200 different memory for each of these fonts and for each of the type faces, and probably in wellblocked handprint- fonts investigated under the subject [RADC] pro- ing. Reading 650 lines a minute from a standard gram is operator selectable." (Greenly, ]966, p. 4). Social Security form, it can read and record in eight 5.24"Sylvania is now developing an extremely hours what a keypunch operator can produce in versatile page reader capable of reading up to 100 days." (Van Dam and Michener, 1967, p. 189). 30 typefaces, where the size of each one may "The Page Reader system processes the employee range from 6 to 36 points, at speeds above 1000 earnings at a rate equal to 120 to 150 keypunch characters per second, on pages of complex format operators with an equivalent or better error rate." such as those found in technical books and journals. (Hennis, 1967, p. 2.) "Early in 1961 a decision was made to build a [IBM 1975 Optical Page Reader]. "New tech- feasibility model of a general-purpose page reader niques in video processing such as contrast control, with the aim of proving the practiciabilityof dynamic video thresholding and data reduction certain newly devised techniques for character circuits have been added to minimize the usual recognition. The initial design goal was to attain difficulty in recognizing seriously degraded forms maximum versatility. This led to the following and printing." (Commun. ACM 9, No. 9, 706 (Sept. specific requirements for this machine: Speed a 1966).) reading speed of at least 1,000 characters and "The 1975 Model 3 Optical Page Reader has preferablyhigher;TypeFontsthemachine an effective speed of 64,800 input lines per hour... must read characters independently of size within Approximately 50 per cent of the total load is broad limits, and must have a capacity for at suitable for scanning. The scanner will be able least 30 distinct type styles; Optics the machine to do the work of approximately 200 key punch should read the original paper document, not a operators in approximately 2,000 manhours and photographic negative; Accuracy the undetected relieve computers used to convert cards to mag- error and reject rates should be as low as possible; netic tape of about 885 hours of work per year.... Flexibility the machine must use a control and "Rough calculations indicate that from 35 to scan techniquecapableof being programmed 40 key punch positions would have to be displaced to read pages with complex formats, such as those if the necessary capital investment [.--- $800,000] encounteredintranslation and informationre- is to be written off over 5 years." (Lannon, 1967, trieval applications. This class of applications is p. 50). characterized by a wide variety of printing and 5.23"A peephole template matching technique formats, and relatively high quality. The tech- developed by General Precision's Link Group niques also show promise of being suited to lower employs a mask or template in which onlya rela- quality print such as typewriting. tively small number of selected sub-areas of the "Design and construction of a reader with most image field are used as apertures for matching... of these properties is complete, and feasibility A laboratory page reader having a vocabulary of of the above goals has been demonstrated. Further four type fonts has been successfully demon- testing and design improvements are now being strated." (Maass, 1965, pp. 18-23). made." (Gray, 1963, p. 85). "Link Division, G.P.I., for severalyears, has been 5.25Rabinow demonstrated this,Anachine at actively engaged in programs directed towards the Interdata 65, see also Business Automation 12, No. development of advanced optical character recogni- 7, 44 (1965): "Control Data Corp., has introduced tion techniques...Several reading techniques the 915 Page Reader which reads information pre- have been investigated; however, the most success- pared on a typewriter equipped with American ful one uses an electronic peephole matching princi- Standard Assn. optical character reading font... pal in which only selected sub-areas of the image can handle documents from 4 to 12 in. wide, 2f to field are used as apertures for matching. The method 14 in. long, and continuous fanfold sheets." (Bus. allows complete disregard for serifs and other typo- Automation 12, No. 7, 44 (1965).) graphic and stylistic embellishments..." (Greenly, 5.26"Multifont Optical Unit Reads 2000 Char- 1966, p. 3.) acters per Second. Tokyo Shibaura (Toshiba) Elec- "Each character is scanned by a column of photo- tric Co. has developed a high-speed multifont opti- detectors and converted into a digital waveform from calcharacter reader capableof reading 2000 which selected portions are matched against a characters per second and 1200 lines per minute. 109 "Developed by Toshiba central research labora- materials are becoming available in the form of tory, the device's character-recognition logic needs computer-readabletapes. Wehaverecently no adjustment to read as many as six of the widely discovered though that having materials computer- used numerical fonts. readable is not the entire answer it is still neces- "A spokesman said the characters are recognized sary towrite programs that bring this material by their features, rather than by stroke, thickness into the form that we desire. This is no trivial or position. task." (Simmons, 1965, p. 220). "He said the new OCR can be used with journal 5.31"The most flexible method is to include tapesproducedbycomparativelyinexpensive a stored - program controller as part of the OCR printers. Another OCR, which that can also rec- system.This permits dynamic format changes onize hand-written numbers, is under development based on the information being read and makes at the Toshiba laboratory. convenience to the htiman user the primary con- "Characters that can be handled by the Toshiba sideration in forms design and data preparation.... OCR are ISOA, ISOB, IBM, NCR, Farrington (Philipson, 1966, p. 128.) 12F, Toshiba Stylized Font and others (optional). "With recent OCR use of programmed con- "Main components of the new OCR are a paper- trollers the use of variable format is suggested feeding mechanism, a photoelectric scanning device since the machine is much more efficient in for- and a recognition logic. mattingthanthehuman."(AuerbachCorp., "The width of journal tapes that can be fed into Source Data Automation, 1967, pp. 2-17). the unit can vary from 56 to 100mm. Line pitch is "The Farrington Page Reader, Model 3030,... 71- inch or more per line. Character pitch is 10"." has a programming repertoire of some 66 instruc- (Electronic News, 7/10/67). tions." (Feldman and Katz, 1967, p. 0210:03). 5.27"Onelarge user theSocialSecurity "Others who have done significant work on Admin. maintainsthatitscustom-made OCR programmable scanning and pattern recognition system, an IBM 1975, will do the work of 120 to are the Link Group of General Precision, Inc., 140 keypunch operators, but the agency says that Scope, Inc., The Cornell Aeronautical Laboratories, no one will lose his job. The operators will be re- and the Stanford Research Institute." (Hustvedt, assigned to other recordkeeping operations." (Drat- 1967, p. 7). tell, 1968, pp. 35-36). 5.28"The U.S. Post Office in Detroit, Michigan, "The Programmed Controller [SDS-910 com- has completed the initial testing of a pilot-model puter] enables the Electronic Retina Character OCR mail sorter developed by Philco-Ford. First Readertoperformcontexteditingfunctions, test results show that 15% of total mails can be control document sorting operations, and effect optically read at present, with a 45% figure antici- other logical operations such as justification and pated in the near future." (Feidelman and Katz, wordhyphenation forautomatictypesetting 1967, p. 0210:16). operations." (LeBrun, 1963, p. 150). 5.29"Since there are a wide variety of font types 5.32Cf., for example, the following opinions in existence, multi-font readers have been developed or recommendations: "Such 'auto-encoding' (up to 128 fonts). These depend however on manual processes would have to be based on considerable modification of plugboards to accommodate dif- insight into the physical properties and make-up ferent styles with the result that mixing of docu- of documents." (Luhn, 1958, p. 211.) ments, containing different fonts, is not possible ... "The U.S. Government is sponsoring investi- and will be possible in the future, if automatic sens- gationsintothefeasibilityof recognizing the ing is developed, only at great expense." (Auerbach formats of intermixed inputs from several technical Corp., Source Dataatomation, 1967, p. 2-10 to journals. This researchisdirected toward pro- 2-11). ducing a system which can recognize the format "The manual method, consisting of altering and then read and assemble the information for the recognition logic by manual replacement of later computer processing." (Stevens, 1965). such machine parts as plugboards and optical 5.33"Amongstthedifficultiesencountered masks, is low in cost but inadequate for reading in the processing of machine readable texts, in- a stack of documents in which the character consistencies inthe use of punctuation marks, fonts are mixed." [As also on a typical journal compounds, capitals, spacing and indentations have page]. (Auerbach Corp., Source Data Automation, been a problem way out of proportion with respect 1967, pp. 3-103). to the simple functions these devices stand for. 5.30"Today's optical readers will take selected For instance, even with the aid of a dozen dif- type fonts; some very sophisticated versions will ferent tests performed by the machine, the true take several type fonts. None, however, will take end of a sentence cannot be dc,',ermined with the number of type fonts found in a scientific certainty. It is hoped that publishers of scientific journal, nor will any handle the number of type literature will hi time sacrifice some of the niceties fonts found in the Encyclopedia Britannica. Due and aesthetic aspects of the printed page for the to the rapid increase in computer technology in sake of clarity in communication." (Luhn, 1955, the publishing industry, more and more textual p. 22). 110 5.33a"Two main typesof degradation are the other three-quarters is for paper movers, format encountered, misplaced characters and unclear control, and other handling mechanisms." (Van or smeared characters. The causes are three- Dam and Michener, 1967, p. 190). fold: mechanical, human, and accidental. Mechan- 5.36"Measurementsofprintcharacteristics ically, for example, typewriter ribbons wear, leaving such as stroke width variation, edge irregularity, light or smeared impressions. Soft paper leaves voids, extraneous ink, paper to ink optical reflec- fuzzy-edge characters. Pinion wear and jammed tivity, etc., are commonly used techniques for speci- keys cause tilt or skew in the typed symbol. Human fying acceptable tolerances of printed text.... mistakes typically cause misregistration, misalign- "[The high-resolution optical] scanner used in this ments and palimpsests (where two characters are study was designed and constructed...and has superimposed, oneonlypartiallyerased), and been used by G.E. for various pattern recognition smears. Accidents account for overlays of foreign studies which require high resolution.... material, ranging from grease to egg foo yong," "One of the most appealing approaches from a (Frank, 1964, p. 2, reprint). theoretical viewpoint would be to classify print 5.34"For optimum reader performance, print quality in terms of print noise statistics. For simple quality is by far the most critical document property. noise statistics, such as independent additive, gaus- Since quality control of printing cannot be done on sian noise, this approach can be realized. It is also a 100% inspection basis, the procedure that is possible to design optimum recognition logic and used must be able to maintain adequate perform- predict the adverse effects of this type of print noise ance of the systei.a with only periodic checks of on machine performance. For practical situations, critical items. Certain conditions, such as proper however, such simplified assumptions for print noise selection, and quality control of ribbons and paper, statistics are inadequate..." (Vitale, 1965, pp. 1, can be handled on a routine basis. Other conditions, 17, 28.) such as end-of-ribbon life and the need for printer 5.37"Ink SmudgeIn OCR, the displacement readjustments must he detected by inspection of of ink under shear beyond the original edges the printed characters. Experience has shown that of a printed character." (Unpublished definition, optimum performancecanl. be achieved in the American Standards Association, X3 Committee reader with a reasonable primer control effort, on Computers and Information Processing). if the character stroke width is maintained between "Ink Squeezeout In OCR, the displacement 10 mils and 18 mils. Characters with stroke widths of ink from the center to the edges of a character that range beyond also can be recog- during printingresultinginacharacter with nized currently. But the probability that unac- `darker' outlines than the center." (Unpublished ceptable characters will occur increases beyond definition, American Standards Association, the level for optimum performance when more X3 Committee on Computers and Information than 25% of the stroke width of the inspected charac- Processing). ter ranges down to 8 mils or up to mils." "Ink BleedIn OCR, the capillary flo-.v of ink (Greanias, 1962, pp. 143-144). beyond the original edges of a printed character." "The essential requirements for optical character (Unpublisheddefinition,AmericanStandards recognition are paper handling techniques, positive Association, X3 Committee on Computers and control of random formats, automatic error correc- Information Processing). tion, system reliability and accuracy, and economical 5.38"Light StabilityIn OCR, the resistance equipment." (Shiner, 1962, p. 335). to change of color of the image when exposed 5.35"Paper feeders, format controls, informa- to radiantenergy."(Unpublisheddefinition, tion buffers and output devices make a practical American Standards Association, X3 Committee reading machine cost many times the cost of the on Computers and Information Processing). recognition logic itself." (Rabinow, 1966, p. 18.) 5.39"TheKidder[PressCompany,Inc.] "The factor limiting present OCR speed is paper Model 081 Optical Character Tester was designed handling. Speed of paper handling can be increased to check the print quality of documents for optical by multiple scanning of documents, or by multi- scanning and can be used for analytic measurements plexing the input. One solution is to overlap the or for routine inspection. Physical measurements functions of reading and transporting documents. are made on a large viewing screen on which A Philco reader under development uses a vidicon amagnifiedimageisdisplayed.Reflectance scanner that deposits an image of the entire docu- measurements, in two spectral ranges, are made ment on a CRT and then scans the CRT surface on a meter giving readings in print contrast signal while the next document is moved into position. directlyand alsoinpercent reflectance .." Another method of increasing speed is selective (Business Forms International, Inc., 1967, p. 33.) scanning (scanning only the necessary parts of the 5.40"Designed to check that printed characters document) with variable-format readers, like the are of sufficient quality to be read by machines, Philco General Purpose Print Reader. The complex the Optical Print Quality Monitor can measure paper-handling mechanisms are also very expensive: paper whiteners and blemishes, and the dimensions, it was stated at the 1966 IEEE Pattern Recognition density and quality of print characters against Workshop that the reading equipment of a typical parameterssetupbeforehand."(TheBritish OCR is only one-quarter of the total cost, and that Computer Society, 1967, p. 126). 111 1

"Equipmentformonitoring OCR characters the character printed on the field." (Uffelman et al., also provides a CRT display, in this case of the 1967, p. 42.) magnified outline of a character as'seen' by "The spatial distributions of additive ink-noise a 0.2 mm scanning spot.The observed density about character strokes have been determined. gradient at a stroke boundary and the severity Ink noise has been shown to be stroke-related, hence of voids or of extraneous inked areas are obviously character dependent." (Uffelman et al., 1967, p. 63.) functions of scanning spot size and thus Monitor 5.44a"The functional value of available readers and Readershouldhavecomparableoptical has been another important factor in the growth characteristics. The displayed boundary can be of optical character recognition. Like anyother selectedfor any brightness level between the commodity, the value of character recognition must white of magnesium oxide and absolute black. be measured by comparing its cost to thatof per- Fig. 14 illustrates the effect of varying thislevel forming equivalent functions by other means. The from 0.51 to 0.70 on the PCS scale where white actual cost of reading is not always the controlling paper gives 0 and absoluteblack gives 1.00." economic factor. Sometimes the value of associated (The BritishComputerSociety,"Character editing and classifying that must be done manually Recognition 1967", p. 22.) may offset the potentialsavings of faster and 5.41"...Handling of paper can result in curl cheaper reading by machine. In other applications, of the sheet, creasing, local distortionsthrough the reduced document preparation cost that can uptake of moisture, and discolouration andscuffing be realized by printing machine readabledocu- of the printed characters; none of theseeffects ments on a single pass through anaccounting can be dealt with veryspecifically. The best way to machine or computer is enough to tip thebalance minimize the effects of handling would be toselect infavor of machine reading." (Greanias, 1962, a fairly stiff, smooth, not tooglossy paper, of good p. 130). dimensional stability and with high tearstrength." "The desirable features of J.C.R. asI see (The British Computer Society, 1967, p.22). In 5.44b particular, "tear resistance is relevant togener2.l them are as follows. handling and may be an important consideration (1) Economically to replace key-punch operators for the preparation of documents on line printers, from the point of view of: in which circumstances the directionof the grain (a) Cost of salaries and cost of equipment. of the paper may be in conflict withthe direction (b) Cost of space and the amount of space desirable from the standpoint of the paper transport occupied. of the reading device. The choice of the optimum (c) Burden of administration and supervision combination of characteristics must be a matter of staff of large punched card sections.(It of selection according to the particularapplication." has been estimated, however, thatwhen (Ibid, p. 33). our centralization iscomplete on all com- 5.42"The inclusion in the body of the paper of puter applications (not justBilling), we will material to which the reading device is sensitive still require 85 operators even withO.C.R. is to be avoided as much as extraneousink...In equipment.) OCR applications this takes the form ofspecified (2) To increase the accuracy of datafed to the maximum dirtcount ...Fluorescent additives computer, since there is lesschance of copy- are sometimes used asbrighteners for paper but ing or punching errors. have a should be avoided for OCR use as they may (3)To reduce the time to prepare and readdata significant effect upon the reflectivity of the paper. into the computer in order to avoidpunching Water marks deserve special attentionbecause bottlenecks, and to lessen the elapsed time they may offend both against the requirementsfor required to produce the final printed output a flat surface andfor minimum opacity and are e.g., bills and meterreading slips so as to better avoided at least in the clear band."(The have more chance of giving a 24 hour service British Computer Society, 1967, p. 33). to consumers. (4) To be able to use the same typefount as will 5.43"Minimum line spacing must be sufficient and a to allow adequate separationbetween the highest be used for the rest of the document, character boundary in a line and the lowestcharacter fount which will be acceptable to consumers. the height To have a high rate of correctly readingthe boundary in the line above, in addition to (5) of of the character itself. The width of theclear band characters and marks, and also a low rate specified is commonly as large as 5/8" (16mm.)... machine breakdown. The permissible character skew maybe as little (6)Not to increase significantly the costof the vertical misalignment as stationery. as 11 degrees and the effort little as 0.007 inch." (The BritishComputer So- (7)Not to increase significantly the operator on the printer orthe maintenance attention on ciety, 1967, p. 30). the printer, and to not reduce thespeed of 5.44"Noise in the character field is dependent printing." (Paine, 1966, p. 221). on the character ...Noise is not scattered about in the character field in a random,equally probable 5.45 Figure 14 illustrates the proposedDickinson- manner. The spatialdistribution of noise resembles Wheeler embellishments. 112 ABCDEFGHJKLMNOPQRSTUVWXY (0) 4)0 ..Jo 40-30 2 N INIIMMIPror MONO NANNEININIVEI =MO .1moo 00= ImONO. SWIM 37 MS*NEIN immulran 1.1. MINNIIIIIVINIMINON MOM= NINOMMIN .ar-:"VIOLAIIIMI EL IMMENEP al1MMINSI ONINIFaltiiMINIMIIIIMIll rm. 1/AW-1 WIN= w JENNE&/"...NNO onion IM MIMEO 7 NONI 2.117M. 4.NO.11111/411 AMMO MIMI MO.m0 .1111111./AO. MIN ArmilL V =MN MN VO-11/011 110/...rININNIIIIIIIIL - Ps- AO NM 11111INNIMI EINEM ob. I== MINED MIMI Ili III a ORO MMO .JNEN 111101111/41111 =111111111111LNVOIMILIMMIIIIMWO == 1=1211Imum --- INNINMIIINIall OMEN NM.1 SMIRIIVINIM MOM=IMO "MOM ManiUMW MOD IMIOM1111/411. g 3 ES FIGURE 14. Dickinson hand-print embellishments. (b)

...2,11.1'1144:11.4311Wy5ti -- 04...1=1.0411, Sk FOSDIC ALPHA-NUMERIC - MARKING ( GENERAL EXAMPLEPURPOSE - SR CHAR.) i DOCUMENT I 0 1 2 3 4 5 6 A B C D E I GH J,K L MHO Number Number R S T U V W X Y Z Period Comma Begli PQ DATA MARKING . ... AREA. . 1 2 3 4 5 6 7 8 9 10 II 12 13 14 111 15 16 17 18 19 70 : 4. I. 1, 70 )1 . 40 4 . 1 S, , s 1 1 " " , 60 , 6 ...... I s . . . 1 . . ." " j 00 FIGURE 15. FOSDIC alpha-numeric marking document. 3i It .A.A. 5.46One type of application is shown in Figure conduction of the marks or crossings constituting 15. the numeral written. Figure 16 is a reproduction 5.46aFor example, "in the production of script of two cards processed in an experimental reader, to be identified three constraints are placed on the based onthisprinciple,demonstratedatthe writer. First, he is asked to observe a base-line and Western Joint Computer Conferenceof1961. a guide line above it. The second constraint is that Figure 16(b) represents source patterns correctly no capitals are used. Finally, the writer is asked to read and recognized. In Figure 16(a), however, there be careful and legible in his script; we seek to is a nonrecognition of the fourth handwritten digit, identify reasonably good writing before considering because that digit, "6". was not properly formed in scrawl and scribble." (Harmon, )962, p. 151). accordance with the system constraints. Provided 5.46b"It is obvious that the use of controlled or that the required vector crossing pattern is not constrained writing may hamper the writer in at violated, however, considerable variation in the least two ways: 1) he might not write as fast as he exactshapeof the handdrawn digits can be could have, had he not been using a controlled writ- tolerated. ing system; 2) the writer may not be able to adhere 5.48"H. C. Vernon and R. R. Walsh were to the control." (Kamentsky, 1961, p. 491). granted a patent on a character recognition method 5.47 An example of the use of this method in and apparatus which attempts to perform recog- close conjunction with suitable constraints on the nition on constrained characters. The character formation and placement of handwritten numeric is scanned by a flying spot scanner in three seg- digits was demonstratedatthe Eastern Joint ments. Each segment is oriented so that its long Computer Conference held in Washington in 1957. dimension is vertical. The scanning within each Dimond of the Bell Telephone Laboratories reported segment proceeds bottom to top in a television at this Conference on a successful method for the raster fashion with the retrace occurring between automaticreading-recognition of numeric digits segments. The three scans are assumed to be handwritten by telephone toll switchboard operators non-overlapping. in the Bell System who produce in the aggregate "Several methods for constraining the handwrit- some two billion toll tickets, with 20 to 30 characters ten characters are suggested, among them a six dot each, per year. This automatic reading may be constraintanda'forbiddenarea'constraint. accomplished by either direct, by-product, machine- A variation of this system is also described in language data generation through use of a stylus which the rasters are oriented horizontally one in combination with a special recording device, above theother. Truth tablesare constructed or by subsequent machine recognition of characters showing the results of these scans on representative recorded on paper in accordance with preprinted characters. guides. "Registration is accomplished by a separate detec- The stylus-recording-inscription device is simple tor and logic system. The characters are assumed in operational principle, easy to use, and portable. to be constrained within a rectangular marked This device has been tentatively termed a "Sty- area which is detected as a separate operation lator", and itsbasic principles are as follows: by a set of horizontal and vertical registration detectors. Once registration has been established "A writing surface is provided on which there the registration logic gives way to the recognition are two guide dots surrounded by a set of criterial scheme." (Stein, 1965, misc. notes). areas consisting of seven conductors embedded in a plastic plate. As a numeral is written with a 5.48aForexample,"handwritingcanbe stylus connected to a source of potential, the separatedintox,y component waveforms for stylus energizes, one at a time, the conductors dynamic analysis or transmission using the analog in the criterial areas involved in the numeral. or digital versions of stylus position." (Rose, 1965, The combination of areas energized causes cer- p. 639). tain flip-flops in a translator to operate and drive 5.49"Groner reports some interesting work the rest of the translator to indicate the correct inhandwritingrecognitionforreal-timecom- numeral." (;imond, 1957, p. 236). puter input. The system involves the use of a In other words, the conductors marking the RAND tabletandafountain-pen-likedevice significant areas serve as vectors which, if crossed together with a CRT display. The recognition by the stylus as the numeric digit is written, will scheme used issurprisingly simple, yet rather in accordance with specific vector-crossing patterns, successful, and a number of editing featuresare serve to identify the digit that was written. In the included to permit the user to make corrections Bell Stylator c'evice, the sequence in which the quickly. This reviewer, who had an opportunity stylus crosses various vectors as the characters to use the system for a few moments, found the are produced is important for recognition purposes. use of a "scribble" character as the rubout command delightfully natural, althoughsome other An independent invention was disclosed in the results of this short trial were not as satisfactory. Johnson patent, assigned to IBM, which provides Some much more meaningful testresultsare for the use of two centering dots and of radial described in the article. An indication that recog- areas extending from these dots for the sensing of nitionof handwrittencharactershasperhaps 115 IBMEXPERIMENTAL CONSTRAINED HANDLETTERING READER

I I 1 1 [4-1 FL7 TV: LLL LC2/ L 1 it 1 1

2 2 2 2 2 I I I _J L_JL_J L _J L L _J L97LTCY _J 3 3 3 3 3

4 4 4 4 4

51 5 5 5 5 LETTER NUMBERS AS SHOWN ABOVE.

6 6 6 6 PINCNEI NILES T1 TIE 11511 IF NUMBERS MIMES NINIER RECHNIZEI.

1 7 7 1 7 ISE ANYSIFT LEO PENCIL.

I I I 8 I

91 I

IBMEXPERIMENTAL CONSTRAINED HANDLETTERING READER

I I 1 I 1

1 11 1 1 ri--7-1 IT I n 2 2 2 2 2 7 7 IT I I I1'1-1 4) I ih 11.10Is l I 1_ L 3 3 3 3 3' LL/ I_ LI 1_

4 4 4 4 4

51 5 5 5 5 LETTER NUMBERS AS SHOWN ABOVE.

6 6 6 61 6 PINCNEI NILES T1 THE ROT IF NUMBERS MUTES NIMIER IECHNIZEI.

1 1 1 1 1 HE ANYSIFT LEA. PENCIL.

I I I 1 1 I I II I I

FIGURE16.Handwritten numerals, verfor crossing technique. not quite 'arrived' commercially is a short new- 5.50"A patent was issued to H. B. Crane on a product announcement on a commercial device device which is hand held and which can be capable of reading the ten digits and five alpha- used as input for the recognition of hand-written numeric characters; the purchase price is quoted numerals or characters as they are written. The as $162,000!" (Mills, 1967, p. 233.) device consists of a series of contacts surrounding "The scheme responds very quickly even though a stylus. As the hand is moved inthe writing of it recognizes a fairly large set of symbols. Moreover, characters, the contacts are made in a sequence itimposes few constraints on style, speed, or which is peculiar to the character being written. position of writing. It makes use of contextual An analysis of the sequence and duration of contact information to distinguish symbols which cannot closures then reveals the identity of the unknown be distinguished by shape alone. This scheme character." (Stein, 1965, misc. notes). has been used daily at The RAND Corporation 5.51"A patent has been issued to L. D. Harmon for using computer code, drawing flow charts, on a device fin- automatic readingof cursive script. andediting.The symbol -recognitionscheme As an operator writes the word in question, his is written in IBM System/360 Assembly Language stylus moves across a writing platen. The surface and runs on an IBM System/360 Model 40." (Groner, is made of electrically conducting strips arranged 1966, p. 591). horizontally. Features are extracted as the stylus 116 is moved. These may include loops, parts of char- "The character recognizer is in the final stages acters extending above or below certain horizontal of development. It uses a character dictionary built strips, etc. The horizontal reference of the logic for each individual user, andcan recognize alpha- is controlled by the stepping ofa time commutator bets in excess of 80 characters. The editor permits which is stepped by every crossing of the stylus characters to be erasedor written over; entire ex- over the zero axis line. Thus differences in the hori- pressions may also be deleted. The analyzer isa zontaldirectionare normalized."(Stein,1965, part of the PLAINT course-writing and computer- misc. notes). assisted instruction system. It will compute the 5.52"An'electronicdraftingboard'con- value of an expression written in ordinary mathe- structed at M.I.T. has been described by L. G. matical notation. Figures 3 and 4 showsome of the Roberts. X-Y coordinate information is produced input forms and resultant displayed output. The in real-time for direct input to the computer bymov- parsing algorithm and allowable notation are still ing a stylus consisting of a pin and a set of coils quite limited, but we are workingon more sophisti- across a 20 X 20 inch sheet of paper laid over a cated parsing and editing facilities that will allow specially wired base. The base is cross-wired in two the user the same freedom of expression he has directions and selectively wound in and out of pulse using a blackboard or a pencil and paper." (Bern- transformer cores to provide coded drivecurrent. stein and Williams, 1968, pp. 28-29). As the stylus is moved over the wires, thecurrents 5.53"As customers make purchases at Higbee's, that are picked up provide coordinate information sales clerks write pertinent informationon specially for the computer by means of the code selected by designed IBM card sales checks...This hand- the wired cores. Approximately 100 machine in- written information indicates the quantity, depart- structions are sufficient to identify a character from ment number, merchandise number, and amount a list of characters that the particular user has of the transaction." (Greanias, 1965,p. 62.) previously drawn." (Stein, 1965, misc. notes). 5.54"IBM MACHINE READS HANDWRIT- "Earnest has developed a set of simpletests on TEN NUMBERS. A machine that reads handwritten the strokes in a word which enablesa computer to numbers is IBM's latest offering. The IBM 1287 pick likely words from a dictionary. Presentlyoptical reader can recognize numbers and five these tests are to count the closed loops, the tails different handprinted alphabetic characters pencil- above and below the small letters, and the axis written on a wide variety of business documents. crossings at the center of the word." (Roberts,The first with this capability to be offeredcom- 1962, p. 210). mercially, it feeds this information directlyinto a "For separated letters, Teager estimates that 100 computer for processing. machine instructionsare sufficient to identify a "There are two models of the 1287. Modl I reads character from a list of characters that theuser hand-pencilled,printed,imprintedandpencil- has previously drawn." (Roberts, 1962,p. 211). marked cut-form documents. Model II reads printed 5.52a"The character recognizer included in cash register and adding machine journal rolls in SHAPESHIFTER was written by Kenneth Ledeen addition to cut-form documents. The 1287 is de- at Harvard University. With a training program the signed to be used with System/360 Model 30, 40 user teaches the computer his own handwriting; or 50.. this information is punched onpaper tape and is "Both 1287 models will be available during the read into SHAPESHIFTER atrun time. In effect first quarter of 1968. Monthly rental of the 1287 each user has a version of the character recognizer model I begins at $3,600; purchase price beginsat tailored to his own writing style....The recognizer $162,000. Monthly rental of the 1287 model II begins must .supply five pieces of information: the name of at $4,000; purchase price begins at $180,000." the character and its four descriptive coordinates, (Commun. ACM 9, No. 11, 829 (Nov. 1966).) the x minimum, x maximum,y minimum, and y maximum of the area where it was written on the tablet. "Only two OCRs capable of reading handprinted data have been announced. IBM recently released "After a character has been printed andrecog- the 1287, which reads handprinted numbers and nized, a canonical form of the character, in thesame five letters but imposes strict constraintson the form and of the same dimensionsas the printed printing. The computer-controlled 1287can read character, replaces the 'ink'on the scope. This 300 machine-printed numbersa second, but slows both 'neatens' the display and tells theuser how down when reading handprinted numbers. An inter- the written characterwas interpreted." (Lewis, esting error-checking mechanism in the 1287causes 1968, p. 721). ..._ , any machine-unrecognizable character to be dis- 5.52b"The major software component of the played on a CRT for the operatorto recognize, rather system is the Q-32 Time-Sharing System (TSS). than immediately rejecting the document. [This TSS currently serves inexcess of 30 users at one same feature is discussed as it applies to film scan- time with a limitation of 47K wordsas the maximum ning in Michael's forum and by Gott.] The 1287 costs program size. To support the terminal, we have $180,000 and rents for $4000 a month." (Van Dam prototype versions of a handprinted-character rec- and Michener, 1967, pp. 189-190.) ognizer, an editor, and an expression analyzer, all "Both International Business Machines Corpora- of which operate under TSS. tionandRecognitionEquipment Incorporated 117 1

announced late in 1966 that they had developed 5.58"A character reader which reads hand- equipment capable of reading hand-printed or handprinted and machine-printed figures (on the same pencilled characters." (Business Forms Interna- line, if desired), characters, and symbols, converting tional. Inc., 1967, p. 5.) the data for magnetic tape has been introduced by Optical Scanning Corp., called the Op Scan 5.55"RecognitionEquipment announcedin 288..." (Data Proc. Mag. 9, No. 6, 72(1967).) late 1966 that it had developed a scanner that "New System Reads Hand-Printed Characters readshand-printedcharactersintermixedwith Computers. A new typed or printed characters in a variety of different, as Data for Direct Input to standard type styles at the rate of 2000 characters optical character scanning system which reads per second and processesdocuments containing both hand- and machine-printed characters and this information at rates up to 1200 per minute. transfers the data directly to magnetic tape for The system can be supplied with the ability to input to computers has been developed by Optical Scanning Corporation, Newtown, Pa. ... recognizeavarietyof hand-printedcharacter shapes tailored to the individual user's specific "The system is free-standing and operates offline. requirements."(BusinessFormsInternational, The basic price is $98,088, or a customer may Inc., 1967, p. 29.) "Recognition Equipment, Inc. rent oi.e for a monthly fee of $1,988.... (Dallas, Texas) have announced a new module "Op Scan 288 reads at a rate of about 600 one-line for its Electronic Retinal Computing Reader which documents a minute. The unit can distinguish enables the automatic recognition of handprinted 10 digits, six letters, and the signs for plus and letters and numbers with a claimed speed and relia- minus. As many as 25 hand-printed or 80 machine- , bilitypreviouslypossibleonlywhenreading printed characters can be accommodated on a machine-printed information. The module, which singlescanningline.Bothhand-printedand can have a vocabularyof up to 40 alphanumerics, machine printed characters can be placed on the permits a considerablevariationin characters same scanning line.... and reportedly provides computer input that is "OpScan288acceptsdocumentsoffrom 2-1/2" X 3-1/2"to 4 -1/2" X 8-1/2". The documents 200 to 300 times as accurate as that provided by tabulating- unverified keypunching." (Composition Information can be on paper ranging from 20-pound to Services Newsletter, Oct. 1, 1967, p. 3). card stock.. .. "OpScan 288iscompatible withallmajor 5.56"InformationInternationalIncorporated computer series." (Commun. ACM 1 1 ,No. 5, ...has a system known as Programmable Film 386, May 1968). Reader-3 which performs a variety of pattern recog- The OpScan 288 allows the 10 numerals, the nitiontaskscontrolledby computer program. letters C, N, S, T, X, Z, and plus and minus signs Included in its capability is character recognition either machine printed or constrained handprinted. of a large number of type fonts and hand drawn (See Fig. 17.) characters. Scanning is done by a programmed 5.59"We haveplanstouseconstrained flying spot. Recognition is done by a variety of hand-printing whereby the hand-printing is placed means: matrix matching,feature analysis, curve in individual character boxes." (Brick, 1964, p. 87.) tracing and others. The recognition logic can be 5.60 "We have work going on in our develop- updated on the basis of experience, and itwill mentlabsonreadinghandwrittenmaterial." purge itself of statisticallyinsignificant erroneous (Huntley, 1964, p. 95.) entries." (Hustvedt, 1967, p. 7). "Other new products are on the horizon including I 5.57"Several manufaCturers, including IBM, a unit from Farrington to read hand-printed numeric Philco, REI, CDC, and III, have scanners which characters. The machine, to be announced shortly, can do fairly well at reading carefullyhand drawn will join hand-printed numeric character recogni- numbers, but at rates considerably slower than for tion systems now marketed by IBM and REI. typed material; some also read upper case letters. The latterwillinstallahand-printing reader None holds out any current hope of reading script." module for its Electronic Retina Computing Reader A (Hustvedt, 1967, p. 2). in July. The device reads ordinary hand-printed letters and numbers at speeds up to 1,200 docu- "Phi lco-Fordhasseveral developments. The ments per minute." (Drattell, 1968, p. 54). company has been testing a hand-print unitthat reads alphanumerics at Educational Testing Service 5.60a"As a basis for the subsequent stroke in Princeton, N.J., a not-for-profit company that, analysis, we decided that our elementary decisions among other things,administers and scores college should supply information on the orientation of the board exams. In addition, Phi lco-Ford will deliver stroke in each elementary area in which it has been next December an optical film reader that rcads detected. This led no to a system yielding . ..a photographic film. The system will be installed at vector field, obtained by a detection process carried Pacific Telephone & Telegraph Co. in San Fran- out on the raw, unquantized video signal. cisco, where it will read photographic film of trunk "Such a field, can be obtain-d from a set Of space line registers. It is believed to be the first successful filters matched to the various possible orientations marriage of microfilm and optical scanning capa- of the charactei strokes. Each time that the scan- bilities." (Drattell, 1968, p. 54). ning pattern sweeps across a stroke, that filter 118 OP SCAN288 DEMONSTRATION FORM

LI0123456789234567891

OPTICAL SCANNING CORPORATION OS 4916 -2 ® 1968 OPTICAL SCANNING CORPORATION0.1 P.O. BOX 40 NEWTOWN, PENNSYLVANIA 18940

FIGURE 17.OpScan document. whichismatched to the corresponding stroke writers with onetime ribbons, these classes will orientation should give an output exceeding a certain correspond Lo our intuitivedecomposition into decision threshold; the other filters should remain classes. For degraded print,or hand-lettered 'block' silent.... letters and numerals, the noise willcause the classes "A suitable measurement for thispurpose is to be enlarged. Thus, if the tail on the '4' gets lost, the measurement of the ink-density gradient, aas happens all too often, it enters into the '0, D' vector perpendicular to the edge of the character class. A second level of description will enableus stroke...This gradient is quantized in direction to distinguish, within each class distinguished at to ±22.5 °, and after the decision threshold, itis the first level, the differences defined by the general normalized in amplitude. The input pattern is thus form of each of the contours composing the elements a set of unit vectors corresponding" to the (direc- of the ideal figure. Thus, in the class '0, D, 4',we tionally quantized) orientationsof the detected shall find that at the left there isan arc in the '0', a character strokes." (Nadler, 1963, p. 815). vertical straight line terminated by two right angles "The model also read (carefully drawn) hand- in the 'D', and an oblique slanting upwards to the written numerals, along withthe three design right in the '4'." (Nadler, 1967, pp. 36-37). fonts." (Nadler, 1963, pp. 815, 821). 5.61"The only commercially available reader "Since we are looking for the boundary between of handwritten characters is the IBM 1287, although the area which 'belongs' to the character and thePhilco-Ford, Control Data, and Burroughs have `outside', it appeared natural to us to seeka meas- developed such readers for the U.S. Post Office urement system which would give us this boundary Department."(FeidelmanandKatz,1967,p. directly, rather than to attempt to deduce it from 0210:11.) the usual type of OCR measurement (thegenera- "Burroughs [Zip reader]uses a five-shift-register tion of a rectangular matrix containing merely and the 'best match' code-comparison metl'od, with information on positions presumed 'black' and a flying-spot scanner." (Kornberg, 1964, p. 116). positions presumed 'white'). We detect the contours "Rabinow's [Zip reader]...light pipes, in an directly in terms of their local orientation in theimage converter tube, pick up the moving address, scanning matrix. We distinguish only eight direc-reproduce it on the back surface of the tube and tions, in jumps of 45°; thus, in place of simpleyes-no shift it up and down at high speed. information, at each pointwe decide whether we "The characters are scanned bya light-sensitive are observing an element of the contour, and its cell behind a pinhole. The cellgenerates electronic direction.... signal patterns corresponding to the specific char- "At the first, and most global, level the principalacter on the envelope. These are recognized and structural elements of the signare determined. compared with the address-directoryon a 'best Thus the '0' is defined as a simple loop, the '4'as a match' basis." (Kornberg, 1964,pp. 116, 119). loop with a tail, descending, the 'd'as a loop a 5.62The Zip Code Reader designed for the Post tail ascending, etc. It is quite clear that at this level Office Department by NCR includes multi-level fea- we cannot do more than to regroup the various ture recognition logic. Eleven blocks of logic are signs into classes, within each of which the struc- used, known as reference functions. Analogcross tural descriptions are equivalent. Thus `B, 8', '0, D',correlation is done in three separate levels. The etc. For high-quality print, such as electric type- result passes throughan A/D encoder, and then goes in to a recognition matrix before being fed to the it has not yet been achieved. It will apparently be final decision logic. (Stein, 1965, misc. notes). necessary to make the most of graphic, linguistic, "National Cash Register [Zip reader] also uses and dynamic constraints on the writing process." a flying-spot scanner, but substitutes a setof photo- (Earnest, 1962, pp. 463-465). graphic 'feature' templates as the analyzing device 5.63For discrimination between two pattern ...The machine has a 2,000-bit optical shift- types, a network of 2,048 cells (simulated on a com- register and a 19-lens beam splitter. Separate puter) was divided into two groups. Each cell, being cathode-ray tubes are used to find the address lines connected randomly to eight input bits, had a weight and scan them. The read scan-signals are fed to preassigned, and reinforcements were applied to a storage tube that displays eachcharacter se- weights of cells in the correct response set for each quentially. The character image is superimposed trial. After normalization with respect to center of on a group of photographic templates, each tem- gravity of the input characters, two character pat- plate looking at a small segment of the character." terns could be distinguished with 95% accuracy (Kornberg, 1964, p. 116). after 20 trials, and after 100 trials the model was 5.62a"In Earnest's system...the writer is able to accommodate rotations and systematic dis- constrained towrite more or less horizontally, tortions. For six character-pattern categories, a and he uses a 'lite' pen to write single words on reinforcement procedure was utilized that modified the face of a cathode-ray tube (thus deferring the random connections between the cells as well certain problems related to actual writing on paper). as cell weights, and thatfacilitated a search for The present [Mitre] system, which is set up on the `good' connections. M.I.T. Lincoln Laboratory TX-2 computer, recog- Further investigations...resulted in models nizes about 10,000 common English words. The in which connections are not random, but in which system deals only with whole words, and so cir- the input pattern elements to a single cell represent cumvents, inpart, the segmentation problem." a specific local operation on a character.In these (Lindgren, 1965, p.107). later models various reward-reinforcement func- "Using the estimated envelope as a guide, a tions were studied. Roberts concluded that it is matrix program finds the following key features: possible to recognize the characters of a complete alphabet with an accuracy of 94%, after a training 1. significant strokes above the envelope (as period of 40 trials per character, using these net- in b, d, f, etc.) and whether such strokes have works and a suitable reward function. (See Roberts, a crossbar (i.e., t), 2. significant strokes below the envelope (as 1960, and Stevens, 1961). 5.63a"The recognition of hand-printed numbers in f, g, j, etc.), was attempted with a linear decisionfunction. The 3. significant closed curves e;;tirely within the set of measurements which was used involved envelope (always in e), quantizing the number into a 12 X 12 binary matrix. 4. the number of times the pattern crosses an A matrix element was given a weight of one if it axis midway between the envelope lines.... contained a mark and weight of zero if it contained "The feature detection programs are based on no mark. The quantized number was thenpositioned connectivity within a given region, so that when in the matrix by aligning its center of gravity with crossbars overlap adjacent tall letters (as in little), the center of the matrix." (Highleyman, 1962, p. the overlapped sequence is detected as a single 276). grand feature.... "The implementation consists essentially of a "Once the key features have been identified, transparency-photomultiplier combination for each a 7-bit category code is constructed as follows: character. The transparency represents the average The highest order bit tells whether any crossbars character. When the image of the unknown char- were found, the next three bits tell the number of acter is focussed upon the transparency the trans- high strokes, and the last three tell the number of mitted light, measured by the photomultiplier, is a low strokes.... function of the desired cross-correlation value ... "The next step in the process is a comparison of Consider a piece of film in which the transparency. the predicted number of axis crossings of each at each point is proportional to the probability of word in the chosen category with the observed occurrence of a mark at that point for aparticular values (within a certain tolerance)...Finally, character. That is, the transparency of the film rep- a crude model of the handwriting process steps resents an average character ..." (Highleyman, across the pattern from left to right, checking the 1961, p. 510). X-coordinates of extracted features against corresponding features (if any) in each word remaining "This method, with centering by shifting and a penalty threshold of 0.04, was applied to the hand- on the list.... printed alphanumeric alphabet of 1,800 samples. "Fivesubjects...wererecentlyaskedto write words selected at random from the dictionary. No rejection criteria were applied. The total rec- Out of a total of 107 words, the computer correctly ognition rate was 77.2 per cent." (Highleyman, 1961, p. 507). listed 65.... "It appears that reliable mechanical recognition 5.63b"While the techniques developed were of cursive writing is not out of the question, though applicable to generalized line patterns, the partic- 120 ular subset chosen for detailed test were hand a different picture, as. found in cartoon strips and printedletters...The printing samples used mail order catalogs...and (2) handwritten arabic were collected from 35 individuals." (Spinrad, 1965, letters, written by the same person. The program p. 125). achieved 95% success on known and 70% on un- "The Scan program starts by determining all the known pictures, and 60% success on known and line segments, at eight angular steps, which can be 55% success on unknown arabic letters (segmented fitted into the source pattern. A series of four handwriting) in the fifth pass....Results from `filtering' programs are then employed to succes- two 5-pattern arrays: (digitized and degraded sound sively 'weed out' those line segments which do not spectrograms of speech (the numbers 'zero', 'one', represent a good fit.... `two', 'three', and 'four', as spoken by different "The representational scheme we have selected speakers) (Uhr and Vossler, 1961c), and (2) seg- has a matrix format in which the rows are a descrip- mented lower-case handwriting, written by different tion of the line segments and the columns describe people. The program achieved 100% success on their relationships to the pattern's centroid. [Angular both known and unknown spoken numerals by the displacement from horizontal reference, bearing fourth pass, and 100% success on known hand- direction].... writing by the third pass. It achieved 60% success "A sample of 243 letters, different from those used on the unknown handwriting, but it is likely that it in the Accumulate program, was run through the would have improved further on these inputs if it classification program. The program correctly iden- had been given more opportunity to learn (once it tified 93.5% of the input set.... achieves 100% success on known patterns it does "The Scan line recognition program, with its not benefit appreciably from subsequent learning constituent filters, acts as a very effective 'noise' experiences)." (Uhr and Vossler, 1962, pp. 322, 326). rejection mechanism...Extraneous blots, spots, 5.63d"Thispaper ...[assumes]thatan etc.., are not fitted with line segments because they adequate structural description of a character has are too small or because the program cannot abstract. been extracted from the character's environment just one representative line segment. Further, be- and passed to the information selector. This struc- cause the Scan program initially fits a multiplicity tural description consists of an ordered sequence of segments, the final abstraction is relative!y tol- of (x, y) coordinate values recorded at points around erant of the void areas within the character's bound- the character's boundary ::here its edge changes aries." (Spinrad, 1965, pp. 127-128, 140-141). direction. The underlying philosophy for this partic- 5.63c"The program that has been tested to date ular structural description is the same as that of generates and searches for its operators at what Zahn. To paraphrase Zahn, a sequential trace of a might be termed the first synaptic level, immediately character's boundary yields the most useful informa- behind the input matrix. A program that is now being tion for distinguishing one character from another written first pre-processes inputs by means of this because curves defining the boundaries of letters first synaptic layer. Specifically, it performs a two- and numerals have certain properties which are dimensional differencing operation, which trans- invariant over several fonts and that any specific forms a surface into its contour (e.g., Raslievsky, letter or numeral observed on the basis of its 1948). Operators are then generated, and successful boundary would hardly be confused with any other and optimally adjusted operators searched for, at letter or numeral. Furthermore, it is felt that dis- this level, to work on these contours. The contour crimination based on the trace of boundary edges differencing operation will allow us to begin to ex- will encounter difficulty at about the same level as amine the problems of recognizing patterns of the human observing characters out of context... . graded intensity.... "The paper presents a mathematical model of a "The speeded up program has been tested on proposed information selector. It offers some in- several different types of input patterns....In teresting possibilities because of its simplicity and most cases, results were quite similar on both power. The model introduces the Angle Versus `known' examples (that is, examples the program Length (AVL) representation, a technique that had previously processed and hence had learned traces the boundary of a character from its struc- from) and 'unknown' examples (that is, different tural description to produce a characteristic wave- from the ones used in learning, and also produced by form. It generates this waveform by plotting the different people)...Results for several different tangent angle versus arc length around the figure's sizes of pattern arrays, all of handprinted capital boundary in such fashion to form a closed contour. letters, printed by different people...show rela- The resultant waveform is unique and complete be- tively little decrease in the program's abilities as cause it preserves the relationship of each boundary the array size is increased, at least up to the 26 letter feature with ever) other boundary feature so that alphabet. Thus, on the sixth pass through the 26 the original and coaly the original character may be letter alphabet the program was 100% correct on reconstructed from it. The Fourier series transform known patterns and 96% correct on unknown pat- of the characteristic waveform is performed and terns." (Uhr and Vossler, 1962, pp. 321-322). the resultant coefficients are used as descriptors CC ...Results for two 10-pattern arrays: These hence the term Fourier descriptors.... were: (1) line drawings of cartoon faces and simple "The AVL representation and Fourier descriptors objects (such as shoes and pliers), each copied from were conceived by Cosgriff in 1960 and subsequently 121 refined and developed by Fritzsche, Raudseps, and detected without resorting to precise measurements Borel. The research reported here is an extension of location and direction. In some instances, such of previous work by the above researchers... . as certain 8's and O's, identity cannotbe established "The recognition scheme was tested with two data by coarse measurements on the outer contour of the samples. One was composed of more than 600 ma- character. For these cases, the scanner is redirected chine-printed characters representing all the nu- to obtain more information from the interiorof the meral classes of 50 different commercial fonts and character...Since accurate recognition can be the other contained over 400 hand-printed numerals accomplished without resorting to precise measure- chosen from 40 styles of handwriting. A reference' ments of line locations or directions, this system set of 21 Fourier descriptor representations was provides large tolerance for shape variations.... used with the machine-printed numerals and a "The reading capabilities of this model were de- different reference set of 17 representations was termined in an experimental study at Tufts Uni- used for the hand-printed numerals. The recogni- versity...Approximately 200 untrained subjects tion technique identified 98.6 percent of the ma- wrote 20,000 numerals, of which 92% werecorrectly chine-printed numerals and 90.5 percent of the recognized." (Greanias et al., 1963, pp. 14-16, 19). numerals from the hand-printed sample." (Brill, 5.64 "A deterministic pattern recognition sys- 1968, pp. 1-2, 7). tem for hand and prilited numbers andcapital block 5.63e"The curve tracing technique is one of the letters was developed and tested by F. Kuhl. Here more recent concepts. Its basicadvantage of being an angularly quantized peripheralcode was used to able to handle relatively wide variations in the approximate contours of character images. The shapes or sizes of the characters makes it a useful transformationiscalled angular mapping trans- method for recognizing hand-printed characters. formation (AMT) and together with redundancy- By following the character outline, the recognition erasing reduced the patterns to a few numbers. The 'unit determines certain features which it uses to condensed numerical description is used in a one- identify the character. Among these features are dimensional correlation scheme. It is insensitive to character splits, line intersections, line magnitudes, major variations in the particular representations and the straightness of lines. The major disad- of a character but still can distinguish between the vantage is difficulty in readingcharacters having characters of interest. broken lines or holes within their block boundaries." "AMT is similar to the edge sequences proposed (Feidelman and Katz, 1967, p. 36). earlier by Unger, except that here each vector en- 5.63f An earlier report on these IBM contour- following techniques provides such details as the codes in incremental change in position, length and following: "The characters are scanned optically directional angle of the trace. along a variable path that depends upon the informa- "Of some 385 test characters drawn in pencil. tion obtained as scanning proceeds. The actual over 95% were recognized correctly.This trans- character on the document serves as the primary formation, likeall contour following recognition store of raw character information. The scanner is sci,emes, is quite sensitive to flourishes and curl- the means for retrieving the information as required icues." (Stein, 1965, misc. notes). from this 'memory'. Specific regions of the character 5.6a "Other parameters which are useful in are scanned selectively(and repetitively, when important special cases, such as that of hand printed necessary), ac,;ording to their significance in char- or written characters, includenumbers of inter- acter identification. The scanner isalso made to sections with a raster, numbers of branch points of seek significant features in new areas of the char various orders, and so on." (Rosenfeld, 1962, p. 114). acter when prior scanning does not provide con- 5.65It is to be noted that "considerable tol- clusive results.... "To accomplish recognition, the contours of erance was provided for: each character are traced two or more times. Character registration. If necessary, characters During the first tracing, the extreme positive and can be located anywhere within an area several negative excursions of the character's outer con- inches square. tour in x and y are stored. On subsequent passes Character size. The height of characters can around the character, the beam is located relative vary over a 4-to-1 range. to the character extremities and therefore is not Shape. The tolerance for shape is defined in dependent on the character location or the point terms of the shapes that occur in unconstrained of initial contact.... handwritten numerals. The initial recognition "The analysis of the filtered beam deflection sig- criteria were selected to cope with more than nals is based on two properties of the line edges 90% of the character shapes found in 3000 that they represent: samples of unconstrained numbers. Line quality. Lines from medium hard pencils (a) The relative location of the line edge within and dark-ink ball point pens in reasonably good the character. condition are acceptable. (b) The approximate direction of the line edge. Character rotation or slant. Normal characters "By considering both properties simultaneously, can be rotated J..- 200." (Greanias et al., 1963, the significance of most line edge signals can be p. 15). 122 5.65a "My invention requires no special nor templates representing the features to be detected. complex scanning. I use any successive line pro- For optimum performance the templates should be ducing scanner, e.g., the scanner shown in the J. matched with the unknown in the correct orienta- Rabinow Patent No. 2,933,246 or a row of photocells tion and position, and the templates should be of he or others. This is possible because my invention is correct size and shape. The basic principle under- only in the nature of a curve tracer. lying detection by auto-correlation is the use of the "To the best of my knowledge, my invention en- features actually present in the input pattern as tails a row system of character identification based their own templates thus making h unnecessary to upon a new philosophy. Specifically, if a scan line position and orientate the pattern." (Clowes, 1962, crosses a character line, my invention remembers in p. 306). time, or voltage, or space, or digits, etc., the crossing For criterial feature extraction by Bomba's local point, and if the next scan line (of the same char- operations, the area operated upon is a radial pattern acter) crosses the character within a given area built up of the appropriate combinations of smaller (of time, voltage value, space, etc.), it is concluded local areas. For example, to find an L-shaped fea- that the character line being investigated is con- ture, a pattern element detector consisting of a cell, tinuous and this information is remembered... P, the seven cells directly above P in a vertical line, "In one embodiment of the invention, when a and the seven cells extending in the horizontal line vertical scan line first crosses a character line the to the right of P, is moved (Ayer the input pattern crossing point is remembered and a gate system is field. This is done in a scanning manner to detect set during a portion of the next (second) scan line. coincidence of black cells in the input pattern local The duration or width of the gate is made to include area with the cells of the L-featureextraction pat- a time, area, etc., slightly above and slightly be- tern. When there is coincidence for all the desig- low... ahorizontal projection of the first remem- nated cells, then an L-feature signal is recorded in a bered crossing point, enabling the said next scan buffer image for this feature at the same respective line to interrogate the character in a restricted area coordinates which the cell 'P' then has. By dividing alongside of the first crossing. Now, if the second these secondary input patterns (buffer images for scan line crosses the character line within the inter- each extracted feature) into zones, the recognition rogated sensitive area, the scanner output falls logic may take account of relative location and con- within the width of the gate system and is remem- nectivity of the features that have been detected. bered; and the same way, but for the next scan line (See Bomba, 1959, and Stevens, 1961, p. 94). and this process continues to the end of the char- 5.66a "SPAC has been suggested for use in acter line. . recognizing handwritten alpha-numeric characters "By examining the behavior of the follower, a (perhaps for inputting handwritten programs to a characteristic trace of the line of the character being digital computer), recognizing diagrams of molec- examined, can be developed and remembered. For ular structures in patent searches and recognizing example, if I assume a horizontally moving character biological cell structures. It could also be used as area and a vertical scan made of a scan element an aid in analyzing photographs of particle tracks moving from bottom to top of the character, my in- as required in higi. energy physics." (Fuller, 1963, vention' easily recognizes whether the characteristic p. 24). trace is sloped up or down, curved or straight by 5.67"To demonstrate the effectiveness of this observing whether the hits are early or late in the design procedure utilizing time-varying threshold width of the gate system.... logic units, the technique was used in the simulated "More than one follower is required to distinguish design of a character recognition device. The input characters from each other...For example, con- patterns were a set of 240 hand-printed alphabetic sider the letter X. The first follower is assigned and characters, divided into 12 classes with 20 char- begins to function when a scan line crosses the lower acters hi each class..." (Astropower Lab., 1964, left corner and there will be a wide space in the p. 194). same scan line, and then another crossing. The 5.68"M. Fischler has reported on experiments second crossing assigns a second follower to operate on several thousand characters, hand-printed by simultaneously with the first but it investigates a approximately 100 people. The decision-making different portion of the character. Now consider the system was a special purpose computer (PIP) letter F. The first crossing is long and the scanner which simulates a network of threshold elements will produce a long pulse. Means are provided for designed to identify a pre-selected set of typical identifying such a pulse. The next few scan lines patterns. The characters were constrained within will produce a hit in the first follower gate system rectangular boxes. corresponding to the first horizontal leg of the F, "A flying spot scanner with quantized output and a second follower will be propagated along the was used. Feature words made up of 350 bits were upper horizontal bar of the F." (Holt, 1964, p. 1). derived to describe the 'upper black', 'right diag- 5.66"The programme described by Bomba onal', 'solid black', etc. Of the identifications at- .employed a cross-correlation technique to extract tempted, some 71% were successful." (Stein, 1965, features from a character. The method may be vis- misc. notes). ualised as the comparison of the two-dimensional 5.69"P. M. Lewis conducted a search for 'good' input pattern with a set of standard patterns or statistics relating patterns to relevant character-

366-107 0 - 70 - 9 123 istics. He chose fifteen alphabets of 62 characters two-dimensional line patterns into one-dimensional each. From these he evolved thirteen characteristics statements. and his resulting system recognized 81.9% of letters. "The most detailed one-dimensional statement The fifteen alphabets included nine machine-printed contains a record of the slope. The most condensed and six hand-printed. One sample of each symbol is a quantized angular description of thepattern. from each alphabet was included. Uppercase, lower The condensed patterns are then compared witha case and numerals were represented in the 920 list in memory to determinean initial classification. samples which also varied in height bya factor of One-third of the characterscan be identified this four. Lewis suggested thatmore samples would have way. The remaining are divided into sub-classes to be looked at in order to derive a more effective containing two to six members. More detailedpat- code." (Stein, 1965, misc. notes). tern statements or additional tests can be applied to 5.70"Defining the vector spaces and deciding extract the proper identity from the sub-class. The whether an unknown branch isan element thereof patterns covered do not cover cases where loops constitutes the main part [of Minneman's 1966 occur. Reasonably well-printed characters were paper on handwritten character recognition]." In used for the study. Constraintswere placed on (1) order to define vector spacesone examines several flourishes and (2) gap size. copies of the same character... asdrawn by "This work was simulatedon a 7090. A curve different people, for 'learning' the referencecon- follower constructed to input thenecessary infor- figuration thereof. The mean of the resultingvector mation would have to be backedup by considerable space is used for point-by-point comparison with an logicto avoid retracing previously encountered unknown. The criterion of acceptance...is based curve segments at crossovers." (Stein, 1965, misc. on a conditional probability and distance relation- notes). ship." (R. Jessup, review, Computing Reviews 8, 5.72"Fifteen modular criteriawere developed 113-114 (1967).) by G.U. Uyehara for reading numerals eithertype- "Samples were obtained of the characters 1, 2, written in multifont or hand printed. Five criteria ...9, 0 from 26 individuals. Theywere instructed were used to describe the inter-connections and ten to letter neatly, but in their own style, avoiding described shape and tread of individual'streams.' breaks (except at top of 8), and to try to make their Uyehara uses a modular criteria detector which characters fit within a l -inch diameter circle. Closed analyzes the character during scanning. The results top 4's were not allowed.. . of this are then inserted intoa buffer store and "The 260 design sampleswere then tested against from there they are organized intoa pattern lattice the 35 standards derived from themto determine network. Two kinds of recogr_ilionare used. The the internal consistency of thesystem. The results Gestalt criteria are used to recognize beginning, were as follows. end, continuation,convergence or divergence of Correctly recognized 253=97.3 percent streams. The single stream criteriaare listed as: Not recognized 2= 0.8 percent first quadrant compound wedge; second, third, Incorrectly recognized 5= 1.9 percent fourth quadrant wedges; first/fourth. quadrantcom- pound wedge; second/third quadrant compound "The number of subclasses necessaryto define wedge; positive diagonal line, negative diagonal all variants of a character typewas found to range line, horizontal linesegment; and vertical line from one for a one or azero to eight for a four. On segment. average, 4.3 variants were found per character "Uyehara reports that the Gestalt criteriaare type." (Minneman, 1966,p. 93). sufficient if the character can be identified from the "To test the performance of thesystem an addi- inner connections alone. Otherwise, the single tional sample of 260 characterswas obtained ... stream criteria as organized in the pattern lattice The results of the testwere as follows: network are necessary to supplement the features. Synchronism between thescanner disc and delay Correctly recognized 254=98.5 percent a Not recognized 0= 0.0 percent lines, and the accuracy of construction of the slits: Incorrectly recognized 4= 1.5 percent in the disc were among the worst engineering problems encountered in the construction of the first [2 characters not processed because ofprogram model of his equipment." (Stein, 1965, misc. notes). deficiency]. (Minneman, 1966,pp. 93-94). 5.73Weeks, 1960. "In these experiments, both 5.71"S. Spilerman has revealed a classification machine printed and handwritten numeric char- scheme for the alpha-numeric set which is insensi- actersservedas source patterns which were tive to common variations in hand-printing. Thi's scanned: 'In television fashion witha raster con- method is based on invariance in angular change sisting of six lines uniformly distributedover the along certain segments ofa line pattern. digit at six different angles uniformly spaced 30 "The system groupsmany variants of a pattern degrees apart.'...There is then processing which together and yet breaks the characterset into a determines the first scan line crossing and the last large number of subsets, each containinga small and then divides the interveningarea ...to obtain number of characters. A method for quantizing and 6 equally spaced areasacross the character image. representing simply connected segments ofa pat- The number of rector crossingsor intersections per tern is first introduced. It is then used to convert scan line are now counted, since different portions 124 of the character may be crossed by thesame scan is usually in on.t piece, but it is occasionally in two line. The counts are compared with probability or more pieces. It is hard to imagine a capital M that tables, which are based on statistics gathered does not have a locus 2012. On the other hand, it is from previously scanned and processed characters hard to imagine an A, B or C that has a locus 2012. and thereby provide the identification formulas." In fact, the locus 2012 is rather characteristic of (Stevens, 1961, p. 50-51.) an M M. .. 5.74 "A model of a pattern recognition com- "To form the feature vector of a pattern,we list puter, whose design is based on electronic simula- the codes in a ainsistent order and, after each code, tion of a nerve cell, has been developed by two write the area of the locus of that code as it appears electrical engineers at Purdue Univ. Thus far, the in the pattern. If the locus does not appear, we write small-scale model, built by Prof. King-Sun Fu, and zero: otherwise we write the relative area of the Wen C. Lin, has recognized hand-printed Roman locus, that is, its percentage of thesum of the areas and Chinese letters. of all the loci. Slice every pattern has each locus "The system consists of a photo-cell grid, which to a greater or lesser degree, even though this may scans and converts the black and white patterns be zero, the vecwr thus formed is somewhat char- into digital signals, joined to a box enclosing 200 acteristic of the pattern... interconnected circuits, or 'neurons', 'andan I/O "The scheme can be implemented bya special- panel board... purpose device atlached to a digital computer. The "The circuits are arranged into three banks: (1) computer can, of course, be the very one that will the transformation layer, which receives andre- use the printed information that is read by the duces the pattern from 144 to 24 bits; (2) thecor- system. Thc; basic cell of a grid of identical cells relation layer, which compares the 24 bits with the has been designed io that, with a pattern impressed pattern references stored in memory, feeds the cor- on the grid of cells, eacl. background cell can com- relation coefficients back to the first layer viaa pute its own four-digit ternary code. The computer learning loop to inform it how to adjust its threshold then communicates with the cell grid to determine for succeeding patterns and fires thesame coeffi- the pattern's vector. With the decision functions cients to (3) the decision layer, which indicates to derived in an earlier training mode stored, thecom- which class the pattern belongs. A second learning puter then classifies the pattern." (Glucksman, loop may correct the decision layer." (Datamation 1968, pp. 13-14). 11, No. 7, 71 (1965).) 5.74b"In exploring the use of nth-order auto- 5.74a "We shall now describea technique, at correlation functions for pattern recognition, we present under investigation at AFCRL's Data Sci- conducted experiments with typewritten samples of ences Laboratory where it was originated by the the characters a, e, s and handlettered samples of author, for reducing alphabetic patternsto feature numerics.... vectors. A printed letter is composed of line and "This [handlettered numerics] was run witha set background. In the case of a letter printedon a white of 3100 characters (distinct from the set used for page, the inked part is the line and the white part finding the subspaces W1), and 38 classification the background. In the followingmanner we give errors were made." (McLaughlin and Raviv, 1968, any background point a four-digit ternary code, pp. 131, 137). where ternary means that each digit is 0, 1,or 2. 5.74c"The samples used in the experiment were We draw imaginary horizontal and vertical lines obtained by asking three subjects to print characters through the background point, forminga cross. The in a two inch square or to write characters in a circle cross has four arms that extend left, up, down and with a two inch diameter. The charactersare fairly right. Each of these arms determinesone of the four centralized and normalized. Eight features... code digits for the point. The digit is 0 if thearm were selected for handprintecl characters, and does not intersect the line of the pattern. It is 1 if eighteen features...for handwritten characters the arm intersects the pattern justonce. It is 2 if the ..Each feature is assigned by a number which is arm intersects two times or more than two. In forming the distance, measured along the predetermined the code, we write the digits from left to right, taking path, from the edge of the squareor the circle to first the digit determined by the leftarm, then the the first intersection of the character." (Chen, 1966, upward arm, then the downwardarm, and lastly p. 557). the right arm... 5.74d"Thefirstprogram, or 'preprocessor' "Now consider the locus of all background points converts any pattern of black-ant!.-white cells on a that...havethecode 2012...Thislocus 20X 20 matrix into a one-dimensLnal description of consists of the lower part of the interior of the its contour.. . triangular upper cavity of the M. Theupper part "Of these classes, the one most extensively used of this cavity is the locus of points whose code is was a set of 27 rather badly distorted examples of 1011. Clearly, every point of the background ofa the hand-printed alphabetic characters D, E, F, pattern must have a ternary code, according to this and G. On this set, versions of theprogram fre- scheme; and clearly, all points havingone code, for quently gave about 75 per cent correct answers example 2012, combine to forma locus. We may after about three learning experiences with each call this locus the locus 2012. The locus ofone code example." (Prather and Uhr, 1964, p. D2.2-1, 5). 125 5.74e"Samples of handwritten numerals were hundred first name signatures. Twenty exemplars collected from a group of salesclerks...These were collected for each of... tensignatures... samples were scanned by a contour analysis Individual signatures were written inside a 5-inch scanner and recorded on digital magnetic tape. The by12-inchrectangle.Although subjectswere measurements provided by the contour analysis instructed to fill the rectangle as much as possible, were all binary and are therefore directly suitable the instruction was difficult to follow, and few fur the present simulation program.... signatures touched the edges of the rectangle. "Two sets of 3,000 samples were used in this Each exemplar... wasprojected against a grid experiment. One set was used to design the recogni- with 20 rows and 48 columens. The image was tion network and the other set to test the network. expanded when necessary to reach three sides of Twenty-four measurements were used to char- the grid, giving preference to the top and left acterize each sample. sides. No other attempt was made to normalize the image.... "Interating on design data, a tree structure... was obtained, and the number of recognition errors "Exemplars of names collected in the second was 42. The error count of the linear structure for session were treated in a different fashion. No the same design data was 84. The computing time attempt at all was made to normalize the figures. on an IBM 7094 was about one hour. Each signature was traced directly on a graphic "Recognition of the test data set was then at- input device developed at the RAND Corporation and attached to the RAND jOHNNIAC computer.... tempted with the tree structure. There were 59 "The ten signatures collected from any subject recognition errors. With the linear network, there in a given session tended to resemble one another were 84 errors in the test data." (Chow and Liu, fairly closely, but there was a considerable difference 1966, p. 79). between sessions. The difference occurred in part 5.75"Research on multifont,handwritten because of changes in the general appearance of the character and cursive word recognition and signa- signatures and in part because of the different ture identification has been done by simulation on methods of processing.... acomputer Telefunken TR 4."("Technisehe "Each signature to be recognized or learned was Hochschule Karlsruhe, Annual Report 1966". 1967, describedwithinthe computer program by a p. 52). property vector with 960 components, where each 5.76"Considerableinformation,frominitial property reflected presence or absence of part of studies performed, appeared to be present in the the figure in a fixed portion of the visual field." dynamic sensing of two signals: pen-tip acceleration (Marzocco, 1965, pp. 210-211, 214). and pen-paper contact. Successive signatures from ...The probability is less than 0.02 that four a subject appeared to be well correlated visually, or more responses will be guessed correctly if no using acceleration and contact signals obtained from systematic effects are present. Since this level of strip recordings.... performance is reached by the time 2 or 3 signatures "The instrumentation of the pen to provide tip- sets are presented under any condition of adminis- acceleration and paper-contact signals was con- tration, it is apparent that learning occurred under sidered to be a step forward. Neither of these vari- every condition.... ables is normally available to the handwriting expert; "A set of properties useful for recognizing the thus, the speed of writing the invaluable time handwritten first name signatures results from the function is lost to the expert when confrOnted a digital encoding without need for further processing posteriori with a specimen.... ". toisolatespecifictopologicalfeaturesof the ..During the course of this study, persons patterns." (Marzocco, 1965, pp. 215-216). found themselves able to identify the signers from 5.77In a later patent issued to Harmon in 1964, the acceleration time plots by looking at the pattern the stylus is me red across a writing platen whose of the acceleration peaks...A plot might be made surface consists of electrically conducting strips of the height of each neak versus its position when arranged horizontally. Features including loops, ranked by size...There would be complete infor- ascenders, and descenders are extractedas the mation as to the peak pattern in the shape of this stylus is moved. curve and in the information which represents the 5.77a"Tests are made for a number of 'local ranking of the peaks. Thus, if the eighth peak were features' such as differential vertical extent,cusps, largest, the fourteenth peak were second largest, the closure and retrograde strokes. Other tests establish sixth peak were third largest and so on, then the individual letter domains, thus grouping the detected numbers 8, 14, 6,...would also identify the features into clusters.... signature.... "The accuracy of letter identification is greater "Using a...vector, whose components consist than 90 percent over a sizable population of writers of such indices as zero-crossing counts, number of and test sentences. Those errors due to segmenta- pen-paper contacts, etc., this comparison system tion inaccuracies amount to less than 3 percent. correctly classified 226 signatures out of...250." Most of the othererrors are due to imprecise (Mauceri, 1965, pp. 8, 9, 50, 54-55.) `feature' extraction.... 5.76a"Patterns whose names were to be learned "Recognition accuracy thus approaches 93per- by the stimulus-sampling program consisted of two cent. Studies are currently being made to use 126 context at a more complex level and to investigate function segments that may be classifiedinto the extent to which the constraints on the writer categories based on topological similarities, thereby may be relaxed." (Harmon, 1962, pp. 151-152), yielding a representation for every cursive letter 5.78". .An attemptis made toidentify by means of a small number of permissible alterna- script on a letter-by-letter basis. We seek a minimum tive sequences of stroke categories. Such categories set of 'local features' which are common to a large are defined by means of statistical averages of the selection of reasonably legible handwriting samples. numerical representations of the member functions, Characteristics such as differential vertical extent, and the likelihood of membership of new function cusp, retrograde strokes and closure are used. segments in each particular category is estimated by Recognition decisions are based on the outcomes of means of the multidimensional distance between the tests for unique combinations of these local fea- representation of the new function and the average tures." (Frishkopf and Harmon, 1961, p. 301.) for the category... "When the same writer is the source of both "Constraints are recognized to exist on two levels, dictionary and test word',, correct identification is those between stroke categories, sequences of which achieved for 32 per cent of all test words written by must form valid letter representations, and those allsubjects...Whencomparisonismade between letters, sequences of which must form between test words written by one writer and a words within the vocabulary of the system. The dictionary written by another, correct identification result of the recognition process is that word which is achieved in 26 per cent of all cases." (Frishkopf is generated from the stroke category sequence, and Harmon, 1961, p. 314.) the constituent strokes of which correspond to a "Method 1 was tested by processing 19 sentences maximum total likelihood. Methods are presented (412 letters total) written by five persons. The correct which implement the stroke sequence to word letter-identification for the entire group was 58.9 mapping in an efficient mat.ner." (Mermelstein and per cent." (Frishkopf and Harmon, 1961, p. 313). Eden, 1964, p. 334). 5.78a"The parameters used are here called the 5.81". .Thedevelopmentofalimited- `loop configurations' of numbers. In the 'loop space', capabi!ityhandwriting recognitionsystemthat Arabic numbers are considered to be made up of can handle a very large font without significant size lines forming fully-closed loops or loops which are or orientation constraints. The remaining constraints open on one, two, or three sides." (Kamentsky, 1961, of language style and segmentation are now being p. 492). removed.... 5.79"Investigations onmultifont and hand- "The user can teach the program to recognize written character recognition are under way. A his set of characters. The learning process for the character recognition scheme forhand-printed program involves modifying individual decision numerals has been realized. For the simulation of trees, changing the weights on each tree, and where handwritten character recognitiOn a special light-pen necessary, introducing new decision trees with off -line input device has been developed. The use of their corresponding properties into the system. shape elements .as features has been studied and Because the program was written as an input linear classifying networks have been considered. device to a larger man-machine system, the descrip- An integrated computer-scanning-classifying system tion of the implementation stresses the human with- feedback control is being developed." (Tech- engineering features. A qualitative evaluation of the nischeHochschule,Karlsruhe,"Researchon system as implemented is offered, together with Pattern Recognition...", 1966, abstract, p, 1). possibilities for expanding and genera'izing the "Methods of contour description and line seg- program." (Teitelman, 1964, p. 559). mentation of characters have been investigated with "Essentially the only things that can be said regard to storage capacity requirements, operation about a program of this type relate to its convenience time and redundancy. Other methods are based and usefulness. The program has been taught to mainly onsequentialdecisiontechniques." recognize, on separate occasions, Russian char- (TechnischeHochschule,Karlsruhe,"Annual acters,Hebrewcharacters,Greekcharacters, Report 1967", 1967, p. 63). upper- and lower-case English characters, and a 5.79a"Handwriting analysis was begun and very large collection of mathematical symbols. It has well characterized by Eden and Halle. They found been used by many different people, and the combi- that only 18 strokes are used to construct all of the nation of control panel and punch-out and read-in English characters. These strokes can be generated features make itat the very least enjoyable to and joined together by simple rules in a computer operate. With a little experience, after the user so as to 'write' any word. By finding the ranges of becomes accustomed to the idiosyncracies of the the stroke parameters for an individual, the com- light pen, and trains the program in the variations puter can forge his handwriting." (Roberts, 1962, p. of his handwriting (surprisingly many people are not 210). aware of the fact that they vary their style of hand- 5.80"An algorithm for the segmentation of the writing from sample to sample), the resulting man- pendisplacementsignalofhandwritinginto machine system becomes quite effective. Again no function segments corresponding to strokes has quantitative results can be offered here other than been presented. This procedure, when repeatedly observations on use of the program by a considerable applied to a collection of handwriting samples yields number of individuals.... 127 "The results seem to indicate that the use of the "The recognition performance on the test sample time sequence is a powerful tool in pattern recogni- of one individual, given a set of weights adapted on tion.It enables a program using fairly simple the analysis sample of that individual, showed that properties to achieve a high rate of recognition. both individuals obtained a recognition performance One way to improve the program might be to include in excess of 99.4 percent." (Simek and Tunis, 1967, more sophisticated pmperties other than the posi- pp. 75, 7-80). tion of the pen. An immediate improvement would "In another study, samples of handwriting were be to include a property that detected sharp corners fed into the system, employing intentionally 'sloppy' by noticing changes in the velocity of the pen. characters to test recognition capability. The set Other properties might note curvature or angle. of weights used was based on a sample of 200 Developing a language that described a wide class alphabets [numerics] of printer one....[approxi- of properties would generalize the program even mately 26%] were not identified correctly. Many further. The user would then be able to communi- of the characters identified incorrectly are notice- cate subtle differences in forms by means of fairly ably smaller; they were undoubtedly not recognized abstract concepts." (Teitelman, 1964, p. 574). because of system resolution.... 5.81a"Although this particular work emphasizes "The handprinting input device and recognition handprintinginput,thetransducerisclearly system described has been shown to be technically applicable to allowing the input of more general feasible. Its use awaio. its economic justification in information such as mathematical symbols, draw- particular applications.... ings, or graphs.... "As might be expected, an increase in alphabet "The output of the printing transducer is trans- size results in a decrease in recognition perform- mitted into an IBM 1620 data-processing system ance. However, with., the largest alphabet size used ...Thedecisionprocedureusesadaptively [26), the recognition rate is still in excess of 92 per- derived linear boundaries...The use here of cent, a level which should be acceptable in many the adaptive system allows the recognition to be applications if some method of feedback, allowing trained for a variety of individual printers and for retransmission or correction, is included.... a variety of printing fonts.... "The performance level achieved by one indi- "The measurement space used describes the vidual, using the weights generated by the analysis character in terms of Straight-line segments... sample of a different individual (but both using the Eight types of segments are allowed, corresponding same font), is in the vicinity of 95 percent." (Simek "to the eight directions of the compass...The and Tunis, 1967, pp. 80-81). printing area is divided into nine zones, or regions, 5.81b"Studies were made into the usefulness of and the region in which a particular segment begins various recognition schemes for handprinted alpha- determines where in the measurement space its numer characters and ease to which subjects occurrence is recorded. The number of times the adapted to program- forced constraints. The Sylvania printing stylus is lifted off the transducer during Data Tablet proved extremely useful in this appli- the creation of a particular character is also recorded cation for several reasons: First, the inking capability ... asis the number of times a particular segment allowed the subjects to write on a sheet of blocked occurs in a particular zone." (Simek and Tunis, paper in a very natural manner. Second, the multiple 1967, pp. 72-74). threshold Z axis feature allowed the programmer to "The circuitry tests the sequence of x, y location use one level (contact with the surface) to indicate pointsfor incremental changes. When apre- stroke and a second level (about 1/8 inch above the dominance of these changes is established in one of surface) to indicate completion of a character. In the eight possible directions, a valid measurement other words, multiple stroke characters are formed has occurred. An indication of the segment is with the tip of the pen always below the 1/8 inch transmitted to the computer; subsequent segments level, and once the pen is raised above that level are recorded only if they differ from the the segment the completion of a character is indicated to the immediately preceding them. For printed alphanu- computer.Third,thehigh resolution provided meric characters, an average of seven such measure- minutedetails needed for recognition of small ments is obtained per character; for the charactercharacters. The 1 per cent accuracy of the tablet `8' a maximum of 14 such measurements occur. Note proved more than adequate for accommodating that we are breaking up curved sections of a char- letter sizes of 3 inch or more." (Teixeira and Sallen, acter into a sequence of straight-line segments.... 1968, pp. 320-321). "The experimental results are based on fairly large samples (approximately 300 alphabets, where 5.82"The word parapropagation is intended to an 'alphabet' is one example each of the characters signify the interruption of a propagation by an interposed barrier. It will be shown that patterns can be to be treated) obtained from two individuals...The two printers were not trained and used their normal classified by a process based on the interruptions of printing style.... propagations." (Glucksman, 1965, p. 435). "The performance on the test sample is depend- 5.82a"Very simple, automatic methods are ent on the size of the analysis sample. In effect, the sufficient to decompose a printed page into text analysis sample must statistically represent the set in one, or a few, predetermined type styles, and test sample.... other material. While there appears to be no con- 128 venient standard for measuring the accuracy of the wherein whole words at a time are recognized by decomposition, it would seem that more compli- an indexing process. The index terms will be cated and time-consuming methods of spectral characteristicssuch as energy distributionat analysis need not be invoked for this purpose. various frequencies for a succession of phonemes. "With documents of the order of complexity of Simple language-structure rules will make it possible technical journals, automated scanning at this level, todistinguish between words which sound the coupled with existing character recognition devices, same but are spelled differently." (Astrahan, 1958, would not produce a computer storable reproduction p. 313). suitable for all of the intended purposes of the 5.82d"Still another area of experiment that has original document. A limited range of functions, grown more germane to the problem of automatic such as automatic indexing, and perhaps extraction, speech recognition are the psycholinguistic studies, could be performed on the coded version. Simpler which have begun to give new insights into how page structures, such as the 'claims' section of humans process speech." (Lindgren, 1965, p. 46). patent documents, could be processed for natural 5.83"Afuturisticsystem must cope with language search procedures and completere- problems such as dialect variations, slurring of display." (Nagy, 1968, p. 486). words in continuous speech, and the resolution of "The IBM experimental character recognition homophones." (Bhimani et al., 1966, p. 275). system on which these experiments were performed "Another difficultyarises from the different already incorporates a number of fairly sophisti- dialects in common use. Here the same words or cated features for registration, normalization, noise phrases spoken by differenttalkerswill have suppression, threshold adjustment, and character differentphoneticcontent. Thus transliteration separation. When presented with a noncharacter from a sequence of phonetic elements to English field, it attempts to convert what it finds into familiar words may involve complex linguistic structure." symbols by means of protracted and agonizing (David and Selfridge, 1962. p. 1098). convulsions. What is needed, then,is a rapid 5.83a"Not only were there variations between `prescan' of the entire page, or a sizable fraction speakers in their acoustic outputs, but therewere thereof, to indicate to the control unit which areas variations by- the same speaker in different circum- are to be read, and which ones skipped or copied stances, in differing emotional states." (Lindgren, onto magnetic tape without further processing. 1965, 121). The logic used to derive this information from the 5.84"Visualexaminationofthequantized prescan is discussed in the third section of this spectrograms on the display matrix revealed that report." (Nagy, 1968, p. 481). the location of the lowest frequency format using "The percentage of correctly identified letters this technique is much more erratic with female in the words underlined varied from 95-98 percent voices. This result demonstrates the fundamental from day to day depending on the care taken to weakness of the short-term Fourier Analysis in adjust the video circuits of the scanner. This locating the poles of the vocal tract." (King and includes only lowercase characters, since there Tunis, 1966, p. 73). was no recognition software available for capitals, "The same words spoken by a man anda woman punctuation, and ligatures (fi,ff,fl, HI, and ffi)." differ drastically in their acoustic content, but the (Nagy, 1968, p. 483). listener has little difficulty in establishing that they "The method to be discussed stems from the are the same words." (David, 1958, p. 294). Simple observation that character fieldsare readily "It is not merely a question of the differences in distinguishable from almost everything else by 'the speech of a man as compared with a womanor a average density of the lines and of the blank spaces child, but also of the variations in dialecticalpro- above and below the lines." (Nagy, 1968,p. 483). nunciation, with the added factor of emotional 5.82b"Because handwriting presents a moreor colouring." (Kusch, 1965, p. 201). less continuous signal, it too raises the specter of 5.85"R. H. Galt in 1951 decided to draw upa the segmentation problem." (Lindgren, 1965,p. 105). set of rules for recognition of ten spoken numbers 5.82c"Automatic recognition of speech would from spectrograms." (David, 1958, p. 301). be particularly u3eful for information, communica- "L. G. Kersta in 1947...assumed a selected tion and control engineering. With its help it would vocabulary consisting of ten spoken numbers, and even be possible to control machines by means of took spectrograms of them from each of ninemen the spoken word, to carry out switching operations and five women. He divided each spectrogram into of all kinds, to dial telephone numbers, to write a mosaic of square elements. If the integrated texts on typewriters, to feed data into computers density in a particular element were 1/2 or greater and to simplify the transmission of messages." than the integrated density then it was represented (Kusch, 1965, p. 201). as being entirely black. [otherwise, white]...For each digit he then compiled two master patterns "Voice recognition for purposes of translation, ...all black elements common to the utterances control, or recording for printing will eventually be of the fourteen speakers....The second master handledwiththeaidof large memories... pattern consisted of all common white elements. Schemes for building up words by recognizing indi- ...Now if the black and white mastersare applied vidual phonemes will be replaced by systems in tandem with a logical 'and' requirement; that is, 129 both black and white elements must match, ... them...Within the limitations imposed by the the maximum error probability for any digit is less number of points within the clusters of the sample than one percent, while the average over all digits space, this experiment resulted in the satisfactory is about 0.2%." (David, 1958, pp. 305-306). separation of eight unknown clusters of complex Lindgren cites other early work at the Bell patterns in an arbitrary sample space, identified Laboratories as follows: aberrant samples, and suggested an output code." "In the Wiren-Stubbs electronic implementation, (Dammann, 1966, pp. 80, 87-88). the properties separated were the voiced sounds 5.85c"Adevelopmentalspeechrecognizer from the unvoiced, the turbulent (noise like) from demonstrated by Philco-Ford Corporation is ex- the nonturbulent; then the nonturbulent sounds pected to be the forerunner of practical and econom- were separated ...into groups ..and the un- ical voice control equipment for applications such voiced turbulent sounds were separated into the as: direct voice input to a computer, phone dialing stops and fricatives ...Fairly good results were by voice, voice control of machinery, and many obtained with this system. For instance, for vowels other specialized applications both military and in short words pronounced by 21 speakers, accuracy non-military in nature. The compact device re- was above 94 per cent .." (Lindgren, 1965, spondstowordsfromalimitedvocabulary p. 118). (presently OH through NINE) with typical acu- "Another version of the Bell Labs system (called racies between 90 and 99 per cent depending upon Audrey) was developed later in the decade (1958) by the speaker, his microphone technique, and his

-Dudley and Balashek. It would recognize acoustic familiarity with the recognizer.... patterns corresponding to 16 different basic linguistic "Input to the word recognition logic comprises elements...The incoming acoustic signals were three measurements that are made continuously broken down into specific patterns, which were on the speech waveform. The resulting unique set compared with patterns stored in the machine. of parameters are based upon a transformation of Best-matches were determined by cross-correlation the informationcarrying resonances(formants) .methods." (Lindgren, 1965, p. 118). of the human voice into a single equivalent fre- 5.85aBerkeley describes briefly a project at quency representation. The three measurements MIT under the direction of William N. Locke are (1) the frequency of this single equivalence supported by the National Science Foundation to resonance, (2) the amplitude, and (3) the state of construct a machine that will recognize spoken voicing, i.e., whether the sound is originated by the sounds and write them down- as English phonemes. vocal chords, or by noise due to air flow over a "A prototype machine distinguishes successfullyconstriction in the vocal tract (as when speaking between vowels and consonants and also between the 's' sound). These three parameters are sub- the consonants sounds `SH', and Efforts sequently quantized to provide digital inputs to are currently being made todistinguish electroni- the word recognition logic." (Killet, 1967, p. 3). cally between the sounds `P', 'T' and The 5.86"AttheElectrotechnical-Laboratory, essential difficulty at present comes in designingwhere analysis and synthesis of both the hearing circuits which will distinguish the sounds when and the speech organs are under investigation they differentspeakersclearly enunciate the same are trying to develop a unique method of recogni- phonemes. The final stage will come in distinguish- tionbased on testing the correlation between ing different sounds spoken by anybody in nlrmal generated and received speech....Researches rapid speech. The principles being used in the MIT [in speech analysis, synthesis, and recognition] study are those of recognizing distinctive differ- are also being done at the University of Kyoto, ences, rather than recognizing patternson the Kokusa Densin Denwa Co., Tohoku University, theory that when a man is trying to find his way and the governmental Electric Wave Research With a map, a small number of judgments made Laboratory. Their present situationisthat not correctly is. sufficient for him to tell where on the more than vowels can be recognized." (Niwa, map he is. None of the workbeing done at MIT at 1962, p. 65). present includes the problem of subsequent correc- 5.86a"At the present time there are several tion of the sounds heard by clues from context." projectsinexistence involving theanalysis of (Berkeley, 1956, p. 9). speech waveforms. Such fields of study are: (a) 5.85b"The clusters for this experiment were development of communication systems that use generated by taking measurements on sustained digitizedand/or parametric speech;(b)experi- spoken vowels. The utterance of each vowel was ments with automatic recognition of spoken com- sampled at a 20-millisecond rate through 15 levels, mands; (c) study of phonetics through models of and the outputs of the filters were quantized into the vocal system and synthetic speech; (d) seg- 13 levels. Each time sample was taken as a pattern, mentation of words, and human perception tests. thus producing a duster of patterns from an utter- For many of these applications conventional de- ance of one vowel. Eight vowels were used ...The vices for examining speech, such as the fixed- set of 250 patterns, viewed as samples of unknown filter spectral analyzer, are inadequate. Additional space for purposes of this experiment wassubjected specialized equipment isrequired and there is to a series of analyses to determine whether clusters increasing interest inNprior digital simulation of existed and whether the method would separate these experimental systems as an aid to design. 130 Alternatively digital, rather than analogue, process- a band -pass filter-bank speech spectrum analyzer, ing of the speech waveform may be used. This with 51 channels that can be set in various combina- method is capable of greater flexibility and accuracy, tionsof center frequencies and different band and new techniques such as zero-crossing analysis widths. Input signals are digitalized after filtering readily lend themselves to computer calculations. in order to allow synchronous sampling of all In the setting-up of digital speech processing facil- channels. Output may be presented visuallyor ities, two problems may be distinguished:(1) recorded on magnetic tape in digital form, or it provision of an input/output device to convert may be reconverted to analog form for display via the analogue speech waveform to a numerical CRT or on X-Y plotter. Investigations can be representation which may be stored on magnetic made of effects of different frequency transpo- tape;(2) development of programs to perform sitions upon input signals, changes in the time code various kinds of analysis on the digital record." system, and the like, and the analyzer results can (Lavington and Rosenthal, 1967, p. 330). be compared with those obtained from the speech "A simple 16-word vocabulary has been used synthesizer. to assess the effectiveness of these measurements. The vocabulary consists of monosyllables formed A number of different techniques for the improve- by combinations of one of four initial consonants ment of the basic low-pass filter have been tried (S, F, T, N), one of two vowel-sounds (ÈE', or and, as a result of the experienceso gained, later investigations have provided for the use of twoor ÒR'), and one of two final consonants (S, N). more filters in parallel so that the disadvantages A recognition program has been testedon a total of a particular system can be compensated by of 70: words spoken by 19 speakers (14 male, 5 circuits female; various accents including Cockney, Lanca- havingcomplementarycharacteristics. shire, Glasgow, American, and Southern English). (Risberg, 1962 see also Stevens, 1968). AveragecomputingtimeontheUniversity's 5.89"(Theoutputofthesonograph [Kay Atlas was 1.2 sec per utterance, of which about Electronics Co.] is a continuous record ofa fre- 0.35 sec was spent in assembling the digitized quency and amplitude vs time; this record is called speech from magnetic tape to the main store prior a sonogram.)" (King and Tunis, 1966, p. 67). to analysis. (The speech was previously put onto 5.90"A simple model for human speech was tape via a tape-recorder or direct Microphone input originally formulated by H. W. Dudley. It relies to the Speech Converter.) The main program was on the observation that during a human speech written in Atlas Autocode, and is at present being utterance (especially during the 'voiced' portions), converted to machine code. Preliminary tests have the acoustic energy is mainly concentrated in only shown that this will reduce the overall tire (in- a few relatively narrow regions of the frequency cluding data-assembly) to below 0.7sec per utter- spectrum ....These energy concentrations- are ance. The size of the Atlas Autocode object-program known as formants. During the utterance of .a word is3,000 machine-instructions."(Lavington and the position and relation of the fi,rmants change, Rosenthal, 1967, pp. 338-339). creating a characteristic 'pattern'distinctive of 5.87"In a case such as the present experi- a word and, to some extent, the speaker." (King ment where the patterns [eight classes of vowel and Tunis, 1966, p. 66.) sounds] emerge from the real world rather than "The second major characteristic of interest is from a pattern-generating rule, there can beno the marked tendency of frequency selective energy assurance that any particular sample belongs to concentrations, referred to as formants." (Pant, its parent class in accordance with any reasonable 1959, p. 6) Other definitions are as follows: "Vowel objectivecriterion."(Dammann, 1966,p.87). 44. sounds are cLaracterized acoustically by formants, The analysis sample of one speaker is not a good statistical representation of the test sample which are frequency regions of high energy concen- of a different speaker. It must be noted, however, tration corresponding to the passbands of the throat andmouthcavities."(Phonetics that the performance of the systemwas signifi- Laboratory, cantly better than chance, implying that there is University College, London, 1963, p. 1). some degree of statistical invariance in the measure- "The vocal tract resonances, called formants, ments ...from one speaker to another." (King are determined by the shape of the vocal tract and Tunis, 1966, p. 74-75.) which, in turn, depends on the position of tongue "A particularly difficult requirement tofulfill and lips. As we pronounce phoneme after phoneme, is that the recognition process must produce the we change the position of these vocal organs, and same results from the voices of different speakers therefore also the formants, the spectrum, and with different personal characteristics." (Kusch, the character of the resulting sounds." (Denes, 1965, p. 204). 1966, pp. 246-247). 5.88"Filter bank analyzers have been much 5.91"A new speech perception theory... used (e.g., by Abramson, et al., Denes and Mathews, known as the Single Equivalent Formant and its Olson and Belar, Davis et al. (1952), and Talbert associated parameters developed by Philco Ad- et al.)..." (King and Tunis, 1966. p. 65). vanced Communications Laboratorycan provide The speech analyzer presently in operationat a sound economical basis for a speech recognition the Royal Institute of Technology, Stockholm is system." (Scient. Info. Notes 7, 21 (1966)) 131 5.91aSee Kusch (1965). For example: "Recog- 5.93a"In considering the recognition system, nition is effected with the help of a coding matrix we intended to accept all the vocabulary commonly for sound groups and another for digits... used and not be limited to a certain category of "The relevant characteristics are obtained from words. To satisfy this requirement the phoneme the form analysis of speech oscillogram...Two was selected as the basic recognition unit. ... vibrations can be extracted from the overall signal zero-crossing pattern of the original signal with adequate reliability by means respectively is closely related to the spectra of the original signal of a iow-pass and a high-pass filter with non-critical and its peak corresponds to the formant. . cut-off frequencies. The vibrations show different "The score of the machine for the Japanese mono- amplitudes, which areto be distinguished as syllables (which contain a consonant and a vowel) `large', i.e., the vibration is present...And as is more than 90 per cent for the vowel part and more `small', i.e., the vibration is absent... than 70 per cent for the consonant part except "The spoken numbers were on the average cor- nasal consonants, and In;'nd /b/ sounds for the rectly indicated to the extent of about 87%." pronunciationofseveralmales."(Sakaiand (Kusch, 1965, pp. 201-203). Doshita, 1963, pp. 836, 839, 845). "The speech signal is essentially resolved into "Segmentation of the vowel interval is performed its component frequencies by means of filters, as by extracting the parameters that describe the for example in the vocoder technique. The fre- aforementioned intervals from the digital pattern quency components that predominate in the fre- of speech sound analyzed by the zero-crossing wave quency analysis are termed formants (characteristic analysis. We defined two quantities called 'dis- frequencies). These formants, which have already tance' and 'stability'. Stability expresses the station- frequently been used as parameters, do not con- ary property of pattern, that is, the property that stitute a reliable basis for the rec'q.nition of all the parameter remains in an almost constant state sounds because of their frequency displacement over a certain interval. On the contrary, distance from one speaker to another. Account has to be is the quantity that expresses the change of the taken not only of the sounds themselves, but also pattern. Stability is useful for discovering the vowel of the sound transitions between consonants and sound and the fricative sound, and distance is vowels." (Kusch, 1965, p. 204). for the sound with burst, such as stop consonant." 5.92In particular, von Keller has tested a/pro- (Sakai and Doshita, 1%3, p. 837). cedure, using 300,us. time intervals, to determine 5.94Further, "the means of automatic discern- the number of zero-crossings in order to separate ment (analysis) and automatic production (synthesis) the different vowels. Using as criteria the mean of sounds of speech require the rules under which lengths of distances as compared to pre-deter- a plurilogical system, i.e., a collection of pluri- mined values, he has been able to separate five of logical elementary units and a manner of thesis the eight vowels. (Stevens, 1968, p. 27;. von Keller, contributing, can be correlated with real sound 1966). manifestations. Apparently, a relative simplifica- 5.93In the word classifier developments, the tion this task is to select a plurilogical system con- objectives are to discover criteria! features that sisting of a collection of differential signs... are not phonemic segments but that will successively discriminate more different classes as addi- "In correlating the constructed system of dif- tional features are considered. A speech recogni- ferential signs with real sound manifestations, the tiou device was constructed for twenty classes main difficulty consists of differences between the consisting of the ten numbers and ten words, in phonetic norms characteristic for the speech of Italian, for Euratom, ISPRA. any person. Similar problems appear in the study "The University of Bonn has developed a machine of a correlation of the system of graphic differential that understands Italian and types out answers signs with the systems of graphic norms, deter- on "command. Known as DAWID, Device for Auto- mining differences between various handwritings matic Word Identification and Discrimination, the and typographical faces. A solution of the latter machine, at the moment, has a vocabulary of 20 task is required for programming operations for Italian words. It is claimed that it makes fewer reading automatic machines." (Ivanov, 1962, pp. errors than IBM's Shoebox system which only understood 16 words and had an error rate of 6 5.95"STAR(StandardTelecommunications per cent. Automatic Recognizer) is a research machine being "Dawid contains circuits which detect the ex- built at STL to further work on speech recogni- plosive sound at the beginning of a word and tion...Four types of pulses (beginning and end measure delays on the remainder of the word. of rise, and beginning and end of fall) are derived The energy level of sound =": below 400 c/s is also from the short-term mean power envelope of the detected and the formants of t he sound are fre- speech wave representing an isolated word. Any quency analysed. A phonetic and easily readable or all of these may then be used to provide group- language in which such a machine could write ing of the other features detected, thus simplifying down messages has also been developed." (See also the order information. There are two main advan- Data and Control Systems, May 3, 1966, p. 65, tages in this approach. The grouping is related to Tillman et al., 1965). the broad structure of the word to be recognized, 132 and there is automatic time normalization meaning- segment. But the parameters of distributions of fully related to the word." (Hill, 1965, p. 357). a segment are affected by the phonetic contextual 5.96It is claimed for this experimental speech effect from the neighboring phoneme. The speech recognition system that "this work extends the sound, therefore, must be recognized as a whole results existing in the literature in that it deals pattern considering phonetic context. This principle, with significantly larger sample sizes than have however, isdifficult to apply to the recognition been commonly used, with a limited number of system of general conversational speech." (Sakai limited vocabularies, and with the effect of trans- and Doshita, 1963, pp. 842-843). formations of the primary measurement space on 5.97See, for example, a report of work by King recognition performance." They state further that and Tunis based on the hypothesis that the spectrum "the recognition performance did not drastically analysis of speech waveforms will provide measure- deteriorate as the number of words in the vocabu- mentscontainingorcomprisingstatistically lary was increased and suggests that good perform- invariant measures of the spoken word. They ance on vocabularies larger than 15 words is achiev- describeaprogramming systemthatapplies able." (King and Tunis, 1966, p. 65). various transformations to the input measurement 5.96a "CYNTHIA III is an adaptive hybrid com- space."Thistransformed measurement space puter, developed by Andromeda, Inc., of Kensing- then is used as the input to the categorizer section ton, Maryland, which can learn to recognize highly of the recognition system. The categorizesection variable statistically distributed signals. It must go consists of linear decision functions in which the through a preliminary learning process before rec- weights are obtained using an adaptive algorithm." ognition is effective. After learning is finished it (King and Tunis, 1966, p. 65). will quickly recognize the respective signals applied 5.98See Wooster, 1964, p.13,as follows: to its input by producing an unambiguous output "It is becoming increasingly clear that speech recog- coded response which identifies the signal class nition can not be done on the basis of the acoustic to which an input signal belongs. The machine is properties of the speech signal alone; that general capable of extrapolating or generalizing so dm. soluticts will rely upon the interplay of linguistics new signals, never presented in the learning opera- and semantics." tion, are recognized correctly if they belong in the Similarly, Lindgren reports: "Somehow. ways learned classes. The machine also provides an must be found of incorporating linguistic informa- internal digital memory and a visual display of the tion in the decision-making functions of possible characteristics of the statistical signals which render speech recognizers." (Lindgren,1965, p.121). them separable.... "Some observers place the resolution of speech "The machine does not learn individual signals recognition problems squarely on the determination when a large number of signals are presented, but of the laws of language." (Lindgren, 1965, p. 46). learns output classes to which certain significant "It becomes clear, then, that quite apart from its signal characteristics belong. All responses which own problems, the success of automatic speech deviate from any of the averages of a set of classes recognition depends in the last analysis on the suc- are extrapolations. The average itself may not be cessful solution of the problems involved in parsing present in an actually previously presented and and semantic analysis." (Spolsky, 1966, p. 494). learned signal, but is established by learning. The It is to be noted in particular that: "Our under- machine ignores other characteristics not in the standing of these contextual factors is, as yet, slight. recognition classes and which are unnecessary for We believe that, the mechanical and anatomical identification. The present machine can deal with properties of the vocal tract play an important part isolated signals such as words or short sentences as well as a wide range of grammatical, linguistic takenasmeaningful wholes, but cannot cope and semantic factors. The anatomical, physiological, with continuous speech. Extension of its powers linguistic and other factors have a great influence are being planned for the recognition of continuous on the acoustic features produced when a phoneme speech." (Lesti, 1963, p. 279). is pronounced in context as opposed to when the 5.96b"To use phonetic notation is clearly simpl- same phoneme is pronounced in isolation; it is also est for the first stage, converting the continuous more than likely that as listeners we make :Ise of speech signal to discrete symbols. It is theoretically our knowledge of the operation of these factors feasible to develop a method of analyzing and seg- acquired unconsciously when we learn to speak to menting the acoustic signal, using phonetic symbols unravel the ambiguities of the acoustic cues. The for specific patterns recognized, but in practice this study of these factors, of the dynamics of speech, is has proved to he extremely difficult. Thus, byat present the major research interest of Speech 1%0, all that could be claimed was the ability to Science." (Denes, 1966, p. 249). recognize spoken digits and some vowels in isola- 5.99See, for example, Dolby et al., 1965, p. 3-3: tion(Marrill,1960)."(Spolsky,1966,p.493). "It is one question to ask whether a particular letter "In the speech recognition system described sequence is an affixing sequence, and quite another ...,the speech sound was divided into several to ask whether it is an affix in a particular word." segments (that seem to correspond to the phoneme) 5.100"The FarringtonManufacturing Com- from the time variation of file analyzed pattern and pany, New York, N.Y., a leading producer of Optical thendiscriminationwasperformedforeach Character Readers, is expanding into the voice 133 identification field by agreement to purchase Voice- matic print identification, such as for a personal print Laboratories, Inc., Norville E. White, Chair- identification system. In his preliminary studies, man and President of Farrington hasannounced. Cuadra uses manual detection of minutiae. Trauring The agreement provides that Voiceprint Labora- assumed idealized feature detectors." (Hankley tories, with headquarters in Somerville, N.J., will and Tou, 1968, p. 417). be operated as a Farrington subsidiary. 5.100d"Van Emden has suggested a single "Voiceprint identification is a method of identify- print classification scheme using multiple ridge ing individuals by a spectrographic examination of counts between selected minutiae, but also assumes their voices which is nearly as accurate as finger- idealized feature detectors." (Hankley and Tou, print identification. Mr. White said that Voiceprint 1968, p. 417). identification 'will become a useful and valuable tool "Scientists and technicians at Litton Industries in the data processing field. In addition to use in Advance Data Systems have been conducting a credit verification systems, it is expected that during series of experiments and studies in the field of the next few years voice identification will be used automatic fingerprinting processing for approxi- for direct communication between depositors and mately three years... a bank's computers via telephone'."(Computers "Consider now the key problem in the field. and Automation 16, No. 10, 55 (Oct. 1967).) What informationisto be extracted from the 5.100a"The Justice Department last week finger ridge pattern? If the FACT [Fingerprint launched a $500,000-plus study of the human voice Automatic Classification Technique] technique is as a means of identifying criminals. The work could employed, the finger ridge pattern is scanned and revolutionize law enforcement. the location of all minutiae that has ridge endings "The program is planned as a 3-year project, but and ridge bifurcations or branches is temporarily the Department is committed only to the first year's enteredintostorage ...Thedataextraction research. Tuesday, it awarded a $146,509 grant to device then returns to each of these locations and the Michigan Department of State Police for the counts the number of ridges intervening between first phase of the project. all combinations of minutiae... "Justice officials said the study will cover such "Using the Fingerprint Automatic Classification voice characteristics as pitch, timbre, resolution Technique, the minutiae table is processed using intensity and the phase spectrum of sound waves. a proprietary technique evolved by the staff of "They also want to learn the degree of precision Litton Industries. In actuality, four different process- needed for adequate voice identification. ing techniques have been evolved using the minutiae " 'You may say a certain word one way one day and table. They all reduce to a binary descriptor word another way the next. We want to know where whose length is a function of the number of minutiae individualpeculiaritiesbecome unreliable',an encountered in the original print. This descriptor official explained. typically runs somewhere in the vicinity of forty "The project, a new venture for the Department, bits but could go as high as 120 bits in some remote was proposed by the Michigan police. cases. This descriptor is well ordered and prelim- "First year's work will be channeled into three inary tests have indicated it to be unique." (Van subcontracts: Emden, 1967, pp. 493, 496, 498). "Dr Department of Audiology and Speech 5.100e"After elementary spatial filtering, an Science at Michigan State University will interpretationprogramsystematicallysearches make a validation study of spectrographic along ridges, correcting for gaps, contiguous ridges, equipment already on the market. etc., to detect the print core and topological fea- "Stanford Research Institute will make a state- tures. A topological coding program generates a ment-of-the-art examination and probe alterna- sequential code sentence which describes the topo- tives to present equipment. logical structure of the center region of the print. "Michigan State's School of Police Adminis- Ordering of code sentences is used to classify finger- tration will cover law enforcement aspects. prints." (Hankley and Tou, 1968, p. 413). "Justice officials said the probe, for all its depth, "All lines are directed downward from locally is not intended as a 'definitive piece of research'. uppermost points, labelled Ti and called top nodes. But after the 3. years, they aim to know if voice The choice of ridge tops as nodes may be justified identification is feasibli for law enforcement and by their use in topological coding and by the fact if more study will be worthwhile." (Electronic News that humans place comparatively heavy emphasis 13, June 24, 1968). on such points of locally maximum curvative." 5.100b"Fingerprint identification is a very real (Hankley and Tou, 1968, p. 423). possibilitywiththecomputer-controlled CRT "The proposed automatic fingerprint interpreta- reader. The technique might well produce a more tion and topological coding schemes were simulated complete method of cataloguing and comparing on an IBM 7094. The Ohio State University SCAT- fingerprints automatically and at very high speeds." RAN language was used, which allowed bit manipu- (Fulton, 1963, p. 40). lating operations and logics to be expressed in 5.100c"Trauring and Cuadra have suggested essentially FORTRAN-likenotation."(Hankley matching of minutiae patterns as a basic for auto- and Tou, 1968, p. 444). 134 Additional features are as follows: mented into a number of small areas, the slope "(1) The proposed scheme is a single-print sys- of the ridges passing through each area allows one tem. Sufficient information is extracted from to recreate, using only the slope data, a rough each print to allow a fine classification. approximation of the print. Having this rough ap- (2) The interpretationprogram learnsprint proximation of the print allows one to divide the structure and performs contextual filtering total fingerprint population into a number of groups, to correct for imperfections such as ridge thus allowing one to determine into which group an gaps and contiguous ridges. unknownprintbeim, eswithonsearching the (3) The interpretation logic is evolved via a entire file." (Shelman, 1967, pp. 467-469). learning mode using training data and inter- "If complete automation is to be accomplished, action with human programmer. the scanner should have provisions for translating (4) The coding system is topological ...The and rotating a print to establish a reference system. topological structure of a fingerprint is in- If a translational and rotational invariant descriptor variant to distortions due to rolling or stretch- can be developed, this constraint may be relieved." ing of the skin or to changes in the size of the (Shelman, 1967, p. 477). finger. it is also invariant to translation and 5.100g"It is the purpose of this paper to con- rotation of the prints. sider optical processing of two-dimensional images (5) The classification technique is highly com- or patterns and to stress its usefulness for finger- patible with the manual mode. The rules are print processing...An angular averaging method, easily learncd., and classifications may be which provides rotational invariance of the spectra, verified by inspection." (Hankley and Toll; in addition to their inherent translational invariance, is shown to be suitable as a coarse classification 1968, p. 413). method. In addition, the use of cross-correlation Hankley (1968) reports further as follows: "This techniques for identification of a fingerprint against paper presents a new scheme for topological classi- a lirnitA set is discussed, and some other possible fication of single fingerprints. As a basis for a finger- schemes of optical processing are briefly men- print file system, it allows a fine ordering of in- tioned." (Marom, 1967, p. 481). dividual prints using a concise coding of only a small 5.100h"A veryspecial hologram, calleda core area of each print. Moreover, the classification spatial filter, has the capability of comparing two is highly compatible with the way in which humans patterns and producing a signal which is a function naturally describe prints. These properties make of the correlation or similarity of the patterns. practical today an automated file system with re- Experimentally, it has been found that complicated, mote manual input of print code. Alternately, the natural objects with irregular patterns can be scheme lends itself to automatic photointerpretation, recognized with greater confidence than can man- which is described elsewhere in the literature." made objects which tend to be geometrically sym- (Hankley, 1968, p. 1). matrical. Fingerprints, because of their random- 5.100f"Under the control of the computer, the ness, appear to be ideal objects for the spatial scanner scans selected areas of 35 mm film, either filtering method of recognition... in the horizontal or vertical direction. The line scan- "Eight fingerprints... wereobtained from the ning density may be varied. The scanner sends to New York State Identification and Intelligence the computer the x and y coordinates of boundary System...All of the prints were very similar and crossings between light and dark areas. It also would have the same general classification. The identifies whether a light or dark area is being prints were copied on 35 millimeter film and a entered... spatial filter was produced from the print desig- "In the experiment we conducted, 128 small nated 'E'. The strip containing the eight similar squares per print were scanned. The square sizeprints was then passed through the recognition was approximately 41 ridges on a side. Proper system with the spatial filter in position and the positioning of the boundary was checked for each recognition plane was photographed. The only print area. This was accomplished in the following man- that produced an intense recognition spot was finger- ner. First an area was scanned. Then the amount print 'E', the correct print. Several of the other of dark area was calculated. A ratio was formed prints produced some light scattering in the rec- between the amount of dark area and the total ognition plane due to partial correlation with the area. If this ratio was between the limits of one- filter, but the light was not focused to a small spot third and two-thirds, the sample was accepted for and measurement showed that the intensity of the add!;;onal processing. If it was outside these limits, spot produced by fingerprint 'E' was 175 times the boundary was changed, the print was rescanned greater than any of the noise produced by the other and the ratio test repeated... prints. There is absolutely no doubt that fingerprint "There are two types of information that a È' was the only print which correlated with the machine can extract, namely ridge endings and filter. Other experiments supported the conclusion ridgeslopes.Ridge endings can be used for that the spatial filtering process is very selective secondary classification and positive identification and is not confused by two very similar prints. In of an unknown print. Ridge slopes can be used fact, attempts were made to generate a filter which for primary classification. If a fingerprint is seg- would recognize all prints of one particular classi- 135 fication and in this way the spatial filtering process A system to perform this function would encode a could be used to classify prints automatically. So fingerprint pattern to be identified so that it could far we have been unsuccessful, largely because the be compared at high speed with similar patterns, spatial filters are too selective and tend to recognize available in a e9mputer type storage facility. An prints by their fine details and ridge spacing rather approach to encodir g a fingerprint, using data about than simply by their general patterns. This is ad- bifurcations, or ridge splits, is described. Some vantageous when you are attempting to recognize relationships between a suggested approach to one particular print, but disadvantageous when you matching fingerprint pattern and correlation detec- are attempting to classify it." (Horvath et al-, 1967, tion.procedures used in various modern communica- pp. 485, 489-490). tion systems are cited." (Gaffney, 1967, abstract). 5.1001"In September 1965, the FBI submitted 5.102"The system concept that was examined in to members of industry who had displayed an inter- our research ...attempts to capitalize on the est, a request for proposal stated in very broad information represented by the ridge character- terms. This proposal solicited ideas for the develop- istics. Unlike the other approaches, which aim at ment of a device that would accurately read and complete automation, it proposes a man-machine classify inked fingerprints. By the end of the year, operation, with the man locating the ridge char- we have received a number of responses . .It acteristics and the machine (computer) performing appeared that the greatest consideration had been computation and matching of the locations ... given to the use of general purpose scanners; how- The computer would transform the locations into. ever, it became apparent that a number of organiza- a standard frame of reference and perform avariety tions were giving serious consideration to the use of calculations on the relationships among core, of coherent light and a system of optics to do the delta,andcharacteristics,toprovide'index' reading... handles for later comparison." (Cuadra, 1966, p. 9). "An indication of the industry's interest in this 5.102a"One advantage of neighborhood logic area can be garnered from the fact that a prepro- over other pattern recognitionschemes is its in- prosal conference held at FBI Headquarters on variance to orientation...The PLUS instruction January 12, 1967, was attended by representatives is a symmetrical operation and as such it does not of more than 30 interested companies." (Voelker, depend upon the orientation of the bifurcations. 1967, pp. 446-447). Thus, a single sequence of instructions extracts all 5.100j"With the current state of the art, it is the bifurcations which are in the fingerprint. Since not possible to tell whether an automatic coding in the inverted image (negative), a ridge ending technique can be developed and put into practical resembles a bifurcation, the algorithm can be operation over a period of the next five to ten years applied to the inverted image to extract ridge end- or so. This depends as much upon interest in the ing locations." (Freedman arid Hietnanen, 1967, problem and financial support as it does upon tech- p. 508). nological development. With this somewhat uncer- "At present, work is continuing at Bendix Re- tain prognosis, it seems desirable to examine alter- search Laboratories to develop an electro-optical nate coding procedures while still maintaining the computing system which will be capable of process- computerized efficiency of the rest of the system. ing fingerprints in parallel using neighborhood logic. An obvious appraoch to such an alternate tech- It is felt that such a system will have a great deal of nique was a hybrid system where a human would potential to solve many other pattern recognition perform some of the more critical pattern recogni- problems including some in the area of surveillance tion functions and a machine would perform the and detection. Thus, the system will be extremely remaining functions. This hybrid system can be flexible, just as the biological pattern recognition called a semi-automated system. If a fully auto- system is, and it would, therefore, fill the gap left mated coding technique were eventually to be de- by existing equipment." (Freedman and Hietnanen, veloped, it seems highly probable that there would 1967, p. 509). still exist many situations in which it could not perform satisfactorily. This is particularly true with 5.102b"Several approaches have been made to automatic fingerprint identification including holo- respect to the coding of many latent fingerprints graphic techniques. The techniques that use details fonnd at crime scenes." (Kingston, 1967, p. 450). of a fingerprint such as cores, deltas, ridge counts, "Thus we have two major reasons for developing ridge endings, and ridge bifurcations tend to be a semi-automated technique:(1)itcould serve specialized according to the method by which the to get a computerized fingerprint processing system information is taken from the fingerprint. Thiebault operational sooner than would be possible if we has described a technique using the directions of were to wait for fully automatic procedures,and (2) minutiae (ridge endings and bifurcations) where the it could serve as an ancillary input to a fully autodata are obtained by a special camera that photo- mated system when this is developed." (Kingston, graphs the finger. Where a human operator identi- 1967, p. 450). fies the core and delta as a reference axis, a match- 5.101"During the past several years, various ing technique has been developed that utilizes organizations have been investigating the possibility only the positions of the minutiae. The technique of identifying a fingerprint by automated means. described in this paper is intended primarily for 136 use with an automatic reader that reads minutiae assumption that an algorithm can be developed directly from a fingerprint card without identify- which will generate a venation pattern in which ing cores and deltas. certain stochastic features perturb an essentially "This paper describes a procedure for matching deterministicprocess. The parameters of the fingerprint impressions by computer using only two algorithm could then be regarded as the features types of minutiae: ridge endings and bifurcations ... serving toclassify the leaf types according to [The] matching procedure is based on matching a species."(QuarterlyProgress Report No.80, constellation or group of minutiae formed about a Research Laboratory for Electronics, M.I.T., 221 particular minutia called the mucleus from one im- (1966).) pression with a corresponding configuration from 5.103a"Seven typesof patterns were con- another impression...Although a partbular considered. Four of these were features of the lunar stellation Hexes and stretches, it retains a certain terrain: craters with flat floors, craters with central pattern identity in impressions taken from the same peaks, rima, and wrinkle ridges...The remain- figure over a period of years. Matching just one con- ing three classes of patterns were types of cloud stellation from one impression with a corresponding cover appearing in NIMBUS imagery: noncumulous, impression appears to be sufficient for concluding cumulous polygonalcells, and cumulous solid that both impressions were made by the same finger. cells." (Darling and Joseph, 1968, pp. 457-458). "Both relative distances and relative angles are Another approach to cloud pattern recognition, used to match a constellation in one fingerprint im- at Astropower Laboratory, has involved the compression with a similar constellation from another puter simulation of a Perceptron-type approach. impression..." (Wegstein et. al., 1968, pp. 1-2). (Joseph et al., 1964). "It is interesting to note that the variation in the 5.104"A groupisconducting research on angle between minutiae directions tends to be quite respiratoryailmentsatSouthwestern Medical different than the variation in the distances. This School in Dallas, Texas. One parameter of interest suggests that differential emphasis might be placed is the number of times that a patient coughs during on angular information versus positional informa- a given interval. Microphones are placed in selected tion in the development of a descriptor. It is clear hospital rooms and audio recordings are made... that any descriptor must be tolerant of these varia- In addition to the requirement to classify the audio tions and would, therefore, be redundant. It would signal as cough or non-cough, the physiological not be used with the expectation of a perfect fit but model of the cough indicates that decision theory rather in the sense of 'best match'. Its development may profitably be applied to the waveform to identify would represent another challenging bench mark the cough as having emanated from individuals along the route to a possible automated fingerprint suffering from different broad classifications of identification system." (Wegstein and Rafferty, 1967, respiratory diseases." (Welch et al., 1967, p. 260). p. 464). 5.105"The recognition of sheet music is being "Various minute details in a fingerprint might be studied as an example of pattern recognition in utilized in a descriptor. For example, the character- partially structured pictures. 'Partially structured' istics of sweat pores are as individual as the minute refers to a picture composed of well-defined forms details of ridges. Unfortunately, special processes or 'characters' and of forms that are only partially are required for making impressions that show sweat defined. The need to handle both types of forms pores. On the other hand, ridge endings, bifurca- and to deal with the relations between a number of tions, islands, enclosures, incipient ridges, and forms simultaneously makes this problem more breaks in ridges readily appear in inke.d fingerprint pictorial in nature than character recognition alone. impressiops. Furthermore, if just a few of these The scheme that is being utilized includes nonlinear minutiae are the same in two different fingerprint background elimination, contour tracing, and feature impressions, then there is a very high probability extraction... that both impressions are from the same person." "With the aid of the scanner and ancillary equip- (Wegstein and Rafferty, 1967, p. 461). ment ...it is possible to program a computer to 5.102cIt is similarly pointed out that: "Although determine note value, duration, phrasing, and cer- simple and economical, the familiar process of tain other music indicators from ordinary printed recording inked fingerprints is liable to a variety of sheet masic.... errors and inconsistencies. There is wide tolerance "At present, the program successfully processes in positioning the finger. Variations in inking cause and recognizes one general class of patterns, namely marked differences in contrast between fingerprints the timing-bar complexes attached to notes con- as well as in different parts of the same fingerprint. taining the time information for the notes. Progress Ridge details are lost through over- and under-inking. continues along these lines with work now being The flexibility of the finger makes the image depend directedtowardrecognitionofgeneralnote on the way the finger is pressed and rolled. The clusters.... extent to which the sides of the finger are recorded "By using this scanner, pictures of music were is highly variable." (Schwartz, 1967, p. 511). scanned into the TX-0 and written on magnetic 5.103"Recognitionof PlantSpecies from tape ina format compatible with FORTRAN. Leaf-Vein Patterns. We begin this study under the Nonlinearverticalandhorizontaldefocusing

137 (background elimination) were effected on these 5.107See Swanson, 1966, p. 4. pictures by using the IBM 7094 computer. Tapes 5.108"At the Royal Institute of Technology, were then generated containing pictures obtained Stockholm, S. Karlsson of the Institute for Tele- by various logical combinations of picture arrays, communication Theory is engaged in the study such ascomplementation and logical `anding'. of problems of pattern recognition with respect The results of using these processes were various to electrocardiographic recordings. Karlsson and types of modified pictures, such as those with hori- Arvidsson have developed a programming system zontal and vertical lines erased, those with nothingfor EKG measurements, using an approach some- present except horizontal lines, and so forth." what similar to that of Pipberger (1965) at the (D. H. Pruslin, Quarterly Progress Report No. 80, U.S.Veteran'sAdministration.Karlssonstarts ResearchLaboratoryforElectronics,M.I.T., with a proceduretoformulateinitialclusters 221, 225-226 (1966).) such that a member .of a cluster has its nearest 5.105a"I have already been approached by neighbor in the same cluster (e.g., in the peaks persons interested in tracking baboons in theirof the recording), and repeats for the next nearest migrations across Africa. The problem is: Can the neighbors and so on. The purpose is to derive a rump prints of baboons be taken by a machine such measure for each pattern category by extracting as the FACT finger scanner, processed and held features common toall members of the same in file for future comparisons when these same cluster." (Stevens, 1968, p. 33) animals are again captured? While this seems very 5.108aSee, for example, Pizer and Vetter, comical, this and other side benefits could accrue 1968, on the perception and processing of medical from the work done in the name of automatic radioisotope scans and the work of Butler (1968) fingerprint processing." (Van Emden, 1967, p. 504). on automatic analysis of bone autoradiographs. 5.105b"Itiswise to remember that many 5.109"The general approachisalso being problems in music are complex. The extent to appliedto problems of medical diagnosis and which data-processing technology will render solu- data interpretation such as are encountered in tions more accessible than do traditional procedures electrocardiographic, phonocardiographic, and remains an open question. For example, it now x-ray interpretation." (Teager, H. M., Quarterly appears that computer-generated graphic displays Progress Report, No. 80, Research Laboratory for offer new resources for the editing procedures Electronics, M.I.T., 269 (1966).) that are central to much work in musicology. Yet, 5.110For example, projects at the University a great deal of work must be done before machine- of Pisa are concerned with automation of the implemented editing can cope with the notational analysis of radiocardiograms. "This involves a non- systems of various historical periods or with such a linear programming problem to determine transfer complex representation of the human creative functions of interest, requiring detections of the process as a page from a Beethoven sketchbook. first three moments, average dispersion, symmetry, In attempting to cope with such problems, however, as characteristic of desired curves. Similarly, there we can expect that traditional music scholarship is a project to extract time sequences of nerve will obtain insights that may determine extensive pulses and to determine coding in the optical criticalrevisionsof conventional methods and system. For this purpose, a special-purpose analog- criteria." (Forte, 1967, p. 329). to-digital converter has been designed. A more 5.105c"Before processing, scores are encoded generalized converter is under development for by hand in a language designed by Stefan Bauer- a variety of biomedical investigations including Mengelberyinconnectionwithacomputer- EEG, EKG, and other problems involving pattern ithplemented project to print scores automatically. recognition with respect to clinical data." (Stevens, This input language has unique attributes: 1968, p. 32). (1)It is isomorphic to standard music notation. 5.110a"The automatic computer analysis for (2)It is highly mnemonic, hence easy to learn. the determination of the mitotic index involves (3)The responsibilitiesof the encoder are scanning with the FIDAC...Since a mitotic cell minimal. The encoding rules are unam- is in the process of fragmenting, it has a much biguous and do not require the encoder more irregular boundary than has a nonmitotic to make arbitrary decisions. cell.To distinguish between mitotic and non- (4)It is economical. The amount of code for a mitotic cells, then, the computer is programmed completerepresentationis remarkably to count the number of concavities smaller than small." (Forte, 1967, p. 4). a given segment length in the silhouette of each 5.106"As an aid to the physician, to release object or cell, and compare this number with the his time for other types of patient care, the In- total number of such segments on the boundary. strumentation Field Station of the United States The greater the proportion of concavities, the more Public Health Service has developed programs for irregular the boundary and the more likely it is automaticanalysisof waveforms such asthe that the cell is mitotic." (Ledley et al., 1968, p. 100). electrocardiogram, phonocardiogram, forced 5.111"The CYCLOPS 1 appears to be among respiratory spirogram, electroencephalogram, and themostsophisticatedand powerfulpattern- other medical signals." (Weihrer et al., 1967,p. 289). recognition devices yet reported;itscapability 138 to do quite elaborate scene analysis is perhaps thatis,toidentify, separate and position the unique." (Hart, 1966, p. 16). different objects or bodies belonging to the scene(s). For example, "the program may be commanded The demands of information will vary: sometimes we from the typewriter to analyze a complex scene. will be interested in knowing if an object is seen Such a scene may consist of known items, i.e., in the scene or not, while at other times we may shapes which have been defined as being significant require a complete description of the scene, in- (the alphanumeric characters); and of other items, cluding information on relative support and (3-dim) some of which may be significant to the viewer position of the different components." (Guzman- but not to the system, and some of which may be Arenas, 1967, p. 1). merely background 'noise'. An arbitrary number of "For planning and control of the overall activity known items may be present simultaneously; they of the machine, we are taking into account texture may be of di Bent sizes and orientations; they may and color, as well as object boundaries and the like. overlap, or be inside each other; they may be super- The system will be able to partition a scene into imposed on an arbitrary background." (Marill et al., objective regions, combine these regions into pro- 1963, p. 28). posed objects, and finally present this collection The CYCLOPS -I scheme involves a series of of pseudo-objects as an abstract model of the objects finding operations by looking for what might be and background in real space." (Minsky, 1966, there and by progressive eliminations of possibilities. p. 16). Thus, "a negating tree system has been used by 5.113Cf., for example, Dreher as follows: "As Marill et al. in which the possibility of an unknown a matter of survival and social interaction, itis being a particular character is examined sequentially necessary for living systems constantly and con- (i.e., if the unknown alphanumeric does not consist currently to perform two types of event classifica- of essentially one segment whose length islittle, tion, termed by anthropologists 'etic' and `emit' more than the distance across the end points, and is distinctions. Under etic classification, similarities approximately vertical, it is not a one)." (Minneman, among events are cancelled out and attention is 1966, p. 86). focussed upon differences; in the emit classifica- 5.112"...We are now able to begin serious tion, differences are disregarded and any possible study of the most difficult problem facing the similarities are focussed upon." Dreher (1966, p. 9). project: The analysis of real-world three-dimensional 5.113a"Pattern recognition is only one aspect scenes." (Minsky, 1966, p. 12). of the much more fundamental problem of analysis "Programs have been developed to read selected and description of classes of patterns (or equiv- parts of the visual scene, analyse them for parts of alently, classes of pictures)." (Narasimhan, 1966, polygonal objects, and then transform them to p. 167). real-world coordinates. The present programs are "The problem of designing logics to recognize still rudimentary, and their extension is vital to the characters can be considered as requiring two main project." (Minsky, 1966, p. 16). efforts. The common starting point of all recognition "Our goal is to develop techniques of machine techniques is'feature extraction' operations on perception, motor control, and coordination that the character to determine the significant character- are applicable to performing real-world tasks of istics of the character. The second investigation object-recognitionandmanipulation."(Minsky, involves the problem of identification, that is, the 1966, p. 11). problem of classifying a character as a member "The sensory equipment includes two visual- of one of several classes." (Liu, 1964, p. 586). input devices: TVA, a vidicon television camera, "The essential problem of character recognition and the more precise TVB, an image-dissector is to find which of two patterns is the more similar deviceforcontrolled-scananalysis."(Minsky, to a third. When this problem has been solved, an 1966, p. 13). unknown pattern can simply be 'recognized' as Further, "processes for recognizing visual objects belonging to the recognition class to whose members by computer programs are being studied. These it is most similar." (Ullmann, 1966, p. 584). processes involve analysis of the visual field into 5.113b"Since pattern recognition isactually objects and background. Objects are to be recog- the operation of classification, the concept of classes nized by generating hypotheses am.; confirming of patterns plays a central role...Each class is them or modifying them by the results of selective represented by a region, or more generally by a dis- attention to parts of the field and of the proposed tribution function in some kind of pattern space." objects. These processes will entail use of stereopsis (Richardson, 1966, p. 10). and color, as well as the construction of abstract "Very popular in decision techniques and in symbolicthree-dimensionalrepresentationsof patternrecognitionisthe minimum Hamming the scene within computer memory." (Quarterly distance (MHD) method using binary feature set Progress Report, No. 80, Research Laboratory for {xi }. Here the categorizing is accomplished by Electronics, M.I.T., 195 (1966).) determination of the MHD between the applied "A picture, scene or view is read with the help of binary sets and all those ideal feature sets stored." an optical device and stored as an array of light (Kazmierczak and Steinbuch, 1963, p. 826). intensities in the memory of the computer. The "A more general adaptive procedure for deter- ultimate goal will be to understand this information, mining linearor piecewiselinear discriminant 139 36.1-107 0 - 70 - 10 functions for multiclass pattern classification is of variables and in eventual selection of a 'few' proposed. The adaptive procedure isa many- `good' ones...For this reason, the process of pattern or group-pattern adaptation. The training pattern recognition can be considered also as a sequence consists of groups of vectors inmatrix sort of information compression,because its purpose form instead of single vectors. The convergence is to ignore that information which is irrelevant proof shows that this procedure terminates in a to class features." (Watanbe et al., 1967, p. 192). finite number of adaptations if a solution exists. 5.117 Some of the results that have been re- A necessary and sufficient condition is developed ported deal with the likelihood functions of separ- for testing the linear separability of each subset able stochastic functions (Brick and Owen, 1967), of (d & 1) samples. Furthermore, the proposed with measures to select variables so that the prob- procedure can be implemented with the addition abilityof misclassification issystematically de- of only a little complexity to existing systems. Com- creased (Rao, 1966), and on algorithmic approaches puter simulations indicate satisfactory results." to the determination of decision boundaries based (Wee and Fu, 1968, p. 178, abstract). on learning samples. (Henrichon, 1967). 5.114"Therearemanypurelyhardware 5.118"The mathematical model used is the considerations that dictate the particular choice classic one of which so many of the statistical of measurement, transformation, or decision func- patternrecognitionstudieshave been based, tion to be used; certain measurement spaces may wherebyann-dimensiona.samplespaceis require much -nore digital storage than others." partitionedinto category regions with decision (King and Tunis. 1966, p. 66). boundaries...Assuming there exists a probability 5.115"Among the challenging problems in the distribution associated with each category describing design of pattern recognition systems, two problems the distribution of its members in n-space, the object are of utmost importance: (1) the extractionof is to partition the space in an optimal fashion. An pattern features, and (2) the optimum classification unknown vector is then assigned to the category in of pattern classes. The first is concerned with the whose regionitfalls." (Cooper, 1966, pp. 1-2). problem of what to measure, and the second deals "Adaptive pattern recognizers incorporate with the problems of making optimum decisions approximations to the class probability distributions in classification." (Tou and Heydorn, 1967, p. 57). that improve as a result of cumulative information "The essential point underlying the validity of derived from known inputs." (Barus, 1966, p. 385). the separation of the system into two parts, is the 5.118a"The reliability of a particular recog- requirement that the quantities [bi] produced by nition procedure can be described quantitatively the property filter define a -pattern space of modest by means of certain misclassification probabilities dimensionality in which the classes overlap to and cost functions. These, in turn, define a loss only a minor degree." (Richardson, 1966, p. 75). function which usually forms the base for choosing "To avoid inaccuracies in prediction caused by the classification regions. For example, the Bayes too many measurements, a method is given for procedure chooses RI,...R. so as to minimize selecting from the available measurements a subset the expected loss, while the minimax procedure that will usually give a more accurate prediction. minimizes the maximum expected loss." (Capon, Essentially, this technique consists in estimating 1965, p. 247). the mean square error associated with various A specific example is given by Allais as follows: numbers of measurements in a particular ordering. "Data for the first experiment were obtained from This selective approach to prediction is most profit- ahandwritten character reader, built by IBM ably applied to problems involving a relatively (Greanias, et al.). The problem treated is that of large number of measurements and a limited discriminating between the handwritten numerals number of learning samples. Two experiments one and zero. One hundred measurements were having these properties were presented to show chosen at random from 160 which were available, the power of the selectivepredictor."(Allais, and of these only 78 were nontrivial. Thus these 1966, p. 128). experiments were performed with q= 78 candidate 5.116"The theorem of the ugly duckling shows measurements. Each measurement was an integer- that if all the predicates which are logical functions valued number less than 20, and predictand y was of the observable attributes of objects are con- a binary indicator. Hence, the data are nonnormal. sidered as having equal importance, or weight, there Measurements were ordered by partial correlation cannot exist any recognizable classes of similar with theclass indicator y, using 100 learning objects. This is because, according to the theorem, samples. Recognition using various numbers of any pair of non-identical objects shares the same measurements was then tested on 284 independent number of predicates as any other pair of non- data samples. Minimum estimated error ep occurred identical objects. This means that if there are with p =29, which in this case resulted in zero recognizable classes of similar objects in our world recognition errors." (Allais, 1966, p. 128) of experience, some predicates are tacitly given "The main result of the foregoing analysis is the more weight than some others. This is the basic practical interpretation of an expression for the mathematical fact underlying the technical knowl- mean squareerror of themaximum-likelihood edge shared by everybody that the essence of predictor. Here, the fractional increase in error pattern recognition lie; in weighting or evaluation due to using a learning sample was found to be 140 474,.:V-.44.42--";,C.-4 44. 44-4

approximately p /(N -p -2), where p is the number The total amount of dust of all kinds over a given of measurements. This result explains why a pre- area of the 'floor' represents the probability that diction is sometimes more accurate if it is based a pattern lies in the area regardless of its class on fewer than the total number of available measure- membership. The relative thickness of the various ments." (Allais, 1966, p. 128). layers of dust then gives the probability of member- 5.118b"The notion of effectiveness of a recog- ship in each class." (Richardson, 1966, pp. 11-12). nition logic may be interpreted from an information- "The use of statistical decision functions is one theoretic point of view. Intuitively, one hopes to of the many possible approaches to the problem obtain effective recognition logics so that the ac- of character recognition by computer." (Chu, 1965, curacy of recognition may be enhanced, or the p. 213). Cf., also, Dressler and Werner as follows: number of logics required to achieve a given "Of the many differentpatternrecognition accuracy may be reduced." (Liu, 1964, p. 586). systems and devices utilizing various principles of 5.118c"Since the shapes of the Arabic numerals operation, this analysis is confined to those which and English alphabet were not designed to make employ a `forced-learning'__ preliminary procedure, the most efficient use of common features, itis and which base decisions upon application of the not possible to code the numerals and alphabet maximum-likelihood principle of statistical decision with six or seven features bits, as in the case of theory. In this type of system, the statistical data an ideal code. Prior studies have shown that sub- compiled from the forced-learning process are stantially more than seven features will be required stored in the memory of some associated computing perhaps as many as twenty or thirty." (Greanias, circuits and used as the basis for calculations of 1962, p. 145). posterior probabilities, and probability calculations 5.118d"Considering pattern recognition as a are made directly from Bayes' Rule. Within this Eattistical decision problem, the structure ofa recognition system can be derived from the functionalparticular procedural framework, there are, among form of the probability distributions. Successive other conceivable possibilities, two basic alterna- approximations to the distribution functions lead tives regarding the nature of the data utilized. to a hiemchy of recognition structures." (Chow (a) The simpler method, now in commonuse, and Liu, 1966, pp. 73-74). employs some fixed threshold, a singlequan- 5.118e"In general, the measurements are not tity, for each mask. If the excitation resulting statistically independent; there exists a certain, from the application of an unknown pattern though usually unknown, dependence among the to a mask exceeds this threshold, then the measurements. The central problem is to determine mask is said to be excited or to fire. The which dependence relations are worth examining probability scores necessary for the calcula- and how to weigh them." (Chow and Liu, 1966, tions are then only the relative frequencies of p. 74). excitation and nonexcitation for each pattern "For certain types of measurements, the chain class of interest... ordering or the dependence tree may be inferred from the physical or geometrical properties of the (b) Contrasted with this, a somewhat more com- measurements. For example, when these are the plicated procedure which could be used would direct video image of the pattern, then the assump- employ thecompleteprobabilitydensity tion of neighbor dependence such as described curves resulting when one entire class of by Chow [1962] is a reasonable choice." (Chow, patterns is presented to each mask, instead 1966, p. 108). of justa fixed threshold procedure. The 5.119"It has been recognized that the pattern statistical decision in this case is then based recognition problem can be discussed within the not merely upon the probability of a firing frameworkofstatisticalclassificationtheory." or a nonfiring, but rather upon a quantitative (Capon, 1965, p. 247). measure of the mask's response to an un- "We turn now to a second way of describing known, in terms of the known probability' classes, namely in terms of statistical distribu- density curve." (Dressler and Werner, 1964, tions...The only reasonable way to describe pp. 471-472). overlapping classes is statistical. In this kind of 5.120Similarly, in the area of speech recog- description, whether the pattern space be complete nition, "one approach of interest assumes the or incomplete, a given pattern is assigned a proba- existence of a perfect measurement X; for each class bility of lying in a given small region of pattern of utterance. It also assumes that the variety of space and this probability is further decomposed measurements is the result of the perfect measure- into a sum of terms each giving the probability of ments being corrupted by additive noise.. .. membership in a given class.... "If an operation on the measurements,X' =--g(X), "Another way, perhaps intuitively appealing, of could be discovered that would eliminate the effects looking at the statistical prescription isto regard of variation in speed of talking, volume, and ex- each class as a pile of dust restingon a multidi- pression between speakers, the problem of recog- mensional floor corresponding to thepattern space. nition of human speech might still be approached Each pile contains a particular kind of dust and dif- from the perfect signal point of view." (King and ferent piles may overlap, one layerover another. Tunis, 1966, pp. 66-67).

. 141 "Talbert, et al., and Dammann have reported on "Adherents of the statistical approach spend the use of adaptive linear decision functions in much time arguing among themselves as to whether speech recognition studies..." (King and Tunis, this or that statistical technique is more appropriate, 1966, p. 65). but those who have a chance to compare them often "Researchinpatternrecognitionmay be find that the difference in output between one characterized as a searchfor invariants. The technique and another is not appreciable." (Doyle, problem is to find attributes that all instances of a 1965, p. 18). given pattern have in common that instances of "All such methods involve two considerations. other patterns do not. The particular class of in-_ First, there is defined a measure of group-density or variants selectedwillultimately determine the of inter-group likeness. Examples of the latter type performance of the pattern recognition system. of measure (the so-called 'similarity coefficients') are However, performanceisnotthesolefactor legion; the best-known have been reviewed by influencing the choice of invariants. Associated Goodman and Kruskal (1954, 1959), Dagnelie (1960) with any research project are certain objectives and Sokal and Sneath (1963), but it is doubtful and goals. In pattern recognition, achievement of whether even these extensive collections are com- these goals is not only a function of what the system plete. For general consideration, suppose that two recognizes, but also how it performs the recognition. groups (i) and (j) fuse to form a group (k); then, Thus the selection of invariants may be more in- extending (and slightly altering) the symbolism of fluenced by the purposes and philosophy of the Williams, Lambert and Lance (1966), we shall need research than by a desire for a high rate of recog- to distinguish between three types of measure: nition. To illustrate this point, let us examine briefly (i)-measures, which define a property of a group, two character recognition schemes which are at (i,j)- measures, which define a resemblance or opposite ends of a spectrum with respect to the difference between two groups, and (i, j, k)-meas- aims of their research. Both perform very well in ures, which define some difference between the that their recognition accuracy is high. Their basic original two groups, considered jointly, and that difference in purpose and philosophy is reflected in formed by their fusion. Of these, (i)-measures are two very diverse sets of invariants. These two confined to clustering techniques except in so far schemes will provide a frame of reference for the as they may be incidentally required in the course research described inthispaper." (Teitelman, of calculation of (i,j, k)-measures." (Lance and 1964, pp. 559-560). Williams, 1967, p. 373). "What is needed, then, is a concise description, 5.122"Goodman and Kruskal (1959) have iden- invariant to size and position of the pattern which tified over 50 different formulas for measuring presents, in a direct manner, the segment associations. Each such formula hasitsown relationships. One would like, further, that any such advantages as well as its drawbacks, and, given description while not being rotationally invariant, ourpresentincompleteunderstandingof the be readily amenable to rotational manipulations. problem of semantic association, it would be pre- This last because many line patterns and letter mature to suggest any one as ideal." (Guiliano, sets, while not wholly rotationally invariant, are 1965, p. 27). largely so, and it would be desirable to be able to 5.123 An optical relization of "PAPA" (Italian conveniently probe this aspect of the recognition acronym for 'Automatic Programmer and Analyzer problem. Clearly, in order to avoid reference .to an of Probabilities') (1962) provided a filter for each arbitrary external point, a point in the pattern of the two classes a and b to select only those optical itself should be chosen as a reference. The pattern's fibers that give a preferential 'yes' answer for the centroid is a good choice." (Spinrad, 1965, p. 128). appropriateclasswith weightsproportionalto 5.121"Only this kind of identification of the their 'goodness' as discriminating criteria. problem[defining the objectives of associative "The first dichotomic separation is based on retrieval in mathematical terms] is likely to yield `the most widely fluctuating property', which is also productiveresultsand we may find...that in most cases the most evident or simple property distinctlydifferentassociation matrices arere- for the sense organs of the machine." (Gamba, quired, depending upon the problem which requires 1961, p. 147). solution.... "The inductive ability of PAPA is related to an Bryant alsoexpresses ". .surprise that so `intelligence' term, proportional tothe average much attention has been given to derivation of correlation factor of different patterns of the same measures of association and correlationin the class." (Gamba, 1962, p. 176). absence of mathematical work which specifies "After PAPA No. 2 has been instructed by that, under certain reasonable assumptions, these presenting toitan equal number of example are the measures needed." (Bryant,1964,pp. patterns of both classes, the tape is run backwards. 503-504). The read-out amplifiers pick up the "yes" (and "It is clear that carefully controlled experiments "no") pulses for each A-unit and check their to evaluate the efficacy and usefulness of the statis- values. If they are the same within a preassigned ticalassociation techniques have not yet been percentage, this means that the A-unit is inefficient. undertaken except in a few isolated instances.... In this case the corresponding noise...is auto-

142 matically erased and replaced by fresh noise. This people has been developed for the Air Force might give rise to another inefficient A-unit, but Systems Command's Aeronautical Systems Division since the process is such that the efficient A-units at Wright-Patterson Air Force Base, Ohio. are maintained, by repeating the process several "Called Conflex 1, the pattern-recognition system times the average efficiency of the A-units for the was built by Scope, Inc., of Falls Church, Va., in a given classes is considerably increased." (Palmieri 30-month program for the Air Force Avionics and Sanna, 1962, pp. 8-9). Laboratory at ASD. In the presently demonstrable model of Gamba's "Decision-making by machine is becominga `PAPA', Figure 18, 4,096 associational 'cells' are critical requirement in aerospace operations where used with 8-bit discriminatory capacity for each,so vast quantities of data must be collected, studied, that 64,000 bits of storageare =squired for this and a conclusion reached almost instantaneously. purpose, and up to 8 categories or :lasses can be "The device is capable of recognizing 4,800 differ- distinguished. It should be noted, ofcourse, that ent previously learned patterns, including pictorial size normalizations are not required in the system displays, letters, numbers, and geometric designs, and that depending upon the trainingsequence, with 99.6 per cent accuracy." (Ordnance Electronics, considerable tolerance to location in the input 212, Sept.-Oct. 1963). field and to rotation can be accommodated. "CONFLEX I was designed to serve as an experi- "The presentdemonstration equipment and mental CR [`conditioned-reflex'] system. For this programming allow a 32-sample training sequence reason, it was decided that the design should proper class. Output consists of the lighting up of an vide for several thousand D-cells [discrimination] indicatorsignal for the machine-selected class and multiple M-fields (classes)." (Uffelman, 1%2, and/or the printout of the numerical values of the p. 137). machine-calculatedprobabilitiesthatagiven "What is required, is a system that doesnot input pattern belongs to class 1, 2,...8. merely discriminate betweenone pattern and "Potentialapplicationinvestigations have in- another, but rather analyzes classes of patterns and cluded studies of knight vs. rook moves in chess, is capable of extracting the properties ofa class and the question of whether the Paschal lines for that is common to all its members.... random interconnections of lines drawn from 6 "Handwritten characters were used to train the points on the circumference of a circle are vertical device. The letters A, D, and Cwere considered. or horizontal. Perhaps more practically, the system Thirty examples of eachwere used for training. has been applied to meteorological chart discrimi- After training, all ninety of these exampleswere nations with respect to isobar or isotherm curve rerun through the system for recognition. There patterns, that are symptomatic, respectively, of were eight system errors using probabilistic sorting. good or bad succeeding weather in a specific Using unknowns (or letters not used in the training location. Here, given training sequences of 50 sequence) the probabilistic scoring techniquewas samples each for the two classes, the 'PAPA' compared to the Perceptron technique. Theerror machine system has achieved 80% valid recogni- rates were 14% for Perceptron and 20% for Probabil- tions," (Stevens, 1968, p. 29). istic scoring." (Hoffman, 1962, pp. 153, 156). To date, the same randomly-generated inter- 5.126"The GPO approach of Coombs and section-detection patterns have been used for the Milne has been applied to badly mutilated examples various tests. There is no reason, however, that of a 10-character multifont vacabulary. A technique other randomly-generated filtering patterns could of `shuggling' (shuffling or jittering)is used to not also be used, but this question has not been determine maximum and minimum correlata of the investigated in detail. On the other hand,at least input image with the same and other patterns in preliminaryconsiderationhasbeengivento the vocabulary. From these determinations, a some- problems of speech recognition, without alteration what irregular 20 X 10 matrix of criterial points of the basic approach. provided successful character discriminationof 5.124 The objective is, "givena set of patterns, 198 of 200 patterns of the training set." (Stevens, [to] determine a set of features, minimal in number, 1968, p. 5). such that each pattern can be formed by thesuper- 5.127"Much more substantial is the nonlinear position of a subset of these features." These nature of transformations realized by the A-units authorsconclude, however: "Unfortunately,we with many inputs. Such transformations, as has have not yet been able to find a reasonably brief been shown by Barus' work, given a great number of algorithm to solve the problem in this degree ofA-units, make it possible to store more detailed generality. However, we shall present two useful information about the variety of patterns utilized algorithms for -the case in which the featuresare during the teaching." (Kovalevsky, 1965, pp. 41-42). large compared to their mutual overlap. Results "The test alphabets were chosen onpurpose to be of experiments by digital-computer simulation will difficult to classify. Early success withan alphabet be used to show that these algorithmscan lead to of 'squares' and 'circles' (`O's') led to further tests valuable results, even when this restriction isnot with a more difficult alphabet, consisting of the met." (Block et al., 1964, p. 84). hand-printed letters '0', 'Q' and 'C'. Although these 5.125"A 'conditioned-reflex' machine which letters are distinguished by simple features, it must can learn to recognize photos of aircraft or even be remembered that we are assuming noa priori 143 1111111ENNWNV10Ropit

r t eNyro' *rot Imo 41,w U0.0 , r ; U lo 4" 04, 11111; , ,,,t,(4,,3 ",!!43s,a e : fa. ' . ' '; A A A, /..1.,,''' .r''' ...,'7, . ".

' ,,AiTA, s Z.(44 I 4 ';2. .01 'a'' "rtIlt

r-- - 111 tt .. i [Ito! i 1, ,1111

Q. ''' :;L'''''''' )' I /0 ',.... * knowledge of the alphabet on the part of the into subsets of similar samples, where similarity is machine or the designer." (Barus, 1962, p. 241). defined as the number of identical bits in the words "Members of a stored sample can also be used to equalling or exceeding a given threshold. Categories define subclasses, each sample member being are formed by taking, any sample, and 'remembering' treated as a subclass centroid. Newly encountered it as 'Mask i', taking another sample and either pattern members, instead of being averaged into ignoring it if similar to Mask i or remembering it as an evolving subclass... areused to evaluate the Mask ii, continuing until all samples have been representativeness of the stored sample, to assign processed. A "signature" is then developed for weights to each subclass, and to adjust parameters each sample consisting of binary indications as to of the assumed subclass distribution. With this which of the categories it does or does not belong to. method, subclasses (each represented by a member (p. 355). of the stored sample) can easily be discarded or new Further,Bonner'sproperty-findingalgorithm ones added as part of a learning program." (Barus, "looks for bits and logical combinations of bits 1966, p. 386). which are valuable in distinguishing the various 5.128"The Multivac has been trained to recog- types of samples in the category. It does this mostly nize and discriminate alphanumeric symbols, play in a nonexhaustive manner by proposing hypotheses simple games, differentiate between sounds (such of possibly valuable properties, then evaluating and as between those of a whale and a dolphin), and even classifyingthesepropertiesandcreating new to run a maze...The Multivac's ability to search hypotheses from the result." (p. 355). photographs for evidence of structures of a partic- 5.132"By a procedure using these principles, ular kind is being applied under a contract to NASA the program builds up properties of such complexity which calls for a 'study of the feasibility of extra- as: a thirty way AND OR-ed to a forty-way AND; terrestrial landmarks', which may lead to an auto- a four-way OR AND-ed to another, and other matic method of classifying the visual appearance properties of lesser complexity down to and includ- of other planets. The contract includes design of a ing individual bits." (Bonner, 1962, p. 356). system that will be trained to recognize a number of 5.132a"The problem of designing discriminant categories of landscapes by processing a library of functions for nonseparable pattern sets is a practical representative slides to be furnished by NASA. In one because in many practical problems the over- actual spaceflight missions, the system will observe lapping probability densities of the different pattern the surface of a planet being explored and classify categorieswill make nonseparable pattern sets the landscape into one of the known categories." inevitable. Current practice for pattern sets which ("Machines That Think?", 1966, pp. 6-7). See also are not completely separable is to use the adaptive Asendorf (1968). methods anyway, because they tendtoyield 5.129 "MINOS II has been successfully trained adequateclassifierseven though they cannot on patterns, converted by preprocessors into digital eliminate all classification errors. form, that represent such diverse subject areas as "Using linear programming in the design is easier handwritten graphical symbols, weather information, than using adaptive methods." (Smith, 1968, p. 367). visual or photographic objects, time-varying wave- "In general, the feature sets of different pattern forms, statistical data, and switching functions." classes are not linearly separable. It is doubtful (Munson et al., 1966, p. 360). whether in this case an approximation by complete "In the meteorological application, ISODATA is linear-decision functions is practical with regard to being used to aid in the analysis of 646 weather the required expense." (Kazmierczak and Steinbuch, records related to ceiling height at Washington, D.C. 1963, pp. 831-833). The goal in this case is the design of preprocessing 5.133For a comprehensive discussion of thres- for short-time (5 hours ahead) predictions of ceiling hold logic techniques, see the text by Hu (1965). We height." (Ball and Hall, 1965, p. 330). note also the following: 5.130"The input to CHILD consists of n analog "Many pattern classification systems use linear values, which can be thought of as an n-dimensional threshold elements to perform linear discriminations analog vector. This input vector can be derived on the patterns of interest. In such an approach, either directly from the sensors or from some pre- n properties of patterns are measured and each processing technique utilized to extract character- property is assigned a real number, Xi. The n values istics (or features) from the sensory pattern... (Xi. . Xn) are used as a descriptor of the It is CHILD's primary purpose, then, to determine pattern.... (1) which components of the input vector are "A linear threshold element (LTE) is a device important, (2) the range of acceptable values each which has n inputs (Xi,.. Xn) and a single out- component may assume, and (3) the degree of im- put, .f. Each Xi is a real number and the value off portanceto be assignedto each component." is determined by forming a weighted sum of the (Choisser and Sammon, 1963, p. 1). input quantities and comparing this sum S to a 5.131In Bonner's method (1962), samples of threshold T." (Mattson and Dammann, 1965, p. 294). members of the recognition vocabulary are repre- "A linear logic unit may be provided with a thres- sented as binary words for the presence absence of hold varying in time over the entire range of possible pre-specifiedproperties. The problemisthen input levels. A logic unit then turns on at a time simplified by breaking the total number of samples directly determined by the input level. Depending 145 on the nature ofthe weights assigned to the input enough to be representative. After a suitablenumber connections, a unit with n binary inputs mayhave of iterations, the process either converges, orelse from nt 1 to 2" distinct input levels. In thelatter the performance of the systems ceases toimprove caseachieved by assigning weights proportional ...Nonconvergence indicates either overlap of to powers of 2 to the connectionseach possible the distributions associated with certain orall pairs binary input state has a unique weighted signal of classes or else linearinseparability." (King level. Thus the time of response of a logic unit may and Tunis, 1966, p. 67). be made to specify exactly the input subpattern to "Two classes are said to be linearlyseparable if the unit. a hyperplane canbe placed between them. Many "The proper utilization of this unit in a large sys- self-organizing machine concepts assume thelinear tem requires that the outputweight associated with separability of classes. It is obvious thatif two a logic unit varysynchronously with the threshold. classes are convex and nonoverlappingthey are To accomplish this, a system clock may be used to linearly separable. However, if the classes are not control the threshold and to specify the output convex the question oflinear separability is harder weight." (Astropower Lab., 1964, p. 19k). to answer. However, in the caseof certain specific "The input weights and hence the logic unit models of classes, linear separability can beshown of the to be impobsible."(Riellardson, 1966, p. 11). hyperplanes were determined by the use "how well will discriminant analysis technique ...For compu- 5.135 Greenberg and Konheim ask tational reasons, each logic unit is permittedonly the separating hyperplane separate patternswhich are not included inthe training set but which are a small number of inputconnections (corresponding the same classes? This is a to nonzero coefficients of thehyperplanes). More to be separated into units are generated than are used in themachine. question involving the error rate on new patterns and one that can be answered only bysubsequent "The selection of logic units to be incorporated hand, since in the machine is done sequentially inorder to testing of the new data. On the other maximize the separation of pattern classes in the the linear functionals are continuous it isreasonable to assume that new patternsthat are close to old recognition space ...each unit being added to The limits of the machine being selected from apopulation of 1-,att.,,,:ns will_ be similarly classified. possible units by the criterion of maximum improve- such a form of `generalization'defierid Ivor, separations. This is well the linear functionals separate the classes. ment over existing interclass There is a well-defined and obvious senseby which accomplished by means of a loss function ...The we may rank thefunctionals which separate the procedure...concentrates on those patterns near the boundaties, assigning large weights tologic two classes A andB. However, the algorithm units active for these patterns, thuseffectively given...yields only one such separating func- increasing the separation of pattern classes. tional and the problem of finding the 'best' separa- "The use of linear logic thresholdelements in tion is a difficult unsolved problem."(1964, p. 306). Inc., ..has current pattern recognitiondevices involves two 5.136"RabinowElectronics, major difficulties: such devices are capableof only received a patent for an optical character recogni- ..The two outputs, and the assignmentof outputs to par- tion system featuring dictionary lookup . linear system checks doubtful words(indistinctly written) ticular inputs is subject to the restraints of by comparing characters against words in storage, separability." (Astropower Lab., 1964, pp. 192, 194). ACM "Threshold logic, a new and active area of switch- then selects the most likely word." (Commun. ing theory, has been primarilymotivated by an 9, 707 (1966). ` threshold "Successful machines of which we are aware, interest in certain computer devices called recognize individual characters and produce outputs gates'. The logical functions performedby such There functions,linear-input (usually clerical) identifying the characters. gates calledthreshold have been a smaller number of attempts torecognize functions, majority functions, etc., have been given of mathematical attention. A switching words, as opposed to the individual characters a great deal of which the word is composed. The S. F. ReedPatent function f of n arguments is a threshold function for a2, . anand a No. 2,905,927 discloses a method and apparatus when there exist weights al, recognizing words as such, and not theindividual threshold T (all real numbers) so that characters of the word. The V. K. Zworykin etal. Patent No. 2,616,983 and L. E. Flory et al., Patent 1 if E T, No. 2,615,992 disclose machines which arein the i=1 nature of word recognizers. TheZworykin and Flory .. xis= disclosuresrelateto equipment for translating 0 if E aixi< T." written data into sound to aid the blind.On the other hand, the Read patent is directly concerned with word recognition in data processing pursuits. (Winder, 1963, p. 108.). "An object of our invention is to provide arecogni- 44. individual characters .Iterative routines which operate on tion system for identifying the 5.134 . look-up a sequence of measurements,Xi, X2,.. X,, where of a word, but which relies on a dictionary the class associations are known apriori and the to ascertain the identityof the word if the reading number of measurements from each class islarge machine is incapable of identifying one or moreof 146 its characters. In principle, we recognize a word (and dictionary designedfor automatictranslation." hence obtain knowledge of all of the characters (Casey and Nagy, 1968, pp. 502-503). thereof) by identifying all of the characters of the 5.137"...Recognition performance can be word which the reading machine is capable of improved by using context. Most obviously, letters identifying, and then systematically interrogating and sounds make words, but even knowing digram a dictionary until we find a word or words which letter or syllable frequencies can improve decisions have all of the machine-read characters plus another about letters or phonemes." (David and Selfridge, character (or characters) in the place of the unknown 1962, p. 1094). character (orcharacters). The dictionary word 5.137a"The machine's recognition accuracy is (or words) are then fed to a utilization device such improved by using simple contextual constraints; as a buffer, printer, computer, etc.... letter digram frequencies are employed." (Harmon, "Another object of our invention is to provide 1962, p. 152). automatically operative means for correcting an 5.138"A good example illustrating the use of erroneously read character, or character reading context in a decision theoretically correct manner machine failure of any other type, where the intelli- is the improvement achieved in the performance gence criteria to determine the identity of the un- of a faulty character recognizer through the utiliza- known character are based on the characters which tionofletter digram or trigram frequencies." are correctly read by the machine. (Sebestyen and Edie, 1964, p. 9). "A broader objective of our invention is to provide 5.138a"When recognizing handwritten or highly an error correction system for a character or char- disturbed characters human beings make use of the acters of a word, regardless of the originating source, redundancy of the words and sentences, i.e., of i.e.., whether or not a reading machine is used. the context. This process can be simulated by iayer- "A novel aspect of our invention is the technique ing learning matrices and thus used for increasing for finding the word in the dictionary. One technique the recognizing ability of an automatic system." involves inserting 'trial' characters into the proper (Steinbuch and Piske, 1963, p. 858). position or positions of a group of identified char- 5.138b"As a preparatory step, a trigram of the acters for a dictionary look-up which continues Japanese phoneme sequence was examined, which until a word having the known characters plus a gave us the data to design the recognition system trial character is found...We have means to ...From the result about 1000 cases of three remember thecharacterswhich areproperly phoneme sequences cover 90 per cent of the data identified and also the positions of the character appearing in the conversational speech...The or characters which are not identified. With this basic principle of recognition is the matching of remembered information we insert 'trial' characters the analyzed pattern of input unknown speech one after the other into the position of the improperly sound with the stored standard patterns correspond- identified character to form 'trial' words. A 'trial' ing to the three phoneme sequences." (Sakai and character as used herein, is a character inserted in Doshita, 1963, p. 843). the space which would be occupied by an unknown 5.139"The specific effect on error reduction is character of a word. A 'trial word' is a word formed impressive. If a scanner given a 5% character error of known characters, i.e., those read by the machine rate, the trigram replacement technique can correct plus one or more trial characters. The trial words approximately 95% of these errors. The remaining are compared to the words in a dictionary for a error is thus ...0.25% overall." (Carlson, 1966, dictionary look-up. The entire dictionary may be p. 191). interrogated and all words made with trial char- 5.139a."One stores the (27)2 digram frequencies acters printed out, or we can stop the dictionary of the English language (space is a character here) look-up when the correct word is found in the and, in the case that a letter sequence is not found dictionary. The dictionary is preferably a high-speed in the dictionary (indicating that an incorrectly magnetic storage device, for example a magnetic spelled word has been detected), uses this table to drum, because it is fast and offers versatility not examine the left and right neighbors of each letter found (or at least as easily used) in other storage in an effort to find the letter least likely to be devices. It is entirely possible and practical to have correct ...The most likely change is then made, a drum store the entire English language in the based on the same digram data, and the dictionary practice of our invention." (Rabinow and Holt, searched again with this change. This process is 1966, p. 1). repeated...until a match is found." (Cornew, 5.136a"The experience gained in developing a 1968, pp. 79-80). recognition method independent of the particular 5.139b"In a sample of text several thousand geometrical configuration of the characters in the words long, the observed letter transition frequencies text has also been applied to other tasks. Portions may be expected to match the stored values quite of the clustering algorithm were used to group closely. For printed text in such quantity, it seems Chinese characters, to design ternary references for reasonable to base recognition on the relatively a commerical multifont reader, to derive Boolean invarianttransition frequencies, rather that' on recognition logic, and to classify the entries, on preconceptions of the structure of characters." the basis of a philological questionnaire, ina (Casey and Nagy, 1968, p. 492).

147 "Tentative identities are assigned to groups of appreciably less similar. Obviously, there is no similarcharacters.Theseidentitiesarethen placeforparadigmsinclustering problems." permuted until the text reads sensibly; good sense (Watanbe et al., 1967, p. 92). in this case means an admissible set of transition 5.140b"The purpose of Taxoinetric Analysis frequencies. Letter pair frequencies are actua'aly is to separate a population into distinct groups or used... clusters, each cluster being defined in terms of the "In solving a cryptogram, each type of character qualities or attributes which the ,nembers of the is labeled with the same (though incorrect) symbol cluster have in common. This is achieved by first to start with. The processor described here, however, computing a measure of association, a similarity must first perform the nontrivial task of determining coPfficient between each and every member of the which samples should be classed together by the population or operational taxonomic unit (0.T.U.). same symbol. The system achieves this by means The simplest coefficient might be defined as: of a series of cluster-seeking algorithms operating in the sample space." (Casey and Nagy,1968, p. 492). number of attributes shared 5.139c "We have been concerned...with the Stotal number of attributes." recognition of the identity of a word or of letter sequences by making use of the fact that only (Shepherd and Willmott, 1968, p. 57). certain letter sequences occur in English with any 5.141"Early workinclustering techniques appreciable probability.... assumed that the clusters were known a priori. "Since certain forms occur in certain contexts Fisher in 1938 made contributions in this area." while others do not, all the available information is (Mattson and Dammann, 1965, p. 296). t, not being used by current pattern recognizers." 5.142Parker-Rhodes says: "Broadly, members (Voss ler and Branston, 1964, p. D2.4 -1). of a clump must be more like each other, and less "The dictionary method has the disadvantage of like non-members, than elements of the universe requiring that any word to be recognized must be picked at random. We can put this more rigorously included in the dictionary. This requires a large in terms of inconnections (i.e., similarities between amount of storage space, as well as a fair amount membersofoneprofferedclump), andout- of time for calculation in a sequential computer, connections (similarities between members of the since the joint probability must be calculated for clumps and those of its complement); the former each word of the given length.... must be above average, relative to the number of "A system combining the dictionary method pairs of elements involved, while the latter must be [3737 words] with use of the digram method for below average." words detected as not occurring in the dictionary, "The analogue of error, or uncertainty of the data was tested on text from newspaper articles and from [in statistical theory], is that [in classification theory] a psychology book. With this combined system a we do not really know withcertainty whethr reduction of about 45% in the number of letter we do not really know with certainty whether or errors was obtained, with an original garbling of 20%. not a given element of the universe belongs to a "A character recognizer performing at the 80% given class or to its complement....the statistical recognition level might be expected to improve to a expression of this is to assign a probability to its recognition performance of 95% or 96% or slightly class membership, assigning thus a vector of real higher through the use of an extensive word diction- numbers to each element....Instead, we conceive ary." (Voss ler and Branston, 1964, pp. D 2.4-1, 7, 8). of our classes as marked off by partitions of the 5.140"Closely related to the use of molecular universe; the uncertainties attaching to them are formula in the recognition process is the use of then represented as partitions of the universe into chemical context to assist in recognition...For three parts, the 'in% the `out', and the `doubtful'." example, hydrogen links with only one other atom, (Parker-Rhodes, 1961, pp. 7-9). whereas nitrogen links with three, four, or five Needham acids: "We suppose that the data are other atoms. Thus, 'H' is not likely to be mistaken presented as a series of objects, to each of which is for 'N'." (Cossum et al., 1964, p. 271). attached a string of properties. We then have two 5.140a"Classification problems may be divided problems: into two categories: pattern recognition and clustering. In pattern recognition, the classifier knows the (1)to find a suitable definition of a 'class' or names of classes and some samples (called here- group which is stated in terms of something after, paradigms) of each class. After the training computable from the data; period is over, during which the paradigmsare (2)to find a practicable algorithm for discovering shown, the classifier is required to placenew objects the classes in a given body of data." (Need- of unknown class affiliation into the classes by a sort ham, 1963, pp. 4-5). of inductive generalization based on the known 5.142aSee also: "Automatic clustering tech- paradigms. In clustering, the classifier is given a niques for subclass determination reported by collection of objects with known propertier and is Bonner and Firschein and Fischler may be adopted required to generate classes in such a way that to subdivide some classes of characters that have objects within each class are strongly similar to large variations among the fonts in consideration." one another, while objects of different classes are (Liu, 1964, p. 593). 148 I 5.143"Dee F. Green, a research associate at the of word or document clusters related according to University of Arkansas Museum...[has devel- some given criterion; for the ranking of documents, oped a code] for reducing the individual attributes obtained in answer to a search request, in order of of some 4,000-odd pottery vessels to a numerical relevance; and for the detection of synonyms by system for computer handling. Once the material comparing association lists attached to the index is classified, the various attributes will be sorted terms." (Salton, 1963, p. 53). into discrete categories and then statistical tech- 5.146Further ".. the requirement that algo- niques applied to lump the attributes into statisti- rithms for automatic classification should bepo- cally meaningful groups, or ceramic types." (Bowles, tentially capable of applicationto very large object 1965, p. 271). universes prompts investigation ofnew procedures Li. The computer is of use in studies of shards that may prove more efficient when appliedon a or fragments of artifacts . In this connection, large scale or mayprove more efficient under certain Jesse D. Jennings of the University of Utah suggests conditions." (Senechalle, 1964,p. 1-1). constructing a matrix of coefficients of similarity of 5.147"A few examples of pattern recognition one artifact to another, and thus to all others within problems that might be handled bya machine of the a given corpus of objects." (Bowles, 1965, p. 270). type envisioned will now be mentioned. The first 5.144"Feurenures of conn,ction (similarity) class uf pattcins ihaicuries to mind is an alphabet between pairs of objects may be computed...The of hand-printed characters drawnon a binary grid, first three assume a binary incidence matrix . the components of the primarymeasurement vector The fourth connection definition is used in cases co corresponding to the cells of the grid. Another where definition is used in cases where the initial problem isthe classification of organisms into incidence data consistsof nonbinary attribute species, where the components ofco might represent values... the presence of strength of certain attributes.The "Each connection matrix is compacted, in the problem of medical diagnosis is closely related, the sense that only nonzero elements of the full n x n attributes becoming symptoms and results ofcer- matrix are recorded." (Dale et al., 1964, p. 9). tain diagnostic tests. Library classificationcan be We note further that ". .An alternate tech- considered a pattern recognition problem, in which nique is suggested in which the properties rather the components ofco might denote the presence or than the objects of the incidence matrix are classi- absence of certain key words in titles,subheadings, fied. The properties are regarded as a second uni- abstracts, and so forth. Aset of waveforms, cor- verse set 'U', and the connections formalize simi- rupted by noise and distortion, but normalizedin larity between properties through reference to their time and band-limited,can be considered a set of coincidence among objects. Clumping in `U' will patterns. Sampling at the Nyquist intervals would accordingly associate properties thatare similar in provide the co-components. Other possibleexamples their properties of incidence among the objects ofinclude recognition of spoken syllables,identifica- U." (Dale et al., 1964, p. 5). tion of languages from spoken samples, identification 5.144a"Rabinow has recently obtained a patent of musical instruments from theirtone qualities, and covering the general concept of an autonomous prediction of weather, economic trends andthe reading machine. He envisages combiningan ele- like." (Barus, 1962,pp. 228-229). mentary clustering procedure with dictionary look- "Pattern recognition should ultimatelyprovide up techniques." (Casey and Nagy, 1968, p. 493). powerful techniques foruse in such research areas 5.145"Word and document associations haNe as form perception, target recognition, language been used variously for the alterationor extension translation, theorem proving,game playing, and the of a set of index terms attached to a given document, testing of psychological models." (Pratherand Uhr, by addition of associated terms; for the generation 1964, p. D2.2-1).

6. Conclusions

6.1For example, " 'Recognize' is used to con- material, which it then displayson a screen for the note the act of identifying an incoming pattern with operator's attention (a far smaller display than that one of a known set of patterns. 'Classify', is meant necessary for an entire page would be sufficient to describe the separation of the incoming patterns here). The operator then simply indicatesto the into several sets according to similarities between scanner what action to take: to scan the material, the patterns which are inferred by the device." as before, in a facsimile mode; to resort to curve (Meyer et al., 1962, p. 181). following; to summon its arsenal for italics, boldface, 6.2"Another area which would bear investiga- or superscripts; or to let the operator key in the tion is that of higher level interaction between the offending word or letter. Aside from the intrinsic control computer and the operator to permiteven economies which may be realizedon some classes greater variability in the input documents and flexi- of character recognition applications, theexperience bility in the output. The scanner could proceed gained here may be useful in other man-machine autonomouslyuntilitencounters'unreadable' interaction situations." (Nagy, 1968,p. 487). 149 6.3"In the case of Russion language mate- of assigning descriptors to schematic representa- rials, in which we have a great deal of interest, we tions of electrical circuits." (Wooster, 1963, p. 156), have found that an optical character reading device but that "much less work has been done on good which would be able to accept just the Russian programming systems for 'picture languages',l'than scientific material, would require the unambiguous has been done on the techniques and media for identification of about 800 characters and symbols. picture production." (Wigington, 1966, p. 87), and This would give us an acceptable input capability that "a picture language is two-dimensional, and as for automatic translation. These 800 characters and yet we have no general method of formalizing its symbols are found in four fonts of the Cyrillic alpha- syntax. A number of investigators are working on bet, upper and lower case; the Greek alphabet; the the problem, but to date useful results are not Latin alphabet; the digits; subscripts and super- available." (Sutherland, 1967, p. 29). scripts; the mathematical symbols; and all of the 6.6"It has been pointed out that there are normal symbols that are used in the expression of theoretical problems; phonemic theory does not scientific discourse." (Howerton, 1962, p. 349). "On the rnarhinP print reariPr, a ifThnieni require- make it possible to set up unique rules according to which the acoustic patterns of an utterance may ment will probably exist to read scientific Chinese yet be correlated with the phonetic elements of the publications. The RADC technical approachis message conveyed by the utterance (Fant and Ris- based on evolutionary research and development." berg, 1962). One of the most serious problems is (Shiner, 1962, p. 337). deciding how to use the information that a human "The first step involves research and develop- listenerobtains from the context (Garvin and ment on a semiautomatic technique to determine Bertram, 1964). Current work is concerned with optimum coding and logiccharacteristics. The analysis in greater detail of the various formants second step then involves the development of an involved in the speech signal, and with investi- optical character recognition technique which will gating what elements of the signal are significant completely mechanize the digital coding of the for speech production and recognition." (Spo 'sky, Chinese ideographs for rapid processing with auto- 1966, p. 494). matic language translation equipment such as the 6.7"In the area .A pattern description, there USAF AN/GSQ-16 language translation machine." is as yet no general theory, and few properties of (Shiner, 1962, p. 337). measurement design algorithms can be adequately 6.4"In foreign language materials we find that there is a tendency to avoid the use of such generalized from one application to another. The things as italics or underlining for emphasis, but problem of deriving efficient sets of measurements rather the foreigners spread the letters of a word needs a proper formulation...." (Chow, 1966, apart to draw attention. If you teach a machine to p. 102). recognize a white space between two characters as 6.8In addition to patents and patent drawings being the boundary between the two words, how do as such, we note the following suggestion: "The you also make that machine recognize that when it large volume of trademarks registered with the sees white barriers between letters, it may be the patent office has prevented any systematic search prior to adopting one's own mark...The computer case of a word having been expanded for emphasis?" (Howerton, 1962, p. 349). here presents an excellent method for prescreen- ing...Advanced input devices to computers will 6.5This is an approach suggested by Kirkh accept an overlay composed of a halftone of the et al. (1964), and earlier. trademark in question...The computer will then Work on picture-language-equivalence inference search its memory for a trademark of similar pattern systems has also proceeded elsewhere, notably in quality. If one is found, it will be reprinted on a experimental investigations reported by Schwartz return screen immediately. Once the basic composi- et al. (1965) and by Narasimhan (1966). tion of a trademark is found to be similar to another, We note further that "such picture processing differentiating colors may then be checked out." techniques may also be able to handle the problem (Sims, 1966, p. 26). Appendix B.Bibliography

Abrahams, S. C., Automation in X-Ray Crystallography, Chem. Bacon, F. R., N. C. Churchill, C. J. Lucas, D. K. Maxfield and Eng. News 41, 108-116 (1963). C. J. Orwant, Application of a Telereference System to Di- Abramson, N. M., W. E. Dickinson and F. B. Wood, The Applica- visional Library Card Catalogs: A Feasibility Analysis, Final tion of Decision Theory to Voice Recognition Machines with Rept., 91 p. (Engineering Research Inst., Michigan Univ., Ann an Illustrative Example, IBM Tech, Rept. 16.01.071.003 (IBM Arbor, May 1958). Corp., Endicott, N.Y., Dec. 1959). Baker, J. D., From the Diet of Worms to the Bucket of Worms: Aicher, J. R., Producing Piping Isometric Drawings via Com- A Protest Concerning Existing Display Dogma for Information plier and Pk. ter, The Computer Bull. 11, 134-138 (Sept. Systems, in Second Cong. on the Information System Sciences, 1967). held at The Homestead, Hot Springs, Va., Nov. 1964, Ed. J. Alden, W. L., Cutting Communication Costs with Facsimile, Spiegel and D. E. Walker, pp. 429-433 (Spartan Books, Wash- Data Proc. Mag. 6, 11-14 (Sept. 1964). ington, D.C., 1965). Allais, D. C., The Problem of Too Many Measurements in Pat- Balkovic, M. D., Applications of Data Transmission Systems, in tern Recognition and Prediction, 1966 IEEE Int. Cony. Rec., Law Enforcement Science and Technology, Vol. 1, Proc. First Pt. 7, pp. 124-130. National Symp. on Law Enforcement Science and Technology, Allen, R. J. and E. Niehenke, The Cool World of Components, Chicago, HI., March 1967, Ed. S. A. Yefsky, pp. 153-157 Electronics 37, 75-79 (June 15, 1964). (Thompson Book Co., Washington, D.C., 1967). Alt, F., Digital Pattern Recognition by Moments, in Optical Ball, G. H. and D. J. Hall, ISODATA A Self-Organizing Com- Character Recognition, Ed. G. L. Fischer, Jr., et al, pp. 153- puter Program for the Design of Pattern Recognition Preproc- 179 (Spartan Books, Washington, D.C., 1962). essing, in Information Processing 1965, Proc. IFIP Congress Amann, C. and M. Klerer, Flexowriter/DIOA System, Rept. No. 65, Vol. 2, New York, N.Y., May 24-29, 1965, Ed. W. A. 124, 59 p. (Hudson Labs. of Columbia Univ., Dobbs Ferry, Kalenich, pp. 329-330 (Spartan Books, Washington, D.C., N.Y., June 1966). 1966). American Institute of Physics Staff, Techniques for Publication Ball, G. H. and D. J. Hall, A Clustering Technique for Sum- and Distribution of Information, in Annual Review of Informa- marizing Multivariate Data, Behay. Sci. 12, 153-155 (Mar. tion Science and Technology, Vol. 2, Ed. C. A. Cuadra, pp. 339- 1967). 384 (Wiley, New York, 1967). Baran, P. and G. Estrin, An Adaptive Character Reader, 1960 Andrews, M. C., Multifont Print Recognition, in Optical Char- IRE WESCON Cony. Rec., Vol. 4, Pt. 4, pp. 29-41. acter Recognition, Ed. G. L. Fischer, Jr., et al., pp. 287-304 Barnes, R. F., Language Problems Posed by Heavily Structured (Spartan Books, Baltimore, Md., 1962). Data, Commun. ACM 5, 28-34 Gan. 1962). Andrews, M. C., On Communications and Data Processing: A Baruch, J. J., Information System Applications, in Annual Re- Foreword, IBM J. Res. & Dev. 9, No. 4, 226-227 (July 1965). view of Inf-mmation Science and Technology, Vol. 1, Ed. C. A. Annual Report 1966, Institut ftir Nachrichtenverarbeitung und Cuadra, pp. 255-271 (Interscience Pub., New York, 1966). Nachrichteniibertragung der Technischen Hochschule Karls- Baruch, J. J., A Medical Information System: Some General ruhe, Karlsruhe. Federal Republic of Germany, Feb. 1967, 53 p. Observations, in Information System Science and Technology, Annual Report 1967, Institut ftir Nachrichtenverarbeitung und papers prepared for the Third Cong., scheduled for Nov. 21-22, Nachrichteniibertragung der Technischen Hochschule Karls- 1966, Ed. D. E. Walker, pp. 145-150 (Thompson Book Co., ruhe, Karlsruhe, Federal Republic of Germany, Dec. 1967, 65 p. Washington, D.C., 1967). Aron, J. D., Real-Time Systems in Perspective, IBM Sys. J. 6, Barus, C., A Scheme for Recognizing Patterns from an Unspeci- 49-67 (1967). fied Class, in Optical Character Recognition, Ed. G. L. Fischer, Jr., et al., pp. 227-248 (Spartan Books, Baltimore, Md. 1962). Asendorf, R. H., The Remote Reconnaissance of Extraterrestrial Barus, C., An Easily Mechanized Scheme for an Adaptive Pat- Environments, in Pictorial Pattern Recognition, Proc. Symp. tern Recognizer, IEEE Trans. Electron. Computers EC I5 , on Automatic Photointerpretation, Washington, D.C., May 385-387 (June 1966). 31June 2, 1967, Ed. G. C. Cheng et al., pp. 223-238 (Thomp- Bauer, W. F., On-Line SystemsTheir Characteristics and son Book Co., Washington, D.C., 1968). Motivations, in On-Line Computing Systems, Proc. Symp., Astrahan, M. M., The Role of Large Memories in Scientific Com- Los Angeles, Calif., Feb. 2-4, 1965, Ed. E. Burgess, pp. 14-24 munications, IBM J. Res. & Dev. 2, No. 4, 310-313 (Oct. 1958). (American Data Processing, Inc., Detroit, Mich., 1965). Astropower Lab., Douglas Aircraft Co., Adaptive Techniques as Berkeley, E. C., Translating Spoken English Into Written Words, Applied to Textual Data Retrieval, Rept. No. RADCTDR- Computers & Automation 5, No. 3, 9 (Mar. 1956). 64-206, Final Rept. on Contract AF 30(602)3134, 219 p. Bernstein, M. I. and T. G. Williams, An Interactive Programming (Douglas Aircraft Co., Newport Beach, Calif., Aug. 1964). System for the Casual User, Computers and Automation 17, Atal, B. S. and M. R. Schroeder, Predictive Coding of Speech 26-29 (Feb. 1968). Signals, 1968 WESCON Technical Papers, 8/2, Aug. 1968, 7 p. Bertram, S., Application of Hybrid Analog and Digital Tech- Auerbach Corporation, Source Data Automation, Final Rept. niques in the Automatic Map Compilation System, AFIPS 1392-200TR-1, 1 v. (Philadelphia, Pa., Jan. 16, 1967). Proc. Spring Joint Computer Conf., Vol. 23, Detroit, Mich., Avruhk, M. L., Automatic Text Reading, in Kompleksnaya May 1963, pp. 105-112 (Spartan Books, Baltimore, Md., 1963). Mekhanizatsiya iAvtomatizatsiya Protsessov Obrabotki, Poiska, Berul, L. H., Survey of IS & R Equipment, Datamation 14, No. Vydachi i Peredachi na Rasstoyanie, Moscow, 1966, pp. 357- 3, 27-32 (Mar. 1968). 369, as translated and excerpted by P. L. Stephen, in Soviet Bessinger, J. B., Jr., S. M. Parrish and H. F. Arader, Eds., Proc. Cybernetics: Recent News Items No. 3, Ed. W. B. Holland. Data Processing Conf., Sept. 8-11, 1964, 329 p. (IBM Corp., pp. 17-24 (The RAND Corp., Santa Monica, Calif., Apr. 1967). White Plains, N.Y., 1964). 1967). Bhimani, B. V., R. D. Merrill, R. P. Mitchell and M. R. Stark, An Axner, D. H., Data Communications the Digitronics Way, Data Approach to Speech Synthesis and Recognition on a Digital Proc. Mag. 9, 44-48 (May 1967). Computer, in Proc. 21st National Conf., ACM, Los Angeles, Axner, D. H., The Facts About Facsimile, Data Proc. Mag. 10, Calif., Aug. 30Sept. 1, 1966, pp. 275-296 (Thompson Book 42-44, 46, 48-49, 52-53 (May 1968). Co., Washington, D.C., 1966). 151 Bishop, A. B., Adaptive Passive Recognition, IEEE WESCON Brown. G. W., J. G. Miller and T. A. Keenan, Eds., EDUNET - Cony. Rec., Pt. 4, 1963. Report of the Summer Study on Information Networks Con- Bledsoe, W. W., Some Results on Multicategory Pattern Rec- ductedbytheInteruniversityCommunicationsCouncil ognition, J. ACM 13, No. 2, 304-316 (Apr. 1966). (EDUCOM), 440 p. (Wiley. New York, 1967). Bledsoe. W. W. and I. Browning. Pattern Recognition and Read- Brom), R. NI.. An Experimental Study of an On Line Man-Com- ing by Machine, Proc. Eastern Joint Computer Conf., Vol. 16, puter System. IEEE Trans. Electron. Computers EC-14, 82- Boston. Mass., Dec. 1-3, 1959, pp. 225-232 (Eastern Joint 85 (1965). Computer Conf., 1959). Bryant, E. C. Redirection of Research into Associative Retrieval, Bliss. W. H., Advancements in the Facsimile Art During 1965. in Parameters of Information Science, Proc. Am. Doc. Inst., 1966 IEEE Int. Cony. Rec., Pt. 1, pp. 2-9. Annual Meeting, Vol, 1, Philadelphia, Pa:, Oct..5-891964, pp. Block. H. D., B. W. Knight, Jr.. and F. Rosenblatt. Analysis of A 503-505 (Spartan 8(mks, Washington, D.C., 1964). Four Layer Series-Coupled Perceptron. in collected technical Buck, C. P., R. F. Pray, III and G. W. Walsh, Investigation and papers by Block, Kessler, Knight and Rosenblatt, Vol. 1. pp. Study of Graphic-Semantic Composing Techniques, Rept. No. 49-73 (Cognitive Systems Research Program, Cornell Univ.. RADC-TR-61-58, Final Rept. Contract AF 30(602)2091, 1 v. Ithaca, N.Y., Jan. 1, 1961). (Syracuse Univ., Research Inst.. June 1961). Block, H. D., N. J. Nilsson and R. 0. Duda. Determination and Buckland, L. F., Machine Recording of Textual Information Detection of Features in Patterns, in Computer and Informa- During the Publication of Scientific Journals, 70 p. (Inforonics, tion Sciences, collected papers on learning, and adaptation and Inc., Maynard, Mass., May 30, 1965). control in information systems, Ed. J. T. Tou and R. H. Wilcox. Burger, J. B., High Speed Display of Chemical Nomenclature, pp. 75-110 (Spartan Books, Washington, D.C., 1964). Molecular Formula and Structural Diagram, Rept. No. C105- Blokh, E. L., The Question of the Minimum Description, Radio R-4, 19 p. (General Electric Co., Huntsville, Ala., Dec. 31, Engr. 15, No. 2, 15-24 (Aug. 1960). Translated from the Rus- 1964). sian Journal Radio-teknnia 15, No. 2, 10-14 (Feb. 1960). Burgess, E., Ed., On-Line Computing Systems, Proc. Symp. Blondeau, P., The IBM Computer in the Graphic Arts, in Re- sponsored by the Univ. of California, Los Angeles, and Infor- search and Engineering Council of the Graphic Arts Industry, matics, Inc., Los Angeles, Calif., Feb. 2-4, 1965, 152 p. Proc. 13th Annual Conf., Chicago, Ill., May 20-22, 1963,pp. (American Data Processing, Inc., Detroit, Mich., 1965). 84-91 (Washington, D.C., 1963). Business Forms International, Inc., The Optical Character Rec- Bobrow, R. J., P. A. Gorman, J. B. Calataynd, S. Abraham, A. L. ognition Report, Third Printing, 48 p. (Fort Lauderdale, Fla., Weihrer and C. A. Caceres, Accuracy of Electrocardiographic Mar. 1967). Measurements by Computer, Am. J. Med. Elec. 4, No. 3, 121- Butler, J. W., Automatic Analysis of Bone Autoradiographs, in 126 (1965). Pictorial Pattern Recognition, Proc. Symp.on Automatic Bock, P., Remote Sensing and Data Relay in Hydrology and Photointerpretation, Washington, D.C., May 31-June 2, 1967, Water Resources, abstract, WESCON 68 Program, Western Ed. G. C. Cheng et al., pp. 75-85 (Thompson Book Co., Wash- Electronic Show and Convention, Los Angeles, Calif., Aug. ington, D.C., 1968). 20-23, 1968, p. 46. Boggiss, L-W.; Can-aria Land Pfin Uses IBM Unit, Electronic News, March 13, 1967, p. 64. Cameron, S. H., D. Ewing and M. Liveright, DIALOG: A Con- Bomba, J. S., Alpha-Numeric-Character Recognition Using Local versational Programming System with a Graphical Orientation, Operations, Proc. Eastern Joint Computer Conf., Vol. 16, Commun. ACM 10, 349 -357 (June 1967). Boston. Mass., Dec. 1-3, 1959, pp. 218-224 (Pub. by Eastern Cantrell, J. L., Infrared Geology, Photogrammetric Information Joint Computer Conf., 1959). 30, 916-922, 941 (Nov. 1964). Bonner, R. E., A "Logical Pattern" Recognition Program, IBM Capon, J., Hilbert Space Methods for Detection Theory and J. Res. & Dev. 6, No. 3, 353-360 (July 1962). Pattern Recognition, IEEE Trans. Inf. Theory IT-11, 247-259 Bonner, R. E., On Some Clustering Techniques, IBM J. Res. & (Apr. 1965). Dev. 8, No. 19 22-32 (Jan. 1964). Carlson, G., Techniques for Replacing Characters That Are Bowers, D. M., W. T. Lennon, Jr., W. F. Jordan, Jr., and D. G. Garbled on Input, AFIPS Proc. Spring Joint Computer Conf., Benson, TELLERTRON- A Real-Time Updating and Trans- Vol. 28, Boston, Mass., April 1966, pp. 189-192 (Spartan Books, action Processing System for Savings Bank, 1962, IRE Int. Washington, D.C., /966). Cony. Rec., Pt. 4, pp. 101-113. Casey, R. G. and G. Nagy, An Autonomous Reading Machine, Bowie, R. M., The Information Science and Industry Fifty Years IEEE Trans. Computers C-17, 492-503 (May 1968). Hence, Proc. IRE 50, 609-610 (1962). Cassotta, L., S. Feldstein and J. Jaffe, AVTA: A Device for Auto- Bowles, E. A., The Role of the Computer in Humanistic Scholar- matic Vocal Transaction Analysis, J. Experimental Analysis ship, AFIPS Proc. Fall Joint Computer 'Conf., Vol. 27, Pt. 1, Behavior 4, 99-104 (1964). Las Vegas, Nev., Nov. 30-Dec. 1, 1965, pp. 269 -276 (Spartan Census Bureau Slashes Record Handling Time, Systems 5, No. Books, Washington, D.C., 1965). 1, 32-33 (Jan.-Feb. 1964). Bozman, W. R., Computer Typesetting Program at NBS, in Elec- Chen, C. H., A Note on Sequential Decision Approach to Pattern tronic Composition in Printing, Proc. Symp., Gaithersburg, Recognition and Machine Learning, Inf. & Control 9, 549 -562 Md., June 15-16, 1967, NBS Special Pub. 295, Ed. R. W. Lee (Dec. 1966). and R. W. Worral, pp. 98-106 (U.S. Government Printing Cheng, G. C., R. S. Ledley, D. K. Pollock and A. Rosenfeld, Eds., Office, Washington, D.C., Feb. 1968). Pictorial Pattern Recognition, Proc. Symp. on Automatic Breuning, S. M. and T. H. Harvey, Implications of Integrated Photointerpretation, Washington, D.C., May 31-June 2, 1967, Transportation Engineering Design Systems for Computer Soft- 521 p. (Thompson Book Co., Washington, D.C., 1968). ware and Hardware Development, Proc. 22nd National Conf., Cherry, C., Ed., Information Theory, papers read at Third Symp., ACM, Washingt(m, D.C., Aug. 29-31, 1967, pp. 255-262 London. Sept. 12-16, 1955, 401 p. (Academic Press, New York, (Thompson Book Co., Washington, D.C., 1967). 1956). Brick, D. B., Sylvania Electric Products, session on optical page Cherry, C., Ed., Information Theory, papers read at Fourth readers, in Research and Engineering Council of the Graphic Symp., held at the Royal Institution, London, Aug. 20-Sept. Arts Industry, Inc., 14th Annual Conf., Rochester, N.Y., May 2, 1960, 476 p. (Butterworths, Washington, D.C., 1961). 18-20, 1964, pp. 85-89 (Washington, D.C., 1964). Cherry, C., M. H. Kubba, D. E. Pearson and M. P. Barton, An Brick, D. B. and J. Owen, Pattern Recognition of Stochastic Experimental Study of the Possible Bandwidth Compression Processes, Rept. No. 9, 37 p. (Information Research Asso- of Visual Image Signals, Proc. IEEE 51, 1507-1517 (1963). ciates, Inc., Lexington, Mass., June 23, 1967). Choisser, J. P. and J. W. Sammon, Jr., A New Concept in Arti- Brown, S. C., A Bibliographic Search by Computer, Physics ficial Intelligence, Rept. No. RADC-TDR-63-228, 6 p. (In- Today 19, No. 5, 59-61, 63-64 (May 1966). formation Processing Lab., Rome Air Development Center, Brill, E. L., Character Recognition via Fourier Descriptors, 1968 Griffiss Air Force Base, N.Y., May 1963). WESCON Technical Papers, 25/3, Aug. 1968, 10p. Chow, C. K., An Optimum Character Recognition System Using British Computer Society, The, Character Recognition, 195p. Decision Functions, IRE Trans. Electron. Computers EC-6, (The Gresham Press, Surrey, England, 1967). No. 4, 247-254 (Dec. 1957).

152 Crane, H. B., Character Reading Circuit, U.S. Pat. 3,145,367, Chow, C. K., A Class of Nonlinear Recognition Procedures, Patented Aug. 18, 1964. IEEE Trans. Systems Sci. & Cybernetics SSC- 2, 101-109 Crooks, E., Facsimile Systems as an Aid in Research, Am. Doc. (Dec. 1966). Chow, C. K. and C. N. Liu, An Approach to Structure Adaptation 7, No. 1, 40-43 (Jan. 1956). & Cyber- Cuadra, C. A., A Feasibility Study for Automated Fingerprint in Pattern Recognition, IEEE Trans. Systems Sci. Identification, Tech. Memo. TM-3007, 64 p. (System Develop- netics SSC-2, 73-80 (Dec. 1966). Chu, J. T., Optimal Decision Functions for Computer Character ment Corp., Santa Monica, Calif., 1966). Recognition, J. ACM 12, No. 2, 213-226 (April 1965). Cuadra, C. A., Ed., Annual Review of Information Science and Chu, J. T. and J. C. Chet), Error Probability in Decision Func- Technology, Voi. 1, 389 p. (Interscience Pub., New York, 1966). tions for Character Recognition, J. ACM 14, No. 2, 273-280 Cuadra, C. A., Ed., Annual Review of Information Science and Technology, Vol. 2, 484 p. (Interscience Pub., New York, 1967). (April 1967). Clark, W. A. and B. G. Farley, Generalization of Pattern Rec- ognition in a Se;f-Organizing System, Proc. Western Joint Dale, A. G., N. Dale and E. D. Pendergraft, A Programming Sys- Computer Conf., Vol. 7, Los Angeles, Calif., Mar. 1-3, 1955, Applications in Lin- pp. 86-91 (Institute of Radio Engineers,New York, 1955). tem for Automatic Classification With Clayden, D. 0., M. B. Clowes and J. R. Parks,Letter Recog- guistic and Information Retrieval Research, Rept. No. LRC- Text, Inf. Control 9, 64, WTM-4, 19 p. (Linguistics Research Center, Texas Univ., nition and the Segmentation of Running Austin, Oct. 1964). 246-264 (June 1966). Dammann, J. E., Application of Adaptive Threshold Elements to Clem, P. L., Jr., AMTRAN- A Conversational-ModeComputer System for Scientists and Engineers, in Proc. IBMScientific the Recognition of Acoustic-Phonetic States, J. Acoust. Soc. Computing Symp. onComputer-AidedExperimentation, Am. 38, 213-223 (1965). Yorktown Heights, N.Y., Oct. 11-13, 1965, pp. 115-150(IBM Dammann, J. E., An Experiment in Cluster Detection, Letter to the Editor, IBM J. Res. & Dev. 10, No. 1, 80-88 (Jan. 1966). Corp., White Plains, N.Y., 1966). Clemens, J. K., Optical Character Recognition forReading Daniels, B. T., Quo Vadis OCR, IEEE Almanack 11, 10-11 Sept. (1967). Machine Applications, Ph.D. Thesis, Dept. EE., MIT., Darling, E. M., Jr. and R. D. Joseph, An Experimental Investiga- 1965. tion of Video Pattern Recognition, in Pictorial Pattern Recogni- Climenson, W. D., File Organization and Search Techniques, tion, Proc. Symp. on Automatic Photointerpretation, Washing- in Annual Review of Information Science andTechnology, Pub., ton, D.C., May 31-June 2, 1967, Ed. G. C.Cheng et al., pp. Vol. 1, Ed. C. A. Cuadra, pp. 107-135 (Interscience 457-469 (Thompson Book Co., Washington, D.C., 1968). New York, 1966). Davenport, W. P., Touch-Tone, Data Proc. Mag. 7, No. 10, 36- Clippinger, R.F., Programming Implications of Hardware Trends, IFIP Congress 65, Vol. 1, New York, N.Y.,May 24-29, 38 (Oct. 1965). Books, David, E. E., Jr., Artificial Auditory Recognition in Telephony, 1965, Ed. W. A. Kalenich, pp. 207-212 (Spartan IBM J. Res. & Dev. 2, 294-309 (Oct. 1958). Washington, D.C., 1965). Das;fil-,-E. E., Jr., and 0. G. Selfridge, Eyes and Ears for Com- Clowes, M. B., The Use of Multiple Auto-Correlation inCharacter Ed. G. L. puters, Proc. IRE 50, 1093-1101 (1962). Recognition, in Optical Character Recognition, David, E. E., Jr., M. V. Mathews and H. S. McDonald, 1958 Fischer, Jr., et al., pp. 305-318 (Spartan Books, Baltimore, WESCON Cony. Rec. Pt. 7, pp. 3-10. Md., 1962). Clowes, M. B., and J. R. Parks, A New Technique in Automatic Davidson, L., A Pushbutton Telephone for Alphanumeric Input, Character Recognition, The Computer J. 4, No. 2, 121 -12a Datamation 12, No. 4, 27-30 (April 1966). (1961). Daviei, D. W.., Rev. Int. Doc. 20-,.66--68 (1962). Coggan, B. B., The Design of a Graphic Display System,Rept, Davis, K. H., R. Biddulph and S. Balashek, Automatic Recog- No. 67-36, 185 p. (Dept. of Engineering, Univ. ofCalifornia, nition of Spoken Digits, Proc. Second London Symp. on Ap- Los Angeles, Aug. 1967). plications of Communications Theory (Butterworths, London, Cole, H. and Y. Okaya, Automated Experiment Systems-A 1953). Also in J. Acoust. Soc. Am. 24, 637-642 (1952). Practical Guide, Lab. Management 3, No. 11, 32 (1965). Davis, K. H. and A. C. Norwine, U.S. Patent No. 2,646,465, Cole, H., Y. Okaya and F. Chambers, Computer Controlled Jul; 21, 1953. Diffractometer, Review Scient.Instruments 34, 872-876 Davis, M. R., and 1'. 0. Ellis, The RAND Tablet: A Man-Machine (1963). Graphical Communication Device, AFIPS Proc. Fall Joint Cole, R. I., Ed., Data/Information Availability, 183 p. (Thompson Computer Conf., Vol. 26, San Francisco, Calif., Oct. 1964, Book Co., Washington, D.C., 1966). pp. 325-331 (Spartan Books, Baltimore, Md., 1964). Communicating By Satellite, Vectors (Hughes Aircraft Co.) Davis, R. M., Man-Machine Communication, in Annual Review VIII, 8-13, Fourth Quarter, 1966. of Information Science and Technology, Vol. 1, Ed. C. A. Conn, R. W., Digitized Photographs for Illustrated Computer Output, AFIPS Proc. Spring Joint Computer Conf., Vol. 30, Cuadra, pp. 221-254 (Interscience Pub., New York, 1966). Atlantic City, N.J., April 18-20, 1967, pp. 103-106 (Thompson Denes, P. 13.Computers in Speech Research, in Proc. IBM Books, Washington, D.C., 1967). Scientific domputing Symp. on Computer-Aided Experimenta- Conway, R. W., and H. L. Morgan, Tele-CUPL: A Telephone tion, Yorktown Heights, N.Y., Oct. 11-13, 1965, pp. 245-263 Time Sharing System, Commun. ACM 10, 538-542 (Sept. (IBM Corp., White Plains, N.Y., 1966). 1967). Denes, P., and M. V. Mathews, Spoken Digit Recognition Using Cooper, D. I., The Automated Laboratory, Int. Sci. Tech.36, Time-Frequency Pattern Matching, J. Acoust. Soc. Am. 32, 20-29 (Dec. 1964). 1450-1455 (1960). Cooper, P. W., Nonsupervised Adaptive Detection for Multi- Deutsch, E. S., The Computer Simulation of a Character Recog- variate Normal Distributions, Rept. No.RADC-TR-66-488, nition Machine, unpublished report, General Post Office 52+22 pp. (Rome Air Development Center, Griffiss Air Force Research Station, Dollis Hill, London, England, 1966, 9 p. Base, N.Y., Sept. 1966). Deutsch, J. A., A Theory of Shape Recognition, British J. Psychol. Cooper, W. S., Fact Retrieval and Deductive Question-Answering 46, Pt. 1, 30-37 (Feb. 1955). Retrieval Systems, J. ACM 11, No. 2, 117-137 (April 1964). Cornew, R. W., A Statistical Method of Spelling Correction, Inf. Dickinson, A. H., and J. N. Wheeler, assignors to International & Control 12,79-93 (Feb. 1968). Business Machines Corporation, New York, Reading Machine, Cossum, W. E., M. E. Hardenbrook and R. N. Wolfe, Computer U.S. Patent 2,261,542; patented Nov. 4, 1941; filed Dec. 21, Generation of Atom-Bond Connection Tables from Hand-Drawn 1937. Chemical Structures, in Parameters of Information Science, Diebold, J., The New Computerized Age-1: The New World Proc. Am. Doc. Inst. Annual Meeting, Vol. 1, Philadelphia, Coming, Saturday Review, July 23, 1966, reprinted in Appendix Pa., Oct. 5-8, 1964, pp. 269-275 (Spartan Books, Washington, 3, U.S. House ...,1967, pp. 296-298. D.C., 1964). Dietrich, W., Optical Character Readers for Automatic Document Council on Library Resources, Tenth Annual Report for the Handling inBanking Applications, Elect. Commun. 40, period ending June 30, 1966, Washington, D.C., 1966, 128 p. 312-227 (1965). 153 Dimond, T. L., Devices for Reading Handwritten Characters, Estrin, G., D. Hopkins, B. Coggan and S. D. Crocker, SNIPER Proc. Eastern Joint Computer Cong., Vol. 12, Washington, COMPUTER- A Computer in Instrumentation Automation, D.C., Dec. 9-13, 1957, pp. 232-237 (Inst. of Radio Engineers, AFIPS Proc. Spring Lint Computer Conf.. Vol. 30, Atlantic New York, 1958). City, N.J., April 18-20, 1967, pp. 645-656 (Thompson Books, Dinneen, G.P., Programming PatternRecognition, Proc. Washington, D.C., 1967). Western Joint Computer Conf., Vol. 7, Los Angeles, Calif., Evans, T. G., A Heuristic Program to Solve Geometric-Analogy Mar. 1-3, 1955, pp. 94-100 (Inst. of Radio Engineers, New Problems, AFIPS Proc. Spring Joint Computer Conf., Vol. 25, York, 1955). Washington, D.C., April 1964, pp. 327-338 (Spartan Books, Dodd, P. D., and F. B. Wood. Image Information, Classification Baltimore, Md., 1964). and Coding, 1966 IEEE Int. Cony. Rec., Pt. 7, np. 60-71. Dolby, J. L. L. L. Earl and H. L. Resnikoff,ThekApplication of Fain, V. S., The Principles of the Construction of a Machine for English -Word Morphology to Automatic indexing and Extract- Pattern Recognition, Radio Engr. 15, No. 3,16 -23 (Sept. 1960). ing, Rept. No. M-21-65-1, 1 v. (Lockheed Missies and Space Excerpts under title: On the Principles of Designing a Machine Co., Palo Alto, Calif., April 1965). for Recognizing Images, Automation Express 2, No. 9, 7-9 Doyle, L. B., Is Automat;: Classification a Reasonable Applica- (June 1960). Translated from Russian periodical Radiotekhnika tion of Statistical Analysis of Text? Rept. No. SP-1753, 34 p. 15, No. 3, 13-17 (Mar. 1960). (System Development Corp., Santa Monica, Calif., Aug. 31, Fain, V. S., The Quantity of Coordinate Descriptions of Images in 1964). Systems for Recognition of Visible Patterns in Soviet Develop- Doyle, W., Recognition of Sloppy, Hand-Printed Characters, ments in Information Processing and Machine Translation, Group Rept. No. 54-12, 12 p. (Lincoln Lab., M.I.T., Lexington, New York, U.S. JPRS, July 1960, 14 p. Excerpts under title: Mass., Dec. 10, 1959). Also in Proc. Western Joint Computer On the Quantity of Coordinate Descriptions of the Images in Conf., Vol. 17, San Francisco, Calif., May 3-5, 1969, pp. Systems Designed for Recognition of Visible Objects, Auto- 133-142 (Western Joint Computer Conf., San Francisco, Calif., mation Express 2, No. 10, 15-18 (July 1960). 1960). Fano, R. M., The Computer Utility and the Community, 1967 Drattell, A., Scahning the World of OCR, Bus. Automation 15, IEEE Int. Cony. Rec., Pt. 12, pp. 30-37. 32-39, 54 (Jan. 1968). Fant, G., Acoustic Analysis and Synthesis of Speech with Ap- Dreher, J. J., Bistatic Target Signatures and Their Acoustic plications to Swedish, reprint from Ericsson Technics No. 1. Recognition -A Suggested Animal Model, DARL No. 3, 34 p. 1959, 108 p. (Douglas Advanced Research Labs., Huntington Beach, Fant, G. and K. N. Stevens, Systems for Speech Compression, Calif., Feb 1966). Fortschritte der Hoch Frequenz-technik 5, 229-262 (1960). Dressler, R. F. and W. Werner, Error Rates for Two Methods of Feidelman, L. A., A Survey of the Character Recognition Field, Statistical Pattern Recognition, J. ACM 11, No. 4, 471-480 Datamation 12, No. 2, 45-52 (Feb. 1966). (Oct. 1964). Feidelman, L. A. and J. L. Katz, Auerbach Special Report Optical Duda, R. 0. air! Fossum, Pattern Classification by Inter- Character Recognition, 1 v. (Auerbach Corp., Philadelphia, Pa., actively Determined Linear and Piecewise Linear Discriminant Mar. 1967). Functions, Trans. IEEE Electron. Computers EC -15, No. 2, Feidelman, L. A. and J. L. Katz, Scanning the Optical Scanners, 220 -232 (Apr. 1966). Data Proc. Mag. 9, 34-36 38, 40, 42, 44 (Oct. 1967). Dudley, H. and S. Balashek, Automatic Recognition of Phonetic Feldmi.n, A., A Proposed Improvement in the Printing of Chem- Patterns in Speech, J. Acoust. Soc. Am. 30, 721-732 (1958). ical Structures, Which Results in Their Complete Computer Dudley, H. W., The Vocoder, Bell Lab. Rec. 18, 122-126 (1936). Codes, Am. Doc. 15, No. 3, 205-209 (July 1964). Dunn, H. K., T. S. Yu and W. D. Chapman, Some Theoretical Feldman, A., D. B. Holland and D. P. Jacobus, The Automatic and Experimental Aspects of Analysis with the Sound Spectro- Encoding of Chemical Structures, J. Chem. Doc. 3, 187-188 graph, Rept. NO. 7 (Communication Sciences, Lab., Univ. of (Oct. 1963). Michigan, Ann Arbor, Aug. 1966). Fiala, F. T., Time-Sharing Operations and Management, Data Processing, Vol. X, Proc. 1966 Int. Data Processing Conf., Earnest, L. D., Machine Recognition of Cursive Writing, in Infor- Chicago, Ill. June 21-24, 1966, pp. 157-166 (Data Processing mation Processing 1962, Proc. IFIP Congress 62, Munich, Management Assoc., 1966). Aug. 27-Sept. 1, 1%2, Ed. C. M. Popplewell, pp. 462-465 Firschein, 0. and M. Fischler, Letter to the Editor, Am. Doc. 18, (North-Holland Pub. Co., Amsterdam, 1963). No. 1, 49 (Jan 1967). Eden, M. and M. Halle, Characterization of Cursive Handwriting, Fischer, G. L. Jr., D. K. Pollock, B. Radack and M. E. Stevens, in Information Theory, Proc. Fourth London Symp. London, Eds., Optical Character Recognition 412 p. (Spartan Books, Aug. 20 -Sept. 2, 1960, Ed. C. Cherry, pp. 287-299 (Butter- Washington, D.C., 1962). Fischer, W. A., The Infinite Variety of Land Resources, abstract, worths, Washington, D.C., 1961). WESCON 68 Program, Western Electronic Show and Con- EDUNET, see Brown et al. vention, Los Angeles, Calif., Aug. 20-23, 1968, pp. 45-46. Edwards, A. W. and R. L. Chambers, Can A Priori Probabilities Fischler, M., Hyperplane Techniques in Pattern Recognition, Help in Character Recognition, J. ACM 11, No. 4, 465-470 Proc. IEEE, Correspondence Section, March 1963. (Oct. 1964). Flanagan, J. L., Speech Analysis Synthesis and Perception (Aca- Edwards, W., Probabilistic Information Processing System for demic Press, New York, 1965). Diagnosis and Action Selection, in Second Cong. on the Flory, L. E. and W. S. Pike, assignors to Radio Corporation of Information System Sciences, held at The Homestead, Hot America, Apparatus for Indicia Recognition, U.S. Patent Springs, Va., Nov. 1964, Ed. J. Spiegel and D. E. Walker, pp. 2,615,992; patented Oct. 28, 1952; filed Jan. 3, 1949. Class 141-155 (Spartan Books, Washington, D.C. 1965). 179-100.3. Eldridge, F. H., The Use of System Analysis in the Acquisition Ford, J. D. Jr., Automated Content Analysis, Tech. Memo. of Information Systems, in Information System Science and TM-904, 12 p. (System Development Corp., Santa Monica, Technology, papers prepared for the Third Con!,scheduled Calif., Feb. 19, 1963). for Nov. 21-22, 1966, Ed. D. E. Walker, pp. 175-1&2 (Thompson Forgie, J. W. and C. D. Forgie, Results Obtained from a Vowel Book Co., Washington, D.C., 1967). Recognition Computer Program, J. Acoust. Soc. Amer. 31, 1480 -1489 (1959). Elkind, J., bee W. Sprick and K. Ganzhorn, An Analogous Method Forte, A., Music and Computing: The Present Situation, AFIPS for Pattern Recognition by Following the Coundary, in Informa- Proc. Fall Joint Computer Conf., Vol. 31, Anaheim, Calif., tion Processing, Proc. Int. Conf. on Information Processing, Nov. 14-16, 1967, pp. 527-329 (Thompson Books, Washing- UNESCO, Paris, June 15-20, 1959, pp. 238-244 (Oldenbourg. ton, D.C., 1967). Munich; Butterworths, London. 1960). Discussion by J. C. R. Forte, A., Syntax-Based Analytic Reading of Musical Scores, Licklidu, ibid., p. 244. Rept. No. MAC-TR-39, 36 p. Cambridge, Mass., English, W. K., D. C. Engelbart and M. L. Berman, 'Display Apr. 1967). Selection Techniques for Text Manipulation, IEEE Trans. FOSDIC III to Assist in 1960 Census, NBS Tech. News Bull. 43, Human Factors HFE-8, 5 (1967). No. 6, 106-107 (June 1959).

154 FOSDIC IV Reads Microfilmed Weather Data for Computer, Gaylord. C. V., Multiprogramming and the Design of On-Line NBS Tech. News Bull. 51, 63-65 (Apr. 1967). Control Systems, Data Proc. Mag. 10, 26-27, 30-31, 34, 36, Franco, A. G., N. Marchand and L. J. Saporta, Error-Control 38 (May 1968). Systems Get The Message Across, Electronics 38, No. 23, Geddes, E. W., R. L. Emrich and J. F. McMurrer, Feasibility 125-136 (1965). Report and Recommendations for New York State Identi- Frank, A. I., Multiple-Font Reading-A System Description, fication and Intelligence System, Tech. Memo. TM-LO- Data Systems Design 1, No. 3, 21-25 (Mar. 1964). 1000/000/00, 276 p. (System Development Corp., Santa Monica, Franklin, R. M., Applied Data Transmission Within Industry, in Calif., Nov. 1, 1963). Data Processing Yearbook 1965, pp. 179-182 (American Genetta, T. L., H. P. Guerber and A. S. Rettig, RCA's Automatic Data Processing, Inc., Detroit, Mich., 1964). Store and Forward Message Switching System, Proc. Western Freeman, M. D. and E. D. Hietnanen, Application of Parallel Joint Computer Conf., Vol. 17, San Francisco, Calif.. May 3-5. Neighborhood Logic to Fingerprint Processing, in Law En- 1960, pp. 365-369 (Pub. by Western Joint Computer Conf., forcement Science and Technology, Vol. 1, Proc. First National San Francisco, Calif., 1960). Symp. on Law Enforcement Science and Technology, Chicago, Gentle, E. C., Jr., Data Communications in Business, 163 p. III., March 1967, Ed. S. A. Yefsky, pp. 507-509 (Thompson (American Telephone and Telegraph Co., New York, 1965). Book Co., Washington, D.C., 1967). Gianopulos, G. N. and F. G. Curl. Use of the Computer in the Frishkopf, L. S. and L. D. Harmon, Machine Reading of Cursive Ranger and Mariner Projects. Proc. !BM Scientific Computing Script, in Information Theory, papers read at Fourth Symp., Symp.onComputer-AidedExperimentation,Yorktown held at the Royal Institution, London, Aug. 20-Sept. 2, 1960, Heights, N.Y., Oct. 11-13, 1965, pp. 21-76 (IBM Corp., White Ed. C. Cherry, pp. 300-316 (Butterworths, Washington, D.C., Plains, N.Y., 1966). 1961). Gibbs, G. and J. MacPhail, Philco Corporation, session on optical Fritzsche, D. L., A Systematic Method for Character Recognition, character page readers, in Research and Engineering Council Rept. No. 1222-4, Contract AF 33(616)6137, 56 p. (Antenna of the Graphic Arts Industry, Inc., Proc. 14th Annual Conf., Lab., Ohio State Univ., Research Foundation, Columbus, Rochester, N.Y., May 18-20, 1964, pp. 95-106 (Washington, Nov. 15, 1961). D.C., 1964). Fry, D. B. and P. Denes, Experiments in Mechanical Speech Giuliano, V. E., The Interpretation of Word Associations, in Recognition, in Information Theory, papers read at Third Statistical Association Methods for Mechanized Documenta- Symp., London, Sept. 12-16, 1955, Ed. C. Cherry, pp. 296-212 tion, Symp. Proc., Washington, D.C., March 17-19, 1964, (Academic Press, New York, 1956). NBS Misc. Pub. 269, Ed. M. E. Stevens et al., pp. 25-32 Fry, D. B. and P. Denes, An Analogue of the Speech Recognition (U.S. Govt. Printing Office, Washington, D.C., Dec. 15, 1965). Process, in Mechanisation of Thought Processes, Vol.1, Giuliano, V. E., P. E. Jones, G. E. Kimball, R. F. Meyer and Proc. Symp. held at the National Physical Laboratory, Nov. B. A. Stein, Automatic Pattern Recognition by a Gestalt 24-27,1958, pp. 375-384; discussion,pp. 385-395 (Her Method, Inf. & Control 4, 332-345 (Dec. 1961). Majesty's Stationery Offi';e, London, 1959). Glucksman, H. A., A Parapropagation Pattern Classifier, IEEE Fry, T. C., Remote Control of the Computer, Letter to the Editor, Trans. Electron. Computers EC-14, 434-443 (1965). Electrical Engineering 73, 297-298 (Mar. 1954). Glucksman, H. A.,Classification of Alphabetic Letters by Fukunaga, K., Feature Selection in Pattern Recognition and Characteristic Loci, Research Rev. 7, 12-14 (Mar. 1968). Communications, 1968 WESCON Technical Papers, 21/1, Goettel, H. H., Bell System Business Communications Seminar, Aug. 1968, 7 p. in Data Processing, Vol. X, Proc. 1966 Int. Data Proc. Con:., Fuller, R. H., Content-Addressable Memory Systems, Rept. Chicago, EL, June 21-24, 1966, pp. 188-197 (Data Processing No. 63-25, 2 v. (Dept. of Engineering, Univ. of California, Management Assoc., 1966). Los Angeles, 1963). Fulton, R. L., Visual Input to Computers, Datamation 9, No. 8, Gold, B., Computer Program lb: Pitch Extraction, J. Acoust. 37-40 (Aug. 1965). Soc. Am. 34, 916-921 (1962). Fussier, H. H., Readjustments by the Librarian, Lib. Q. 23, Goldberg, E., Statistical Machine, U.S. Patent 1,838,389; patented 216-229 (1953). Dec. 29, 193); filed April 5, 1928; filed in Germany April 12, 1927. Gabor, D. and P. J. C. Hill, Television Band Compression by Comolak, L. S., Better and Faster Design by Machine, Electronics Contour Interpolation, Proc. IEE 108, Pt. B, 303-315 (1961). 37, 64-71 (June 1964). Gaffney, J. E. Jr., Some Approaches to the Identification of Goodman, L. A. and W. H. Kruskal, Measures of Association for Fingerprint Patterns by Machine Methods, abstract, paper Cross-Classification, Am. Stat. Assoc. .J. 49, 732 (1954). presented to the Pattern Recognition Society, Dec. 13, 1967. Goodman, L. A. and W. H. Kruskal, Measures of Association Galli, E. J., The Stenowriter- A System for the Lexical Proc- for Cross-Classifications, II, Further Discussions and Ref- essing of Stenotypy, IRE Trans. Electron. Computers EC -11, erences, Am Stat. Assoc. J. 54, 123-163 (1959). 187-199 (1962). Graham, D. N., Image Transmission by Two-Dimensional Contour Gamba, A., The Papistor- An Optical PAPA Device, Institute Coding, Proc. IEEE 55,336-346 (Mar. 1967). di Fisica dell' UniversityGenour (Italy), Mimeo, n.d., 3 p. Gamba, A., A Multilevel PAPA, Suppl. Nuovo Cimento 26, Graham, R. E., Predictive Quantizing of Television Signals, 176-177 (1962). 1958 IRE WESCON Cony. Rec., Pt. 4, pp. 147-157. Gamba, A., G. Palmieri and R. Sanna, Self-Learning in PAPA, Gray, S. B., A Highly Flexible, Many-Font General-Purpose Suppl. Nuovo Cimento 20, 146-147 (1961). Page Reader, in Automation and Scientific Communication, Garfield, E., Can Citation Indexing Be Automated?, in Statistical Short Papers, Pt. 1, papers contributed to the Theme Sessions Association Methods for Mechanized Documentation, Symp. of the 26th Annual Meeting, Am. Doc. Inst., Chicago, Ill., Proc., Washington, D.C., March 17 - -19, 1964, NBS Misc. Pub. Oct. 6-11, 1963, Ed. H. P. Luhn, pp. 85-86 (Am. Doc. Inst., 269, Ed. M. E. Stevens et al., pp. 189-192 (U.S. Govt. Printing Washington, D.C., 1963). Office, Washington, D.C., Dec. 15, 1965). Greanias, E. C., Some Important Factors in the Practical Utili- Gar .tash, V. A., Seminar on Reading Machines, in Foreign zation of Optical Character Readers, in Optical Character DevelopmentsinMachineTranslationandInformation Recognition, Ed. G. L. Fischer, Jr., et al., pp. 129-146 (Spartan Processing, No. 17, Washington, U.S. JPRS, Jan. 5, 1961, Books, Washington, D.C., 1962). pp. 11-12. Greanias, E. C., Progress in Handwriting Recognition, in Data Garvin, P. L., Ed., Natural Language and the Computer, 398 p. Processing Yearbook 1965, pp. 59-62 (American Data Proc- (McGraw-Hill, New York, 1963). essing, Inc., Detroit, Mich., 1964). Garvin, P. L. and W. Karush, Linguistics, Data Processing, and Greanias, E. C., The Computer in Medicine, Datamation 11, Mathematics, in Natural Language and the Computer, Ed. No. 12, 25-28 (Dec. 1965). P. L. Garvin, pp. 357-369 (McGraw-Hill, New York, 1963). Greanias, E. C., P. S. Meagher, R. j. Norman and P. Essinger, Gattner, G. and R. Jurk, Application and Technique of Automatic The Recognition of Handwritten Numerals by Contour Anal- Character Recognition, Siemens Review xxx, 390-393 (1963). ysis, IBM J. Res. & Dev. 7, No. 1, 14-21 (Jan. 1963).

366-107 0 - 70 - 11 155 Green, P. E. Jr., R. A. Frosch and C. F. Romney. Principles of Hennis, R. B., Recognition of Unnurtured Characters in a Multi- an Experimental Large Aperture Seismic iarray (LASA), font Application, Tech. Pub. 07.212, 13 p. (IBM Systems in Proc. IBM Scientific Computing Symp. on C,mputer-Aided Develop.nent Div., Lab., Rochester, Minn., May 9, 1967). Experimentation, Yorktown Heights, N.Y., Oct. 11-13, 1965, Henriclion, E. G. Jr., An Algorithm for Nonparametric Pattern pp. 327-339 (IBM Corp., White Plains. N.Y., 1966). Recognition, Tech. Rept. No. 14, 16 p. (Information Research Greenberg, H. H. and A. G. Konheim, Linear and Nonlinear Associates, Inc., Lexington, Mass., Nov. 17, 1967). Methods in Pattern Classification, IBM J. Res. & Dev. 8, Herbert, E., Information Transfer, Int. Sci. & Tech. 51, 26-37 No. 3, 299-307 (July 1964). (1966). Greenly, J. F., Standardization of Typewriter Fonts for Auto- Hickey, P. R., The A-1200: A Narrow Band System Test Vehicle, matic Reading, Rept. No. RADC-TR-65-523, 47 p. (Griffiss in Data Processing, Vol. X, Proc. 1966 Int. Data Processing Air Force Base, N.Y., March 1966). Conf., Chicago, III., June 21-24, 1966, pp. 175-187 (Data Grimsdale, R. L., F. H. Sumner, C. J. Tunis and T. Kilburn, A Processing Management Assoc., 1966). System for the Automatic Recognition of Patterns, Proc. Highieyman, W. H., An Analog Method for Character Recogni- Inst. of Electrical Engineers, 106, part B (Mar. 1959) pp- tion, IRE Trans.Electron.Computers EC-10, 502 -512 210 -221. (Sept. 1961). Groner, G. F., Real-Time Recognition of Hand-Printed Text, Highleyman, W. H., Linear Decision Functions with Application AFIPS Proc. Fall Joint Computer Conf., Vol. 29, San Francisco, to Pattern Recognition, in Optical Character Recognition, Calif., Nov. 7-10, 1966 pp. 591-601. (Spartan Books, Wash- Ed. G. L. Fischer, Jr., et al., pp. 249-285 (Sparta Books, ington, D.C., 1966). Washington, D.C., 1962). Guzman-Arenas, A., Some Aspects of Pattern Recognition by Highleyman, W. H. and L. A. Kamentsky, A Generalized Scam:er Computer, Rept. N. MAC-TR-37 (Thesis) 118 p. (M.I.T., for Pattern- and Character-Recognition Studies, Proc. Western Cambridge, Mass., Feb. 1967). Joint Computer Conf., Vol. 15, San Francisco, Calif. Mar, 3-5, 1959, pp. 291-294 (Inst. of Radio Engineers, New York, 1959). Haagen, W. F. Edge Enhancement of Photographic Imagry -A Hill, D. R., STAR-A Machine to Recognize Spoken Words, in Mathematical Representation and Optical Implementation, Information Processing 1965, Proc. IFIP Congress 65, Vol. 2, 968 WESCON Technical Papers, 25/2, Aug. 1968, 10 p. New York, N.Y., May 24-29, 1965, Ed. W. A. Kalenich, p. 357 Hagan, W. K. and S. D. Larks, Long Distance FM Telephone (Spartan Books, Washington, D.C., 1966). Transmission of Fetai Electrocardiogram, Am. J. Med. Elec. 2, Hillegass, J. R. and L., F. Melick, A Survey of Data Collection No. 2, 147-151 (Apr.-June 1963). Systems, Data Proc. Mag. 9, No. 6, 50-54 (June 1967). Hall, D. J., A Simple Model of a Pattern Recognition System Hittel, L. A., Some Problems in Data Communications Between (Stanford Research Institute, Menlo Park, Calif., April 1964). the User and the Computer, AFIPS Proc. Fail Joint Computer Hammer, D. P., National Information Issues and Trends, in Conf., Vol. 29, San Francisco, Calif., Nov. 7-10, 1966, pp. 395- Annual Review of Information Science and Technology, Vol. 2, 402 (Spartan Books, Washington, D.C., 1966). Ed. C. A. Cuadra, pp. 385-417 (Interscience Pub., New York, Hodes, L. Machine Processing of Line Drawings, 15 p. (Lincoln 1967). Lab., M.I.T., Lexington, Mass., Mar. 7, 1961). Hankley, W. J., Fingerprint Classification for Automated Proc- Hodges, D., Spark Chamber Film Measuring Using the CHLOE essire., 1968 WESCON Technical Papers, 18/1, Aug. 1968, 7 p. System, inPictorial Pattern Recognition, Proc. Symp. on Hankl.,W. J. and J. T. Tou, Automatic Fingerprint Interpreta- Automatic Photointerpretation, Washington, D.C., May 31- tion and Classification via Contextual Analysis and Topo- June 2, 1967, Ed. G. C. Cheng et al:, pp. 199-205 (Thonipson % logical Coding, in Pictorial Pattern Recognition, Ed. G. C. Book Co., Washington, D.C., 1968). Cheng et al., pp. 411-456 (Thompson Book Co. Washington, Hoffman, A., The 'Whirling Dervish', A Simulation Study in D.C., 1968). Learning and Recognition Systems, 1962 IRE Int. Cony. Rec., Hargi-eaves, B., J. D. Joyce, G. L. Cole. E. D. Foss, R. G. Gray, Pt. 4, pp. 153-160. E. M. Sharp, R. J. Sippel, T. M. Spellman and R. A. Thorp., Hoffman, A., The Information and Data Exchange Experimental Image Processing Hardware for a Man-Machine Graphical Activities (IDEEA) Program and Its Relation to the National Communication System, AFIPS Proc. Fall Joint Computer Interests, in Toward a National Information System, Second Conf., Vol. 26, San Francisco, Calif., Oct. 1964, pp. 363-386 Annual National Colloquim on Information Retrieval, Phila- (Spartan Books, Baltimore, Md., 1964). . delphia, Pa., Apr. 23-24, 1965, Ed. M. Rubinoff, pp. 87-1P; Haring, D. R., A Display Console for an Experimental Computer- (Spartan Books, Washington, D.C., 1965). Based Augmented Library Catalog, Proc. 23rd National Conf., ACM, Las Vegas, Nev., Aug. 27-29, 1968, pp. 35-43 (Brandon/ Hogan, D. L., Speech as Computer Input and Output, 196E IEEE Systems Press, Inc., Princeton, N.J., 1968). Int. Cony. Rec., Vol. 14, Pt. 3, Computers, Presented at IEEE Harmon, L. D., A Line-Drawing Pattern Recognizer, Proc. Int. Cony., New York, N.Y., Mar. 21- 25,1966, pp. 91-93 (The Western Joint Computer Conf., Vol. 17, San Francisco, Calif., Inst. of Electrical and Electronics Engineers, Inc., New York, May 3-5, 1960, pp. 351-364 (Pub. by Western Joint Computer 1966). Conf., San Francisco, Calif., 1960). Also in Electronics 33, Holland, W. B., Ed., Soviet Cybernetics: Recent News Items No. 36, 39-43 (Sept. 2, 1960). No. 2 (The RAND Corp., Santa Monica, Calif., March 1967). Harmon, L. D., Automatic Reading of Cursive Script, in Optical Holland, W. B., Ed. Soviet Cybernetics: Recent News Items Character Recognition, Ed. G. L. Fischer, Jr., et ai., pp. 153- No. 3 (The RAND Corp., Santa Monica, Calif., April 1967). 152 (Spartan Books, Washington, D.C., 1962). Holmes, W. S., Automatic Photointerpretation and Target Harmon, L. D., Automatic Reading of Cursive Script, U.S. Pat. Location, Proc. IEEE 54, 1679-1686 (Dec. 1966). No. 3.127.588, issued Mar. 31, 1964. Harris, C. M., Spectrum Analyis of Speech, Information Systems Holt, A. W. (assignor to Control Data Corporation), Character Summaries, p. 43 (Office of Naval Research, Washington, D.C., Recognition Using Curve Tracing, U.S. Patent 3,142,818, Sept. 1966). issued July 28, 1964, filed Feb. 21, 1961, 9 p. Hart, P. E., A Brief Survey of Preprocessing for Pattern Recogai- Horowitz, P. and E. M. Crane, HECSAGON: A System for tion, Technical Documentary Report, 37 p. (Stanford Research Computer Storage and Retrieval of Chemical Structure, 33 p. Institute, Menlo Park, Calif., Dec. 1966). (Eastman Kodak Co., Rochester, N.Y., 1961, unpublished). Hayes, R. M., Mathematical Models for Information Retrieval, Horvath, V. V., J. M. Holeman and C. Q. Lemmond, Fingerprint in Natural Language and the Computer, Ed. P. L. Garvin, Recognition by Holographic Techniques, in Law Enforcement pp. 268-309 (McGraw-Hill, New 'York, 1963). Science and Technology, Vol. 1, Proc. First National Symp. on Heiman, D. R., E. E. Barrett, R. Hayum and F. 0. Williams, Law Enforcement Science and Technulogy, Chicago, Ill., WADE': A Versatile Auto- atic Data Exchange, Comm. & March 1967, Ed. S. A. Yefsky, pp. 485-492 (Thompson Book Elec. 82, 478-482 (1963). Co., Washington, D.C., 1967). Hemami, H., Qualitative Pattern Recognition via Optical Systems, Horwitz, L. P. and G. L. Shelton, Jr., Pattern Recognition Using 1968 WESCON Technical Papers, 25/1, Aug. 1968, 8 p. Autocorrelation, Proc. IRE 49, No. 1, 175-185 (Jan. 1961). Henle, R. A. and L. 0. Hill, Integrated Computer Circuits-Past, Hospitals Share Computer to Aid Patient Service, Systems 7, Present, and Future, Proc. IEEE 54, 1849 -1860 (Dec. 1966). No. 6, 12-14, 30, 39 (June 1966). 156 Howerton, P., User Requirements, in Optical Character Recogni- figs. (Office of Telecommunications Management, Executive tion, Ed. G. L. Fischer, Jr., et al., pp. 348-353 (Spartan Books, Office of the President, Washington, D.C., 1967). Baltimore, Md., 1962). Johnson, R. B., assignor to Lau-national Business Machines Hu. S. -Z., Threshold Logic, 338 p. (Univ. of California Press, Corp.. New York.Indicia-Controlled Record Perforating Berkeley, Calif., 1965). Machine, U.S. Pat. 2,741,312; patented Apr. 10, 1956; filed Huang, T. S.. PCM Picture Transmission, IEEE Spectrum 2, Sept. J8, 1951. Class 164-115. 57-63 (Dec. 1965). Johnson, T. E., Sketchpad III, A Computer Program for Drawing Huang, T. S. and 0. J. Tretiak, Research in Picture Processing, in Three Dimensions, AFIPS Proc. Spring Joint Computer in Optical and Electro-Optical Information Processing, Ed. Conf., Vol. 23, Detroit, Mich., May 1963, pp. 347-353 (Spartan Tippett et al., pp. 45-57 (11.1.T. Press, Cambridge, Mass., Books, Baltimore, Md., 1963). 1965). Joseph, R. D., S. S. Viglione and H. F. Wolf, Cloud Pattern Huang, 'F.S., 0. J. Tretiak, B. Prasada and Y. Yamaguchi, Recognition, Proc. 19th National Conf. ACM, Philadelphia, Pa., Design Considerations in PCM Transmission of Low-Resolu- Aug. 24 -27, 1964, pp. D2.3-1 to D2.3-17 (Assoc. for Computing tion Monochrome Still Pictures, Proc. IEEE 55, 331-335 Machinery, New York, 1964). (Mar. 1967). Julesz, B., A Method of Coding Television Signals Based on Edge Hughes, G. F. and J. A. Lebo, Data Reduction Using Information Detection, Bell Sys. Tech. J. 38, No. 4, 1001-1020 (July 1959). Theoretic Techniques, Rept. No. RADC-TR-67-67, 144 p. Julesz, B., Towards the Automation of Binocular Depth Percep- (Rome Air Development Center, Griffiss Air Force Base, New tion, in Information Processing 1962, Proc. IFIP Congress 62, York, Mar. 1967). Munich, Aug. 27-Sept. 1, 1962, Ed. C. M. Popplewell, pp. 439- Hunt, M. S., Laser System for Lunar Ranging. Research Rev. 444 (North-Holland Pub. Co., Amsterdam, 1963). Office of Aerospace Research 6, No. 12, 5-6 (Dec. 1967). Huskey, H. D., On-Line Computing Systems: A Summary. in Kalenich, W. A., Ed., Information Processing 1965, Proc. IFIP On-Line Computing Systems, Proc. Symp. sponsored by the Congress 65, Vol. 1, New York, N.Y., May 24-29, 1965, 304 p. Univ. of California, Los Angeles, and Informatics, Inc., Feb. (Spartan Books, Washington, D.C., 1965). 1965, Ed. E. Burgess, pp. 139-142 (American Data Processing, Kalenich, W. A., Ed., Information Processing 1965, Proc. IFIP Inc., Detroit, Mich., 1965). Congress 65, Vol. 2, New York, N.Y., May 24-29, 1965, Hustvedt, E. H., private communication, October 1967, p. 7, 2. pp. 305-648 (Spartan Books, Washington, D.C., 1966). Kamentsky, L. A., The Simulation of Three Machines Which Information International, Inc., III Computer Eye, brochure, n.d., Read Rows of Handwritten Arabic Numbers, IRE Trans. 5. Electron. Computers EC -10, 489-501 (Sept. 1961). Innesp,D. J., FILTER-A Topological Pattern Separation Com- Kamentsky, L. A., Character Recognition System, U.S. Pat. puter Program, Proc. Eastern Joint Computer Conf., Vol. 18, No. 3,018,471, issued Jan. 23, 1962. New York, N.Y., Dec. 13-15, 1960, pp. 25-37 (Pub. by Eastern Kamentsky, L. A. and C. N. Liu, Computer-Automated Design Joint Computer Conf., New York, 1960). of Multifon Print Recognition Logic, IBM J. Res. & Dev. 7, Inose, H. and Y. Yasada, A Unity Bit Coding Method by Nega- No. 1, 2-13 (Jan. 1953). tive Feedback, Proc. IEEE 51, 1524-1535 (1963). Kanal, L. N. and N. C. Randall, Recognition System Design by Irwin, M. R., The Computer Utility, Datamation 12, 22-27 (Nov. Statistical Analysis, Proc. 19th National Conf. ACM, Phila- 1966). delphia, Pa., Aug. 25-27, 1964, pp. D2.5-1 to D2.5-10 (Assoc. Ivanov, V. V., Linguistics as a Theory of Relations Between for Computing Machinery, New York, 1964). Language Systems and Its Modern Practical Applications, Kanal, L., et al., Basic Principles of Some Pattern Recognition translation of Russian Linguistic Studies on Machine Transla- Systems, Proc.National Electronics Conf., Chicago, Ill., tion, No. 2, 1961, in Joint Publications Research Service, 1962, pp. 279-295. Foreign Developments in Machine Translation and Information Kaplan, S. J., The Advancing Communication Technology and Processing No. 80, 1962, pp. 1-35. Computer Communication Systems, AFIPS Proc. Spring Joint Computer Conf., Vol. 32, Atlantic City, N.J., Apr. 30-May 2, Jacks, E. L., A Laboratory for the Study of Graphical Man- 1968, pp. 119-133 (Thompson Book Co., Washington, D.C., Machine Communications, AFIPS Proc. Fall Joint Computer 1968). Conf., Vol. 26, San Francisco, Calif., Oct. 1964, pp. 343-350 Xasher, A., Ed., Proc. International Symp. on Automation of (Spartan Books, Baltimore, Md., 1964). Population Register Systems, Vol. 1, Jerusalem, Israel, Sept. Jacobellis, B. R., Impact of Computer Technology on Com- 25-28, 1967. munications, Proc. 19th National Conf., ACM, Philadelphia, Kazmierczak, H. and K. Steinbuch, Adaptive Systems in Pattern Pa., Aug. 25-27, 1964, pp. N2.1-1 to N2.1-4 (Assoc. for Recognition, IEEE Trans. Electron. Computers EC -12, 822- Computing Machinery, New York, 1964). 835 (Dec. 1963). Jaffe, J., Electronic Computers in Psychoanalytic Research, Kazmierczak, H. and F. Holdermann, The Karlsruhe System presented at the Annual Meeting of the Academy of Psycho. forAutomaticPhotointerpretation,inPictorialPattern analysis, May 4-6, 1952, Toronto, Canada, undated, 20 p. Recognition, Proc. Symp. on Automatic Photointerpretation, Jaffe, J., Computer Analysis of Verbal Behavior in Psychiatric Washington, D.C., May 31-June 2, 1967, Ed. G. C. Cheng et al., Interviews, presented at the Annual Meeting of the Association pp. 45-61 (Thompson Book Co., Washington, D.C., 1968). for Research in Nervous and Mental Disease, New York, Killet, H. G., Experimental Speech Recognizer for Limited Word Dec. 1962, 23 p. (Columbia Univ., College of Physicians and Input. Electronic Communicator 2, No. 6, 3 (Nov.-Dec. 1967). Surgeons, New York, undated). King, J. H., Jr. and C. J. Tunis, Some Experiments in Spoken Jakowitz, C. V., R. L. Shuey and G. M. White, Adaptive Wave- Word Recognition, IBM J. Res. & Dev. 10, No. 1, 65-79 form Recognition, in Information Theory, papers read at Fourth (Jan. 1966). Symp., held at the Royal Institution, London, Aug. 20-Sept. 2, Kingston, C.R., Problemsin Semi-Automated Fingerprint 1960, Ed. C. Cherry, pp. 317-326 (Butterworths, Washington, Classification, in Law Enforcement Science and Technology, D.C., 1961). Vol. 1, Proc. First National Symp. on Law Enforcement Science Janning, T. B., Optical Scanning-Some of Its Applications and and Technology, Chicago, Ill., Mar. 1967, Ed. S. A. Yefsky, Limitations, in Data Processing, Vol. X, Proc. 1966 Int. Data pp. 449-457 (Thompson Book Co., Washington, D.C., 1967). Processing Conf., Chicago, Ill., June 21-24, 1966, pp. 108-118 Kirsch, R. A., Computer Interpretation of English Text and (Data Processing Management Assoc., 1966). Picture Patterns, IEEE Trans. Electron, Computers EC -13, Jaspert, W. P. Ed., Advances in Computer Typesetting, Proc. No. 4, 363-376 (Aug. 1964). Int. Computer Typesetting Conf., Sussex, England, July 14-18, Kirsch, R. A., L. Calm, C. Ray and G. H. Urban, Experiments in 1966, 306 p. (The Institute of Printing, London, 1967). Processing Pictorial Information with a Digital Computer, Johnson, H. R., Concept for a National Security Information Proc. Eastern Joint Computer Conf., Vol. 12, Washington, System, 1966 IEEE Int. Cony. Rec., Pt. 1, pp. 98-105. D.C., Dec. 9-13, 1957, pp. 221-229 (The Institute of Radio Johnson, H. R., Computers and the Public Welfare, Law Enforce- Engineers, Inc., New York, 1958). ment, Social Services and Data Banks, preprint, paper for Klerer, M. and J. May, A User Oriented Programming Language, UCLA Conf. on Computers and Communications, 21 p., plus The Computer J. 8, 103-109 (July 1965).

157 Klerer, M. and J. May, Two-Dimensional Programming, AFIPS tion, Short Papers, Pt. 2, papers contributed to the Theme Proc. Fall Joint Computer Conf.. Vol. 27, Pt. 1, Las Vegas, Sessions of the 26th Annual Meeting, Am. Doe. Inst., Chicago, Nev., Nov. 30Dec. 1,1965, pp. 63-75 (Spartan Books, Ill., Oct. 6-11, 1963, Ed. H. P. Luhn, pp. 201-202 (Am. Doc. Washington, D.C., 1965). Inst., Was ington, D.C., 1%3). Klerer, M. and F. Grossman, Further Advances in Two-Dimen- Ledley, R. S.. J. Jacobsen and M. Belson, BUGSYS: A Program- sional Input-Output by Typewriter Terminals, AFIPS Proc. ming System for Picture ProcessingNot for Debugging, Fall Joint Computer Conf., Vol. 31, Anaheim, Calif., Nov. Commun. ACM 9, 79-84 (Feb. 1966). 14-16, 1967, pp. 675-687 (Thompson Books, Washington, Ledley, R. S., M. Legator and J. B. Wilson, Automatic Determina- D.C., 1967). tion of Mitotic Index. in Pictorial Pattern Rceognition, Proc. Klinge:aan, J. and H. V. Pipberger, Comp..r Classifications of Symp. on Automatic Photointerpretation, Washington, D.C., Electrocardiograms, Computers & Biotned. Res. 1, 1-17 (1967). May 31June 2, 1967, Ed. G. C. Cheng et al., pp. 99-103 Kochen, M., Ed., Some Problems in Information Science, 309 p. (Thompson Book Co., Washington, D.C.,.1968). (The Scarecrow Press, New York, 1965). Lee, F. F., Machine-to-Man Communication by Speech Part 1: Koltovoj, B., Literate Automation, in Izvestiya, Jan. 5, 1967, Generation of Segmental Phonemes from Text, AFIPS Proc. p. 4, as translated a!;..1 summarized by P. Stephan, in Soviet Spring Joint Computer Conf., Vol. 32, Atlantic City, N.J., Cybernetics: Recent News Items No. 2, Ed. W. B. Holland, Apr. 30May 2. 1968, pp. 333-338 (Thompson Book Co., pp. 38-41 The RAND Corp., Santa Monica, Calif., Mar. 1967). Washington, D.C., 1968). Kornberg, W., Scanning the Mails, Electronics 37, 115-119 Lee, M. K., The Demise of the Keypunch, Datamation 14, No. 3, (June 15,1964). 51-55 (Mar. 1968). Kortman, C. M., Data Compression by Redundancy Reduction, Lehmann, W. P., Machine Language Translation Study, Sixth IEEE Spectrum 4, 133-139 (Mar. 1967). Quarterly Progress Report, Aug. 1Oct. 31, 1964, Rept. No. Kovalevsky, V. A., Present and Future of Pattern-Recognition LRC-64P22, 1 v. (Linguistics Research Center, Texas Univ., Theory, in Information Processing 1965, Proc. IFIP Congress Austin, Dec. 1964). 65, Vol. 1, New York, N.Y., May 24-29, 1965, Ed. W. A. LeSchack, A. R., The Determination of Clusters by Matrix Kalenich, pp. 37-43 (Spartan Books, Washington, D.C., 1965). Analysis, in Information Storage and Retrieval, Rept. No. Kovalevsky, V. A., Pattern Recognition Method Based on Refer- ISR-7, Ed. G. Salton, June 1964, pp. XIV-1 to XIV-43 (Compu- ence Pattern Transformations, in Information Processing 1965, tation Lab., Harvard Univ., Cambridge, Mass., 1964). Proc., IFIP Congress 65, Vol. 2, New York, N.Y., May 24-29, Lesti, A., Adaptive and Learning Recognition Machine to 1965, Ed. W. A. Kalenich, pp. 566-567 (Spartan Books, Produce Codes or Punch Cards in Response to Spoken Words, Washington, D.C.. 1966). in Automation and Scientific Communication, Short Papers, Kroger, M. G., Introduction to Tactical Information Systems, in Pt. 2, papers contributed to the Theme Sessions of the 26th Second Cong. on the Information System Sciences, held at Annual Meeting, Am. Doc. Inst., Chicago, III., Oct. 6-11, 1963, The Homestead, Hot Springs, Va., Nov. 1964, Ed. J. Spiegel Ed. H. P. Luhn, pp. 279-280 (Am. Doc. Inst., Washington, and D. E. Walker, pp. 267-273 (Spartan Books, Washington, D.C., 1963). D.C., 1965). Lewin, M. H., A Magnetic Device for Computer Graphic Input, Kubba, M. H., Automatic Picture Detail Detection in the Presence AFIPS Proc. Fall Joint Computer Conf., Vol. 27, Pt. 1, Las of Random Noise, Proc. IEEE 51, 1518-1523 (1963). Vegas, Nev., Nov. 30Dec. 1, 1965, pp. 831-838 (Spartan Kuhl, F., ClassifiCation and Recognition of Handprinted Char- Books, Washington, D.C., 1965). acters, IEEE Int. Cony. Rec., Pt. 4, 1963. Lewis, H. R., SHAPESHIFTER: An Interactive Program for Kuney, J. H., Computer Typesetting as Input to Information Experimenting with Complex-Plane Transformations, Proc. Systems, in Advances in Computer Typesetting, Proc. Int. 23rd National Conf., ACM, Las Vagas, Nev., Aug. 27-29, 1968, Computer Typesetting Conf., Sussex, England, July 14-18, pp. 717-724 (Brandon/Systems Press, Inc., Princeton, N.J., 1966, Ed. W. P. Jaspert, pp. 136-140 (The Institute of Printing, 1968). London, 1967). Lewis, P. M., II, The Characteristic Selection Problem in Kusch, H., Automatic Recognition of Spoken Numbers (Digits), Recognition Systems, IRE Trans. Inf. Theory IT-8, 171-178 NTZ Commun. J. 4, 201-206 (1965). (Feb. 1962). Lichtenstein, I. Computer Automation of Chromatographic Data Lance, G. N. and W. T. Williams, A General Theory of Clas- Processing, 1966 IEEE Int. Cony. Rec., Pt. 3, pp. 6-18. sificatory Sorting Strategies1.Hierarchical Systems, The Licklider, J. C. R., Libraries of the Future, 219 p. Press, Computer J. 9, 373-382 (Feb. 1967). Cambridge, Mass., 1965). Lannon, E.R., Optical Character Recognition in the U.S. Licklider,J.C. R., Interactive Information Processing, in Government, in Advances in Computer Typesetting, Proc. Computer and Information SciencesII, Proc. 2nd Symp. on Int. Computer Typesetting Conf., Sussex, England, July 14-18. Computer and Information Sciences, Columbus, 0., Aug. 22- 1966, Ed. W. P. Jaspert, pp. 48-53 (The Institute of Printing, 24, 1966, Ed. J. T. Tou, pp. 1-13 (Academic Press, New York, London, 1967). 1967). Lannon, E. R., Computers & Composition in the U.S. Govern- Liljencrantz, J., A Filter Bank Speech Spectrum Analyzer, ment Past, Present & Future, in Advances in Computer Paper A-27, Fifth International Congress on Acoustics, Liege, Typesetting, -Proc. lnt. Computer Typesetting Conf. Sussex, Sept. 1965. England, July 14-18, 1966, Ed. W. P. Jaspert, pp. 80-84 (The Lindgren, N., Machine Recognition of Human Language. Part I Institute of Printing, London, 1967). Automatic Speech Recognition, IEEE Spectrum 2, 114-136 z, Larson, A. G. and R. C. Singleton, Real-Time Spectral Analysis (Mar. 1965). on a Small General-Purpose Computer, AFIPS Proc. Fall Lindgren, N., Machine Recognition of Human Language. Part II Joint Computer Conf., Vol. 31, Anaheim, Calif., Nov. 14-16, Theoretical Models of Speech Perception and Language, 1967, pp. 665-674 (Thompson Books, Washington, D.C., 1967). IEEE Spectrum 2, 44-59 (A pr. 1965). Lavington, S. H. and L. E. Rosenthal, Some Facilities for Speech Lindgren, N., Machine Recognition of Human Language. Part Processing by Cmnputer, The Computer J. 9, 330-339 (Feb. IIICursive Script Recognition, IEEE Spectrum 2, 104-116 1967). (May 1965). Lax, B., Progress in Semiconductor Lasers, IEEE Spectrum 2, Lipkin, L. E., W. C. Watt and R. A. Kirsch, The Analysis, 62-75 (July 1965). Synthesis, and Description of Biological Images, Annals N.Y. LeBrun, T. Q., An Electronic Retina Optical Character Reading Acad. Sciences 128, Art. 3, 984-1012 (Jan. 1966). System with Integral Logical Editing and Indexing Capabilities, Liu, C. N., A Programmed Algorithm for Designing Multifont in Automation and Scientific Communication, Short Papers, Character Recognition Logics, IEEE Trans. Electron. Com- Pt. 2, papers contributed to the Theme Sessions of the 26th puters EC-13, 586-593 (Oct. 1964). Annual Meeting, Am. Doc. Inst., Chicago, III., Oct. 6- 11,1963, Lockemann, P. C. and W. D. Knutsen, A Multiprogramming Ed. H. P. Luhn, pp. 149-150 (Am. Doc. Inst., Washington, Environment for Online Data Acquisition and Analysis, D.C., 1963). Commun. ACM 10, 758-764 (Dec. 1967). Ledley, R. S., Automatic Coding of Chemicals Directly from Lucky, R. W., Adaptive Redundancy Removal in Data Trans- Structure Pictures, in Automation and Scientific Communica- mission, Bell Sys. Tech. J. 47, 549-573 (Apr. 1968).

158 Luhn, H. P., Indexing, Language anMeaning, in Information NIarzocco, F. N., Computer Recognition of Handwritten First Storage and Retrieval Theory, Systems and Devices, Air Names, IEEE Trans. Electron. Computers EC -14, 210-217 Force Office of Scientific Research Symp., Washington, D.C., (Apr. 1965). 1958, Ed. M. Taube and H. Wooster, pp. 208-218 (Columbia Mattson, R. L. and J. E. Dammann, A Technique for Determining Univ. Press, New York, 1958). and Coding Subclasses in Pattern Recognition Problems, IBM Luhn, H. P., Potentialities of Auto-Enconding of Scientific J. Res. & Dev. 9, No. 4, 294-302 (July 1965). Literature, Rept. No. RC-101, 22 p. (IBM Research Center, Mauceri, A. J., Feasibility Study of Personnel Identification by Yorktown Heights, N.Y., May 15, 1959). Signature Verification, Rept. No. RADC-TR-65-33, 91 p. Luhn, H.P., Keyword-In-Context for Technical Literature (Rome Air Development Center. Griffiss Air Force Base, N.Y., (KWIC Index), presented at American Chemical Society, Apr. 1965). Division of Chemical Literature at Atlantic City, N.J., Sept. 14, Mayeda, T.. Methodology in Presentation-A New Objective: 1959, Rept. No. RC0127, 16 p. (IBM Corp., Yorktown Heights, The Multi-Media Network, draft of lecture delivered to the N.Y., 1959). Also in Am. Doc. 11, No. 4, 288-295 (Oct. 1960). Washington, D.C., Chapter. The Institute of Management Luhn, H. P., Ed., Automation and Scientific Communication, Sciences, Nov. 1, 1967, 20 p. Short Papers, Pt. 1, papers contributed to the Theme Sessions Mayo, C. R., Electrostatic Printing Today and Tomorrow, in of the 26th Annual Meeting, Am. Doc. Inst., Chicago, Ill., Research and Engineering Council of the Graphic Arts Oct. 6-11, 1963, pp. 1-128 (Am. Doc. Inst., Washington, D.C., Industry, Proc. 14th Annual Conf., Rochester, N.Y., May 18-20, 1963). 1964, pp. 74-79 (Washington, D.C., 1964). Luhn, H. P., Ed., Automation and Scientific Communication, Mazzarese, N. J., Experimental System Provides Flexibility and Short Papers, Pt. 2, papers contributed to the Theme Sessions Continuity, Data Proc. Mag. 7, No 5, 68-70 (1965). of the 26th Annual Meeting, Am. Doc. Inst., Chicago, Ill., McCormick, B. H., S. R. Ray, K. C. Smith and S. Yamada, Oct. 6-11, 1963, pp. 129-352 (Am. Doc. Inst., Washington, ILLIAC III: A Processor of Visual Information, in Information D.C., 1963). Processing 1965, Proc. IFIP Congress 65, Vol. 2, New York, Lynch, W. C., Description of a High Capacity, Fast Turnaround N.Y., May 24-29, 1965, Ed. W. A. Kalenich, pp. 359-360 University Computing Center, Commun. ACM 9, 117-123 (Spartan Books, Washington, D.C., 1966). (Feb. 1966). McDonald, R. F. Audio Response Unit Using a Digitally Stored Lyon, R. J. P., Remote Sensing in Geology, abstract, WESCON 68 Vocabulary, 1966 IEEE Int. Cony. Rec., Pt. 1, pp. 52-55. Program, Western Electronic Show and Convention, Los McDonough, J. T. Jr., Homer, the Humanities, and IBM, Proc. Angeles, Calif., Aug. 20-23, 1968, p. 46. Literary Data Processing Conf., Sept. 9-11, 1964, Ed. J. B. Bessinger, J., et al., pp. 25-36 (IBM Corp., White Plains, Maass, K. K., Optical Character Recognition Based on Peephole N.Y., 1964). Template Matching, Computer Design 4, No. 12, 18-23 (1965). McGee, W. C. and H. E. Petersen, Microprogram Control for Maass, W. G., New Information Services from a Not-So-Old the Experimental Sciences, AFIPS Proc. Fall Joint Computer Publishing House, J. Chem. Doc. 2, No. 1, 46-48 (Jan. 1962). Conf., Vol. 27, Pt. 1, Las Vegas, Nev., Nov. 30-Dec. 1, 1965, Macaulay, M., Man-Machine Interaction-Graphics, in Informa- pp. 77-91 (Spartan Books, Washington, D.C., 1965). tion Processing 1965, Proc. IFIP Congress 65, New York, N.Y., McLaughlin, J. A. and J. Raviv, Nth-Order Autocorrelations in May 24-29, 1965, 'Vol. 2, Ed. W. A. Kalenich, pp. 578-579 Pattern Recognition, Inf. & Control 12, 121-142 (Feb. 1968). (Spartan Books, Washington, D.C., 1966). McManis, H. J.,Electrical Communications inthe Future, Machines That Think?, Vectors (Hughes Aircraft Co.) 8, 4-7, Datamation 12, No. 11, 28-30 (Nov. 1966). 4th Quarter, 1966. Meeker, W. F., A. L. Nelson and P. B. Scott, Voice to Teletype Maeda, K. and T. i'akai, Phonetic Typewriter, Electro-Technology Code Converter Research Program. Part II. Experimental (Jan.-Feb. 1960). Verification of a Method to Recognize Phonetic Sounds, Rept. Maguire, T., An Ear to the Underground, Electronics 38, No. 15, No. ASD-TDR-61-666, Part II, 210 p. (Radio Corporation of 91-93 (1965). America, Camden, NJ., Sept. 1962). Marill, T. and D. M. Green, Statistical Recognition Functions and Meeker, W. F. and L. S. Green, Voice to Teletype Code Converter the Design of Pattern Recognizers. IRE Trans. Electron. Research Program. PartI. Experimental Verification of a Computers EC -9, No. 4, 472-477 (Dec. 1960). Method to Recognize Phonetic Sounds. Rept. No. ASD-TR- Marill, T., A. K. Hartley, T. G. Evans, B. H. Bloom, D. M. R. Park, 61-666, 125 p (Radio Corporation of America, Camden, N.J., T. P. Hart and D. L. Darley, CYCLOPS-1: A Second Genera- Dec. 1961). tion Recognition System, AFIPS Proc. Fall Joint Computer Meissner, P., P. Shoenfeld and D. Lottridge, Automatic Process- Conf., Vol. 24, Las Vegas, Nev., Nov. 1963, pp. 27-33 (Spartan ing of Hydrographic Data, unpublished National Bureau of Books, Baltimore, Md., 1963). Standards Report, Washington, D.C.,June 1963,41 p. Marks, D., Data Transmission, Systems 5, No. 1, 13-17, 60-61 Menkhaus, E. J., The Ways and Means of Moving Data, Bus. (Jan. 1964). Automation 14, 30-37 (Mar. 1967). Markus, J., Computers in Commercial Publishing, 15 p. (McGraw- Mermelstein, P. and M. Eden, Experiments on Computer Hill, New York, Mar. 15, 1965). Recognition of Connected Handwritten Words, Inf. & Control Markuson, B. E., Ed., Libraries and Automation, Proc. Conf. held 7, 225-270 (1964). at Airlie Foundation, Warrenton, Va., May 26 -30, 1963, 268 p., Mermelstein, P. and M. Eden, A System for Automatic Recogni- under sponsorship of the Library of Congress, the National tion of Handwritten Words, AFIPS Proc. Fall Joint Computer Science Foundation, and the Council on Library Resources Conf., Vol. 26, San Francisco, Calif. Oct. 1964, pp. 333-342 (Library of Congress, Washington, D.C., 1964). (Spartan Books, Washington, D.C., 1964). Markuson, B. E., Automation in Libraries and Information Merz, M. J. Jr., Optical Character Page Readers, in Res-arch and Centers,in Annual Review of Information Science and Engineering Council of the Graphic Arts Industry, Proc. 14th Technology, Vol. 2, Ed. C. A. Cuadra, pp. 255-284 (Inter- Annual Conf., Rochester, N.Y., May 18-20, 1964, pp. 81-85 science Pub., New York, 1967). (Washington, D.C., 1964). Marom, E., Fingerprint Classification and Identification Using Optical Methods, in Law Enforcement Science and Technology, Meyer, E., Encoding of Organic-Chemical Structural Formulas Vol. 1, Proc. First National Symp. on Law Enforcement Science and Reactions by Machine, in Automation and Scientific and Technology, Chicago, Ill., Mar. 1967, Ed. S. A. Yefsky, Communication, Short Papers, Pt. 2, papers contributed to the pp. 481-484 (Thompson Book Co., Washington, D.C., 1967). Theme Sessions of the 26th Annual Meeting, Am. Doc. Inst., Moron, M. E., Mechanized Documentation: The Logic Behind a Chicago, III., Oct. 6-11, 1963, Ed. H. P. Luhn, pp. 131-132 Probabilistic Interpretation, in Statistical Association Methods (Am. Doc. Inst., Washington, D.C., 1963). for Mechanized Documentation, Symp. Proc., Washington, Meyer, E., Mechanization of Chemical Documentation, Angew. D.C., Mar. 17-19, 1964, NBS Misc. Pub. 269, Ed. M. E. Stevens Chem. Int. (English edition], 4, 347-352 (Apr. 1965). et al., pp. 9-13 (U.S. Govt. Printing Office, Washington, D.C., Meyer, R. F.,-V. E. Giuliano and P. E. Jones, Analytic Approxi- Dec. 15, 1965). mation and Translational Invariance in Character Recognition, Martin, E. J. and W. S. McKee, Commercial Satellite Communi- in Optical Character Recognition, Ed. G. L. Fischer, Jr., et al., cations Experience, IEEE Spectrum 4, 63-69 (July 1967). pp. 181-195 (Spartan Books, Baltimore, Md., 1962).

159 Mikhailov, A. I., Studies on Automatic Indexing and Abstracting Nagy, G., Preliminary Investigation of Techniques for Auto- in the USSR, Working paper for the Unesco-VINITI Symp. on mated Reading of Unformatted Text, Commun. ACM 11, Mechanized Abstracting and Indexing, Moscow, Sept. 28- 480-487 (July 1968). Oct. 1, 1966, 12 p. Also in Symp. on Mechanized Abstracting Narasimhan, R., Syntax-Directed Interpretation of Classes of and Indexing-Papers and Discussion, Invited Paper, pp 33- Pictures, Commun. ACM 9, 166-173 (Mar. 1966). 41. Unesco Doc. No. SC/WS/172, issued Paris, Jan. 12, 1968 Nathan, R., Picture Enhancement for the Moon, Mars and Man, (Distribution Limited). in Pictorial Pattern Recognition, Proc. Symp. on Automatic Mills, R. G., Communications Implication of the Project MAC Photointerpretation, Washington, D.C., May 31-June 2, 1967, Multiple-Access Computer System, 1965 IEEE Int. Cony. Rec., Ed. G. C. Cheng et al.. pp. 239-266 (Thompson Book Co., Pt. 1, pp. 237-241. Washington, D.C., 1968). Mills, R. G., Man-Machine Communication and Problem Solving, National Physical Laboratory, Boundary and Contrast Enhance- in Annual Review of Information Science and Technology. ment in Automatic Pattern Recognition, Teddington, England, Vol. 2, Ed. C. A. Cuadra, pp. 223-254 (Interscience Pub., n.A., 4 D. New York, 1%7). NCIC Progress Report, FBI Law Enforcement Buil., reprint, Minneman, M. J., Handwritten Character Recognition Employing Sept. 1967, 7 p. Topology, Cross Correlation, and Decisior Theory, IEEE Needham, R. M., Theory of Chimps H, Rept. No. ML-139, 48 p. Trans. SSC-2, 86-96 (Dec. 1966). (Cambridge Language Research Unit. England, Mar. 1961). Minot, 0_ N., Automatic Devices for Recognition of Visible Nettlham, R. M., Methods of Classification (Preprint) NATO Two-Dimensional Patterns: A Survey of the Field, Tech. Advanced Study Institute on Atuomatic Document Analysis, Memo. TM-364, 59 p. (U.S. Naval Electronics Lab., San Diego, Venice, July 1963. Calif., June 25, 1959). Needham, R. M., Practical Techniques and Experiments (Pre- Minsky, M., Steps Toward Artificial Intelligence, Proc. IRE 49, print Abstract) NATO Advanced Study Institute on Automatic No. 1, 8-30 (Jan. 1961). Document Analysis, Venice, July 1963. Minsky, M. and 0_ G. Selfridge, Learning in Random Nets, in Needham, R. M., A Method La Using Computers in information Information Theory, papers read at Fourth Symp. held at the Classification, in Information Processing 1962. Proc. IFIP Royal Institution, London, Aug. 20-Sept. 2,1960, Ed. C. Cherry, Congress 62, Munich, Aug. 27-Sept. 1, 1962, Ed. C. M. Popple- pp- 335-347 (Butterworths, Washington, D.C., 1961). well, pp. 284-287 (North-Holland Pub. Co., Amsterdam, 1963). Minsky, M. L, An Autonomous Manipulator System, M.I.T. Neilsen, I. R., A Simple Data Transmission System Using the Project MAC Progress Report III, pp. 11-17, 1966. Office Telephone, Commun. ACM 8, 634-635 (Oct. 1965). Minsky, M. L. and S. Papert, Perceptrons and Pattern Recceini- Neisser, U. and P. Weene, A Note on Human Recognition of tion, Rept. No. MAC-M-358, 1 v. (M.I.T., Cambridge, .dasz Hand-Printed Characters, Inf. & Control 3, No. 2, 191-196 Sept. 1967). (June 1950). "Modified Tape-Recording Typewriter", NBS Tech. News Bull. Newman, W. M., A Graphical Technique for Numerical Input, 50, 118-119 (July 1966). The Computer J. H, 63-64 (May 1968). Monds, F. C. and G. L. Carayannopolis, Demo I, A Supervised Nkenoff, N., Hardware for Information Processing Systems: or Unsupervised Learning Receiver, 1968 WESCON Technical Today and in the Future, Proc. IEEE 54, 1820-1835 (Dec. Papers, 21/3, Aug. 1968, 12 p. 1966). Montgomery, W. D., Reconstruction of Pictures from Scanned Niwa, X., Present State of Mechanization of Documentation in Records, IEEE Trans. Inf. Theory IT-11, 204-206 (Apr. 1965). Japan, Rev. Int. Doc. 29, 63-65 (1962). Moore, G. A., Direct Quantitative Analysis of Photomicrographics Noguchi, S., J. Oizumi and S. Tomita, Recognition of Hand- by a Digital Computer, Photo. Sci. Eno-. 8, 152-161 (1964). written Katakana Characters,in Advances in Computer Moore, G. A., Automatic Scanning and Computer Presses for Typesetting, Proc. Int. Computer Typesetting Conf., Sussex. the Quantitative Analysis of Micrographs and Equivalent England, July 14-18, 1966, Ed. W. P. Jaspert, pp. 40-47 (The Subjects, in Pictorial Pattern Recognition, Proc. Symp. on Institute of Printing, London, 1967). Automatic Photointerpretation, Washington, D.C., May 31- Northrup, R. A., D. M. Motley and G. K. McAuliffe, Implementa- June 2, 1967, Ed. G. C. Cheng et al., pp. 275-326 (Thompson tion and Performance of ADEM, 1968 WESCON Technical Book Co., Washington, D.C., 1968). Papers, 12/1, Aug. 1968, 9 p. Moore, R. T., M. C. Stark and L. Cahn, Digitizing Pictorial Information with a Precision Optical Scanner, Photogrammetric Ohlman, H., State-of-the-Art: Remote Interrogation of Stored Engineering 30, 923-931 (Nov. 1964). Documentary Material, in Automation and Scientific Communi- Morris, D., F. H. Sumner and M. T. Wyld, An Appraisal of the cation, Short Papers, Pt. 2, papers contributed to tne Theme Atlas Supervisor, Proc. 22nd National Conf. ACM, Washing. Sessions of the 26th Annual Meeting, Am. Doc. Inst., Chicago, ton, D.C., Aug. 29-31, 1967, pp. 67-75 (Thompson Book Co., III., Oct. 6-11, 1963, Ed. H. P. Luhn, pp. 193-194 (Am. Doc. Washington, D.C.. 1967). Inst., Washington, D.C., 1%3). Mounts, F. W., Frame-to-Frame Digital Processing of TV Pictures Okajima, M., L. Stark, G. Whipple and S. Yasui, Computer to Remove Redundancy, abstract, in WESCON 63, Official Pattern Recognition Techniques: Some Results with Real Program. Western Electronic Show and Convention, Los Electrocardiographic Data, IEEE Trans. BNIE-10, 106-113 (July 1963). Angeles, Calif., Aug. 20-23, 1968, p. 28. Okaya, Y., Refinement of the Crystal Structure of Taurine: Muerle, J. L. and D. C. Allen, Experimental Evaluation of An Example of Computer-Controlled Experimentation, Acta Techniques for Automatic Segmentation of Objects ina Cryst. 21, 726 (1966). Complex Scene, in Pictorial Pattern Recognition, Proc. Symp. Olson, H. F. and H. Belar, Syllable Analyzer, Coder and Synthe- on Automatic Photointerpretation, Washington, D.C., May 31- sizer for the Transmission of Speech, IRE Trans. Audio, June 2, 1967, Ed. G. C. Cheng et al., pp. 3-13 (Thompson Book AU-10, 11-17 (1962). Co., Washington, D.C., 1968). O'Sullivan, T. C., Shadow Telephone Networks for Time-Sharing Munson, J. H., A. E. Brain, G. E. Forsen, D. J. Hall, N. J. Nilsson Terminals, Computers & Automation 15, 10, 33-39 (1966). and C. A. Rosen, A Pattern-Recognition Facility with a Otten, K. W., Automatic Speech Recognition, Rept. No. 33-IF. Computer-ControlledLearningMachine,inInformation AL TR-65-205, 114 p. (National Cash Register Co., Dayton, 0., . Processing 1965, Vol. 2, Proc. IFIP Congress 65, New York, N.Y., May 24-29, 1965, Ed. W. A. Kalenich, pp. 360-361 Sept. 1965). (Spartan Books, Washington, D.C., 1966). Paine, R. M., Preparation for Optical Character Recognition, The Computer J. 9, 221-229 (Nov. 1966). Nadler, M., An Analog-Digital Character Recognition System, Palmieri, G. and R. Sauna, A New PAPA Machine, Suppl. IEEE Trans. Electron. Computers EC -12, 814-821 (Dec. 1963). Nuovo Cimento, Series X. 23, 266-275 (1962). Nadler, M., The Perspectives for Practical Optical Character Parker, D. B., Graphical Communication in an On-Line System, Recognition, in Advances in Computer Typesetting, Proc. Int. in On-Line Computing Systems, Proc. Symp. sponsored by Computer Typesetting Conf., Sussex, England, July 14-18, the Univ. of California, Los Angeles, and Informatics, Inc., 1966, Ed. W. P. Jaspert, pp. 36-39 (The Institute of Printing, Los Angeles, Calif., Feb. 2-4, 1965, Ed. r. Burgess, pp. 89-100 London, 1967). (American Data Processing, Inc., Detroit, Mich., 1965).

160 Parker, D. C. and M. F. Wolff, Remote Sensing, Int. Sci. Tech. and B. D. Waxman, pp. 377-407 (Academic Press, New York, 43, 20-31 (1965). 1965). Parker-Rhodes, A. F., Contributions to the Theory of Clumps. Pizer, S. M. and H. G. Vetter, Perception and Processing of The Usefulness and Feasibility of the Theor), Rept. No. ML 138, Medical Radioisotope Scans, in Pictorial Pattern Recognition, 34 p. (Cambridge Language Research Unit, Cambridge, Proc. Symp. on Automatic Photointerpretation, Washington, England, Mar. 1961). D.C., May 31June 2, 1967, Ed. G. C. Cheng et al., pp. 147-156 Parker-Rhodes, A. F. and R. M. Needham, The Theory of Clumps, (Thompson Book Co., Washington, D.C., 1968). Rept. No. ML 126, 1 v. (Cambridge Language Research Unit, Poppelbaum, W. J. and M. Faiman, Uncommon Electro-Optics Cambridge, England, Feb. 1960). for Display and Processing, 1968 WESCON Technical Papers, Patrick, R. L. and D. V. Black, Index Files: Their Loading and 16/6, Aug. 1968, 4 p. Organization for Use, in Libraries and Automation, Proc. Popplewell, C. M., Ed., Information Processing 1962, Proc. Conf. on Libraries and Automation, Airlie Foundation, Warren- IFIP Congress 62, Munich, Aug. 27Sept. 1, 1962, 780 p. ton, Va., May 26-30, 1963, Ed. B. E. Markuson, pp. 29-48 (North-Holland Pub. Co., Amsterdam, 1963). (Library of Congress, Washington, D.C., 1964). Porter, J. W. and L. E. Johnson, United Air Lines' Electronic Penterman, D. G. and D. J. Casey, A Challenge to Law Enforce- Information System (EIS), in Data Processing, Vol. X, Proc. mentReduce the Service-Need System Capability Gap by Int. Data Processing Conf., Chicago, III., June 21-24, 1966, Telemobility and Program Integration, in Law Enforcement pp. 74-82 (Data Processing Management Assoc., 1966). Science and Technology, Vol. 1, Proc. First National Symp. Potter, R. J.J., On Optical Character Recognition of Text, Proc. on Law Enforcement Science and Technology, Chicago, Ill., Literary Processing Conf., Sept. 9-11, 1964, Ed. J. B. Mar. 1967, Ed. S. A. Yefsky, pp. 83-88 (Thompson Book Co., Bessinger et al., pp. 306-323 (IBM Corp., White Plains, Washington, D.C., 1967). N.Y., 1964). Perlman, J. A., Digital Data Transmission: The User's View, in Prather, R. C. and L. M. Uhr, Discovery and Learning Techniques ComputersKey to Total Systems Control, Proc. Eastern for Pattern Recognition, Proc. 19th National Conf., ACM, Joint Computer Conf., Vol. 20, Washington, D.C., Dec. 12-14, Philadelphia, Pa., Aug. 25-27, 1964, pp. D2.2-1 to D2.2-10 1961, pp. 209-212 (Macmillan Co., New York, 1961). (Assoc. for Computing Machinery, New York, 1964). Perry, J. H., Jr., Integrated Automation in Newspaper and Book Press, F. and W. F. Brace, Earthquake Prediction, Science 152; Production, AFIPS Proc. Fall Joint Computer Conf., Vol. 29, No. 3729, 1575-1584 (1966). San Francisco, Calif., Nov. 7-10, 1966, pp. 125-136 (Spartan Prince, M. D., Man-Computer Graphics for Computer-Aided Books, Washington, D.C., 1966). Design, Proc. IEEE 54, 1698-1708 (Dec. 1966). Perry, J. H., Jr., Applications of Computers and Optical Character Pruslin, D. H., Quarterly Progress Report No. 80, Research Lab. Recognition Devices, in Advances in Computer Typesetting, for Electronics911.I.T., 221, 225-226 (1966). Proc. Int. Computer Typesetting Conf., Sussex, England, Pyke, T. N., Jr., Computer Technology: A Forward Look. NBS July 14-18, 1966, Ed. W. P. Jaspert, pp. 85-87 (The Institute Tech. News Bull. 51, No. 8, 161-163 (Aug. 1967). Also in Yale of Printing, London, 1967). Scientific, pp. 14-15, Oct. 1967. Peterson, G. E., Research on Speech Analysis, Rept. No. AFOSR J11 62, 24 p. (Univ. of Michigan, Ann Arbor, Oct. 1963). Quirk, W. B., Communications ServicesPresent and Future, Peterson, G. E., Studies in Speech Analysis and Synthesis, AFIPS Proc. Fall Joint Computer Conf., Vol. 31, Anaheim, Final Report, 46 p. (Communications Sciences Lab., Univ. Calif., Nov. 14-16, 1967, pp. 520-522 (Thompson Books, of Michigan, Ann Arbor, Aug. 1966). Washington, D.C., 1967). Petrick, S. R. and H. M. Willett, A Method of Voice Communi- cation with a Digital Computer, Proc. Eastern Joint Computer Rabinow, J., Optical Character Recognition, Data Proc. Mag. 8, Conf., Vol. 18, New York, N.Y., Dec. 13-15, 1960, pp. 11-24 No. 1, 18-24 (Jan. 1966). (Pub. by Eastern Joint Computer Conf., New York, 1960). Rabinow, J. and A. W. Holt, (assignors to Control Data Corpora. Pfaltz, J. L. and A. Rosenfeld, Computer Representation of tion), Reading System with Dictionary Look-Up, U.S. Patent Planar Regions by Their Skeletons, Commun. ACM 10, 3,259,883, issued July 5, 1966, filed Sept. 18, 1961, 11 p. No. 2, 119-122 (Feb. 1967). Rabinowitz, G., Current Status of Automatic Scanning at Brook- Pfaltz, J. L., J. W. Snively, Jr., and A. Rosenfeld, Local and haven National Laboratory and Columbia University, in Global Picture Processing by Computer, in Pictorial Pattern Pictorial Pattern Recognition, Proc. Symp. on Automatic Recognition, Proc. Symp. on Automatic Photointerpretation, Photointerpretation, Washington, D.C., May 31 -tune 2, 1967, Washington, D.C., May 31June 2, 1967, Ed. G. C. Cheng Ed. G. C. Cheng et al., pp. 159-173 (Thompson Book Co., et al., pp. 353-371 (Thompson Book Co., Washington, D.C., Washington, D.C., 1968). 1968). Rao, P. S. R. S., On Selecting Variables in Succession for Best Philbrick, 0., Shape Description with the Medial Axis Trans- Discrimination, Tech. Note No. 6, 14 p. (Information Research formation, in Pictorial Pattern Recognition, Proc. Symp. on Associates, Inc., Lexington, Mass., Dec. 29_1966). Automatic Photointerpretation, Washington,D.C., D.C. May 31 Rath, G. J. and D. J. Werner, Infosearch: Studying the Remote June 2, 1967, Ed. G. C. Cheng et al., pp. 395-407 Use of Libraries by Medical Researchers, in Levels of Inter- Book Co., Washington, D.C., 1968). action Between Man and Information, Proc. Am. Doc. Inst., Philipson, H. L.,L. Jr., Optical Character Recognition: The Input Annual Meeting, Vol. 4, New York, N.Y., Oct. 22-27, 1967, Answer, in, Processing, Vol. X, Proc. 1966 Int. Data pp. 58-62 (Thompson Book Co., Washington, D.C., 1967). -Processing Conf., Chicago, III., June 21-24, roa-,ifi. 119-=130 Ray, L. C. and R. A. Kirsch, Finding Chemical Records with (Data Processing Management Assoc., 1966). Digital Computers, Science 126, 814 (1957). .Phillips, E. T. A., The Use of Telephone Circuits for Facsimile Reagan, F. H.,H. Jr., Data Communications What It's All About, Tests on kc/s Group Circuits for Newspaper Page Trans- Data Proc. 8, 20-24, 26, 66-67 (Apr. 1966). mission, Mtiirhead Technique, A Journal of Instrument Engi- Reddy, D. R., Segmentation of Speech Sounds, J. Acoust. Soc. neering 16, No. 2, 11-13, 16 (Apr. 1962). Sii1 -312 (1966). Phonetics Laboratory, University College, London, Progress Reddy, D. R.R., Pitch Period Determinalion of Speech Sounds, Report, July 1963, 36 p. Commun. 10, No. 6, 343-348 (June 1967). 13148 Grants '$1,239',910 for NOvel'SYiem to Identify Micro- Reinfelds, J., L. A. Flenker, R. N. Seitz and P. L. Clem, Jr., Organisms, The Am. J. Med. Electronics 4, No. 3, 7-9 (1965). AMTRAN, A Remote-Terminal Conversational-Mode Com- Pickett, J. M., Transmitting Speech Sounds by a Tactual Vocoder puter System, Proc. 21st National Conf., ACM, Los Angeles, and by Lip-Reading, Rept. No. 27, 35 p. (The Speech Trans- Calif., Aug. 30Sept. 1, 1966, pp. 469-477 (Thompson Book Co., mission Lab., Div.of Telegraphy-Telephone, The Royal Washington, D.C., 1966). Institute of Technology, Stockholm, Sweden, Mar. 1963). Research on Pattern Recognition, Self-Correcting-Circuits and Pierce, J. R., The Transmission of Computer Data, Scient. Am. Adaptive Systems, Institut fiir Nachrichtenverarbeitung and 215, 145-156 (1966). NachrichteniibertragungderTechnischen Hochschule Pipberger, H. V., Computer Analysis of the Electrocardiogram, Karlsruhe, Karlsruhe, Federal Republic of Germany, Feb. in Computers in Biomedical Research, Vol. I, Ed. R. W. Stacy 1966, 57 p.

161 Reynolds, C. H. Programming Languages, Data Proc. Mag. 9, Sakai, T. and S. Doshita, The Automatic Speech Recognition 28-29 (July 1967). System for Conversational Sound, IEEE Trans. Electron. Richardson, J. M., Theory of Property Filtering in Pattern Computers EC-12, 835-846 (Dec. 1963). Recognition, Rept. No. RADC-TR-66-531, 164 p. (Rome Salton, G., A Combined Program of Statistical and Linguistic Air Development Center, Griffiss Air Force Base, New York, Proceduresfoi Automatic Information Classification and Sept. 1966). Selection, in Automation and Scientific Communication, Short Rippy, D. E., D. E. Humphries and J. A. Cunningham, MAGIC - Papers, Pt. 1, papers contributed to the Theme Sessions of A Machine for Automatic Graphics Interface to a Computer, the 26th Annual Meeting, Am. Doc. Inst., Chicago, Ill., Oct. 6- AFIPS Proc. Fall Joint Computer Conf., Vol. 27, Pt. 1, Las 11, 1963. Ed. H. P. Luhn, pp. 53-54 (Am. Doc. Inst., Wash- Vegas, Nev. Nov. 30-Dec. 1, 1965, pp. 819-830 (Spartan ington, D.C., 1963). Books, Washington, D.C., 1965). Salton, G., Ed., Information Storage and Retrieval, Scientific Risberg, A., Fundamental Frequency Tracking, Offprint, Proc. Rept. No. ISR-7 (Computation Lab., Harvard Univ., Cambridge, Int. Congress of Phonetic Sciences, 1961, pp. 228-231 (Mouton Mass., June 1964). & Co., 's Gravenhage, 1962). Samuel, A. L., Time-Sharing on z Multiconsole Computer, Rept. Roberts, K. V., The Readability of Computer Programs, The No. MAC-TR-17, 23 p. (M.I.T., Cambridge, Mass., Mar. 1965). Computer Bull. 10, 17-24 (Mar. 1967). Sayre, K. M., Human and Mechanical Recognition, Methodos 14, Roberts, L. G., Pattern Recognition by Adaptive Networks, IRE No. 54, 27-40 (1962). Int. Cony. Rec. 8, Pt. 2, 66-70 (1960). Schatz, S., Facsimile Transmission in Libraries: A State of the Roberts, L. G., Picture Coding Using Pseudo-Random Noise, Art Survey, Tech. Memo. No. 1, 24 p. (Information Systems IRE Trans. IT-8, 145-154 (Feb. 1962). Office, Library of Congress, Washington, D.C., June 1967). Roberts, L.G., Recent Development in Optical Character Scheele -, G., Ed., Information Retrieval-A Critical View, 282 p. Recognition a M.I.T., in Optical Character Recognition, (Thompson Book Co., Washington, D.C., 1967). Ed. G. L. Fischer, Jr., et al., pp. 209-212 (Spartan Books, Schreiber, W. F., T. S. Huang and 0. J. Trctiak, Contour Coding of Washington, D.C., 1962). Images, 1968 WESCON Technical Papers, 8/3, Aug. 1968, 6 p. Roberts, L. G., Graphical Communication and Control Languages, Schroeder, M. R. and E. E. David, A Vocoder for Transmitting in Second Congress on the Information System Sciences, held 10 KC Speech Over a 3.5 KD Channel, Acoustia 10 (1960). at The Homestead, Hot Spria,gs, Va., Nov. 1964, Ed. J. Spiegel Schroeder, M. R., J. L. Flanagan and E. A. Lundry, Bandwidth and D. E. Walker, pp. 211-217 (Spartan Books, Washington, Compression of Speech by Analytic-Signal Rooting, Proc. D.C., 1965). IEEE 55, 396-401 (Mar. 1967). Roberts, L. G., Machine Perception of Three-Dimensional Solids, Schultz, G.L., Character Recognition Machine, U.S. Pat. in Optical and Electro-Optical Information Processing, Ed. 3,074,050, issued Jan. 15, 1963. J. R. Tippett et al., pp. 159-197 (M.I.T. Press, Cambridge, Schwartz, J. I., E. G. Coffman and C. Weissman, Potentials of Mass., 1965). Large-Scale Time-Sharing System, in Second Cong. On the Roberts, L. G., The Lincoln Wand, AFIPS Proc. Fall Joint Information System Sciences, held at The Homestead, Hot Computer Conf., Vol. 29, San Francisco; Calif., Nov. 7-10, Springs, Va., Nov. 1964, Ed. J. Spiegel and D. E. Walker, 1966, pp. 223-227 (Spartan Books, Washington, D.C., 1966). pp. 15-32 (Spartan Books, Washington, D.C., 1965). Rochester, N., J. R. Johnson, G. M. Amdahl and W. E. Mutter, Schwartz, R. E., System Considerations in Automated"Fingerprint assignors to IBM Corp., New York. Recognition of Recorded Classification, in Law Enforcement Science and Technology, Intelligence, U.S. Pat. 2,889,535, patented June 2, 1959; Vol. 1, Proc. First National Symp. on Law Enforcement Science filed Oct. 20, 1955. Class 340-149. and Technology, Chicago, Ill., Mar. 1967, Ed. S. A. Yefsky, Rockwell, M. A., Jr., H. Shubin, M. H. Well and P. F. Meagher, pp. 511-515 (Thompson Book Co., Washington,D.C., 1967). SHOCK III, A Computer System as an Aid in the Management Sebestyen, G. S. and D. Van Meter, Investigation of Automation of Critically Ill Patients, Commun. ACM 9, No. 5, 355-357 of Speech Processing for Voice Communication, Rept. No. (May 1966). AFCRL-62-946, 81 p. (Data Systems Div., Litton Systems, Rose, G. A., light-Pen' Facilities for Direct View Storage Inc., Waltham, Mass., 1962). Tubes-An Economical Solution for Multiple Man-Machine Sebestyen, G. S. and J. L. Edie, Pattern Recognition Research, Communication, IEEE Trans. Electron. Computers EC-14, Final Rept. AFCRL-64-821, Contract AF 19(628)1604, 120 p. 637-639 (Aug. 1965). (Litton Systems, inc., Waltham, Mass., June 14, 1964). Rosenfeld, A., Automatic Recognition Techniques Applicable to Selfridge, 0. G., Pattern Recognition and Modern Computers, High-Information Pictorial Inputs, 1962 IRE Int. Cony. Rec., Proc. Western Joint Computer Conf., Vol. 7, Los Angeles, Pt. 4, pp. 114-123. Calif., Mar. 1-3, 1955, pp. 91-93 (Institute of Radio Engineers, Rosenfeld, A., Picture Processing by Computer, Rept. No. 68-71, New York, 1955). 1 v. (Computer Science Center, Univ. of Maryland, College Senechalle, D., Experiments with a New Classification Algorithm, Park, June 1968). Rept. No. LRC-64-WTM-6, 1 v. (Linguistics Research Center, Rosenfeld, A. and J. L. Pfaltz, Sequential Operations in Digital Texas Univ., Austin, Dec. 1964). Picture Processing, J. ACM 13, No. 4, 471-494 (Oct. 1966). Shafritz, A. B., The Use of Computers in Message Switching Rothauser, E. H., The Integrated Vocoder and Its Application in Networks, Proc. 19th National Conf., ACM, Philadelphia, Pa., Computer Systems, IBM J. Res. & Dev. 10, No. 6, 455-461 Aug. 25-27, 1964, pp. N2.3 -1 to N2.3-6 (Assoc. for Computing (Nov. 1966). Machinery, New York, 1964). Rubinoff, M., Ed., Toward a National Information System, Shelman, C. B., Machine Classification of Fingerprints, in Law Second Annual National Colloquium on Information Retrieval, Enforcement Science and Technology, Vol. 1, Proc. First Philadelphia, Pa., April 23-24, 1965, 242 p. (Spartan Books, National Symp. on Law Enforcement Science and Technology, Washington, D.C., 1965). Chicago, Ill., Mar. 1967, Ed. S. A. Yefsky, pp: 467-477 Rubio, J. E., The Clustering and Recognition of Handwritten (Thompson Book Co., Washington, D.C., 1967). Numbers, Rept. No. 1288 (Polytechnic Institute of Brooklyn, Shepard, D. H., assignor to Intelligent Machines Research Corp., Brooklyn, N.Y., July 1965). Arlington, Va., Apparatus for Reading, U.S. Pat. 2,663,758; Rushforth, C. K., Recognizing Patterns in Photographs, Applied patented Dec. 22, 1953; filed Mar. 1, 1951, Serial No. 213,338, Optics 4, 379-381 (1965). Class 178-17. Rutovitz, D., Data Structures for Operations of Digital Images, Shephard, R. N., Application of IPLIV to the Simulation of in Pictorial Pattern Recognition, Proc. Symp. on Automatic Perceptual Learning, paper presented at the Conf. on Informa- Photointe .pretation, Washington, D.C., May 31-June 2, 1967, tion Processing Languages for Digital Computers: A Technique Ed. G. C. Cheng et al., pp. 105-133 (Thompson Book Co., for Model Building, Amer. Psychological Assoc., Cincinnati, 0., Washington, D.C., 1968). Sept. 1959, preprint 5 p. (Bell Telephone Labs., Murray Hill, N.J., 1959). Sakai, T. and S. Doshita, Phonetic Typewriter, in Information Shepherd, M. J. and A. J. Willmott, Cluster Analysis on the Atlas Processing 1962, Proc. IFIP Congress 62, Munich, Aug. 27- Computer, The Computer J. 11, 57-62 (May 1968). Sept. 1, 1962, Ed. C. M. Popplewell, pp. 445-455 (North- Sherman, H., A Quasi-Topological Method for the Machine Holland Pub. Co., Amsterdam, 1%3). Recognitior of Line Patterns, in Information Processing;

162 Proc. Int. Conf. on Information. processing, UNESCO, Paris, Stevens, M. E., The Interconnection of Two DigitalComputers, June 15-20, 1959, pp. 232-237; addendum, pp. 237-238; Conf. Paper CP-55-736, American Institute of Electrical discussion, p. 238 (UNESCO, Paris; Oldenbourg, Munich; Engineers, 1955. Butterworths, London, 1960). Preprint, 18 p. (Lincoln Lab., Stevens, M. E., Abstract Shape Recognition by Machine, in M.I.T., Lexington, Mass., May 7, 1959). ComputersKey to Total Systems Control, Proc. Eastern Shiner, G., User Requirements, in Optical Character, Ed. G. L. Joint Computer Conf., Vol. 20, Washington, D.C., Dec. 12-14, Fischer, Jr., et al., pp. 335-338 (Spartan Books, Washington, 1961, pp. 332-351 (The Macmillan Co., New York, 1961). D.C., 1962). Stevens, M. E., Automatic Character Recognition, A State-of- Shonnard, J. R., High-Speed Communication of Graphic Intelli- the-Art Report, NBS Tech. Note 112, 168 p. (U.S.Govt. gence with Hard Copy Readout,Commun. & Electronics, Printing Office, Washington, D.C., May 1961). No. 61, 176-178 (July 1962). Stevens, M. E., Automatic Indexing: A State-of-the-ArtReport, Simek, J. G. and C. J. Tunis, Handprinting Input Device for NBS Monograph 91, 220 p. (U.S. Govt. Printing Office,Wash- Computer Systems, IEEE Spectrum 4, 72-81 (July 1967). ington, D.C., Mar. 30, 1965). Simmons, R. F., Natural Language Processing and the Time- Stevens, M. E., Nonnumeric Data Processing in Europe: A Field Shared Computer, in Toward a National Information System, Trip Report, AugustOctober 1966, NBS Tech. Note 462, 63 p. Second Annual National Colloquium on Information Retrieval, (U.S. Govt. Printing Office, Washington, D.C., Nov. 1968). Philadelphia, Pa., April 23-24, 1965, Ed. M. Rubinoff, pp. 217- Stevens, M. E., V. E. Giuliano and L. B. Heilprin, Eds., Statistical 227 (Spartan Books, Washington, D.C., 1965). Association Methods for Mechanized Documentation, Symp. Simms, R. L., Jr., Trends in Computer/Communication Systems, Proc., Washington, D.C., Mar. 17-19, 1964, NBS Misc. Pub. Computers & Automation 17, No. 5, 22-25 (May 1968). 269, 261 p. (U.S. Govt. Printing Office, Washington, D.C., Simonton, W., Ed., Information Retrieval Today, papers presented Dec. 15, 1965). at the Institute conducted by the LibrarySchool and the Center Stroke, G. W., Lensless Photography, Int. Sci. & Tech. 41, for Continuation Study, Univ. of Minnesota, Sept. 19-22,1962, 52-60 (1965). Strollo, T. R., Pictorial Input System for a DigitalComputer 176 p. (Minnesota Univ., 1963). Research, Sims, S. P., Tile Ccmputer and the Law, Data Proc. Mag.8, Information System Summaries, 73 p. (Office of Naval Washington, D.C., Sept. 1966). 20-26 (Juiy 1%6). Strom, R., Methodology for Research in Concept-Learning, in Singer, J. R., A Self-Organizing Recognition System, Proc. Kochen, Western Joint Computer Conf., Vol. 19, Los Angeles, Calif., Some Problems in Information Science, Ed. M. Conf., pp. 105-116 (The ScarecrowPress, Inc., New York, 1965). May 9-11, 1961, pp. 545-554 (Western Joint Computer Sublette, I. H. and J. Tults, Character Recognition byDigital Glendale, Calif., 1961). Programming, Feature Detection, RCA Review, March 1962. Smith, F. W., Pattern Classifier Design by Linear Sutherland, I. E., Sketchpad, A Man-Machine Graphical Com- IEEE Trans. Computers C-17, 367-372 (Apr. 1968). munication System, AFIPS Proc. Spring Joint ComputerConf., Sparck-Jones, K. and D. Jackson, Current Approaches to Books, Classification and Clump-Finding at the Cambridge Language Vol. 23, Detroit, Mick, May 1963, pp. 329-346 (Spartan Baltimore, Md., 1963). Research Unit, The Computer J. 10, 29-37 (May 1967). Sutherland, W. R., Language Structure and GraphicalMan- Specialized Computational Equipment from General-Purpose Machine Communication, in Information System Scienceand Modules, NBS Tech. News Bull. 51, 175-176 (Aug. 1967). Technology, papers prepared for the Third Cong., scheduled Spilerman, S., A Recognition Logic Designed for Hand-Printed (Thompson Character, WESCON Cony. Rec., Vol. 6, Aug. 21-24, 1962. for Nov. 21-22, 1966, Ed. D. E. Walker, pp. 29-31 Book Co., Washington, D.C., 1967). Spinrad, R. J., Machine Recognition of Hand Printed Block Sutter, H., Method and Apparatus for ReadingHandwritten Letters, Ph. D. Thesis (M.I.T., Cambridge, Mass., 1963). Symbols, Particularly Numerals, U.S. Pat. 2,928,074;patented Spinrad, R. J., Machine Recognition of Hand Printing, Inf. & Mar. 8, 1960; filed Sept. 27, 1955, Class 340-149. Control 8, 124-142 (Apr. 1965). Swanson, R., Information Sciences 1965, Rept. No.AFOSR Spolsky, B., Some Problems of Computer-Based Instructions, 66-0130, 109 p. (Office of Aerospace Research, USAF, Wash- Behay. Sci. 11, 487-496 (Nov. 1966). ington, D.C., Jan. 1966). S prick, W. and K. Ganzhorn, An Analogous Method forPattern Swanson, R. W., Information System NetworksLet's Profit Recognition by Following the Boundary, in Information from What We Know, in Information RetrievalACritical Processing, Proc.Int.Conf. on Information Processing, View, Ed. G. Schecter, pp. 1-52 (Thompson Book Co.,Wash- UNESCO, Paris, June 15-20, 1959, pp. 238-244 (UNESCO, Paris; Oldenbourg, Munich; Butterworths, London, 1%0). ington, D.C., 1967). Discussion by M. S. Maxwell, ibid., p. 244; Discussion by Talbert, L. R., G. F. Groner, J. S. Koford, R. J. Brown, P. R. Low J. C. R. Licklider, ibid., p. 244. and C. H. Mays, A Real-Time Adaptive Speech Recognition Spring, W. C., Jr., Applications in Medicine, in Annual Review of System, Rept. No. ASDTDR-63-660 (StanfordElectronics Information Science and Technology, Vol. 2, Ed. C. A. Cuadra, Laboratory, May 1963). pp. 311-338 (Interscience Pub., NewYork, 1967). Tanimoto, T. T., An Elementary Mathematical Theory ofClassifi- Stacy, R. W. and B. Waxman, Eds., Computers in Biomedical cation and Prediction (IBM Corp., New York, 1958). Research, Vol. I, 562 p. (Academic Press, New York, 1965). Tanimoto, T. T., The General Problem of Classification and Stark, L., J. F. Dickson, G. H. Whipple and H. Horibe,Remote Indexing, in Machine Indexing: Progress and Problems, papers Real-Time Diagnosis of Clinical Electrocardiograms by a presented at the Third Institute on Information Storage and Digital Computer System, Ann. N.Y. Acad. of Sci. 126, 851 Retrieval, American Univ., Washington, D.C., Feb. 13-17, (1965). 1961, pp. 233-235 (American Univ., Washington, D.C.,1962). Stark, L., M. Okajima and G. H. Whipple, Computer Pattern Tappert, C. C., On the Neural Modeling of Speech Processes, RecognitionTechniques:ElectrocardiographicDiagnosis, Cognitive Systems Research Program Rept. No. 9(Cornell Commun. ACM 5, No. 10, 527-532 (Oct. 1962). Univ., Ithaca, N.Y., Dec. 15, 1966). Stein, E. Miscellaneous Notes on Character Recognition, un- Tate, F. A., Handling Chemical Compounds in Information published,no pagination, Submitted to theNational Bureau Systems, in Annual Review of Information Scienceand Tech- of Standards, 1965. nology, Vol. 2, Ed. C. A. Cuadra, pp. 285-309 ( Interscience Stein, P. G., Transistorized Building Blocks for Data Instrumenta- Pub., New York, 1967). tion, NBS Tech. Note 268, 118 p. (U.S. Govt. Printing Office, Taube, M. and H. Wooster, Eds., Information Storageand Washington, D.C., May 28, 1966). Retrieval Theory Systems, and Devices, Air Force Office of Steinbuch, K. and U. A. W. Piske, Learning Matrices andTheir Scientific Research Symp., Washington, D.C., 1958, 228 p. Applications, IEEE Trans. Electron. Computers EC-12, 846- (Columbia Univ. Press, New York, 1958). 862 (Dec. 1963). Teager, H., Multidimensional Visual Information Processing. Stephan, P. L., Soviet Research in AutomaticCharacter Recogni- Conf., Advances Biomedical Computer Applications, New tion: A Survey of Organizations, in SovietCybernetics: Recent York Academy of Science, N.Y., June 3-5, 1965. News Items No. 3, Ed. W. B. Holland, pp. 5-16 (TheRAND Teager, H. M., Quarterly Progress Report No. 80, ResearchLab. Corp., Santa Monica, Calif., Apr. 1967). for Electronics, M.I.T., 269 (1966). 163 Teitelman, W., Real Time Recognition of Hand-Drawn Char- Ungeheuer, G., R. Rupprath and L. Freidrich, Zur Ent wick- acters, AFIPS Proc. Fall Joint Computer Conf., Vol. 26, lungeinesVerbund SystemsvonPeriodizatiitsanalysator San Francisco, Calif., Oct. 1964, pp. 559-575 (Spartan Books, (Tonhiihenschriebe), und Intersimeter, Paper J-11. Fifth Baltimore, Md., 1964). International Congress on Acoustics, Liege, Belgium. Sept. Teitelman. W., PILOT: A Step Toward Man-Computer Symbiosis, 7-14. 1965, 3 p. Project MAC, Rept. No. MAC-TR-32, 193 p. (M.I.T., Cam- Unger, S. H.. A Computer Oriented Toward Spatial Problems, bridge, Mass., Sept. 1966). Proc. IRE 46, No. 10. 1744-1750 (Oct. 1958). A New Type of Teixeira, J. F. and R. P. Sallen, The Sylvania Data Tablet: Computer Oriented Toward Spatial Problems, Proc. Western A New Approach to Graphic Data Input, AFIPS Proc. Spring Joint Computer Conf., Vol. 13, Los Angeles, Calif., May 6-8. Joint Computer Conf., Vol. 32, Atlantic City, N.J., Apr. 30- 1958, pp. 234-239 (American Inst. of Electrical Engineers, May 2, 1968, pp. 315-321 (Thompson Book Co.. Washington, New York. Mar. 1959). D.C., 1968). Unger. S. H., Pattern Detection and Recognition, Proc. IRE 47, Thiebault, R., Automatic Process for the Identification of Finger- No. 10, 1737-1752 (Oct. 1959). prints, Proc. Int. Symp. on Automation of Population Register United Nations Educational Scientific and Cultural Organization, Systems, Vol.1, Jerusalem, Israel, Sept. 25-28, 1%7, Ed. InformationProcessing, Proc.Int.Conf. on Information A. Kistler, pp. 207-226. Processing, Paris, June 15-20, 1959, 520 p. (UNESCO. Paris; Thomas, R. B. and M. Kassler, Character Recognition in Context. Oldenbourg, Munich; Butterworths, London, 1960). Inf. & Control 10, No. 1, 43-64 (Jan. 1%7). United States Air Force, Office of Aerospace Research, Research Tillman, H. G., Zur Klassifikation der Individuungebundenen Objectives '65, Rept. No. OAR-65-3, 66 p. (U.S.A.F., OAR, Merkmale am Sprachschall, Paper A-47, Fifth International Washington, D.C., June 1965). Congress on Acoustics, Liege, Belgium, Sept. 7-14, 1965, 4 p. Urquhart, A. B., Voice Output from IBM System/360, AFIPS Tillman, H. G., G. Heike, H. Schelle and G. Ungeheuer, Proc. Fall Joint Computer Conf., Vol. 27, Pt. 1, Las Vegas, DAWID I- Ein Beitrag zur Automatischen `Spracherkennung', Nev., Nov. 30-Dec. 1. 1965, pp. 857-865 (Spartan Books, Paper A-12, Fifth International Congress on Acoustics, Liege, Washington, D.C., 1965). Belgium, Sept. 7-14, 1965, 4 p. Uyehara, G. U., A Stream-Following Technique for use in Char- Tippett, J. R., D. A. Berkowitz, L. C. Clapp, C. J. Koester and acter Recognition, IEEE Int. Cony. Rec., Pt. 4, 1963. A. Vanderburgh, Jr., Eds., Optical and Electro-Optical Informa- Van Dam, A. and J. C. Michener, Hardware Developments and tion Processing, 780 p. (M.I.T. Press, Cambridge, Mass., 1%5). Product Announcements, in Annual Review of Information Titus, J. P., Computers, Communications, and the FCC. Commun. Science and Technology, Vol. 2, Ed. C. A. Cuadra, pp. 187-222 ACM 10, 62-67 (Jan. 1%7) (Interscience Pub., New York, 1967). Tobler, W. R., Computer Use in Geography, Behay. Sci. 12, Vander Lugt, A., F. B. Rotz and A. Klooster, Jr., Character- 57-58 (Jan. 1967). Reading by Optical Spatial Filtering, in Optical and Electro- Tobler, W. R., Problems and Prospects in Geographical Photo- Optical Information Processing, Ed. J. R. Tippett etal., interpretation, in Pictorial Pattern Recognition, Proc. Symp. pp. 125-141 (M.1. T. Press, Cambridge, Mass., 1955). on Automatic Photointerpretation, Washington, D.C., May 31- Van Emden. B. M., Advanced Computer Based Fingerprint June 2, 1%7, Ed. G. C. Cheng et al., pp. 267-273 (Thompson Automatic Classification Technique (FACT), in Law Enforce. Book Co., Washington, D.C., 1968). ment Science and Technology, Vol. 1. Proc. First National Tou, J. T., Ed., Computer and Information Sciences-11, Proc. Symp. on Law Enforcement Science and Technology, Chicago, 2nd Symp. on Computer and'Information Sciences, Columbus, Ill., Mar. 1967, Ed. S. A. Yefsky, pp. 493-505 (Thompson Book 0., Aug. 22-24, 1966, 368 p. (Academic Press, New York, 1%7). Co., Washington, D.C., 1967). Tou, J. T. and R. P. Heydorn, Some Approaches to Optimum Van Geffen, L. M. H. J., A Review of Keyboarding Skills, in Feature Extraction, in Computer and Information Sciences- Advances in Computer Typesetting, Proc.Int. Computer Proc. 2nd Symp. on Computer and Information Sciences, Typesetting Conf., Sussex, England, July 14-18, 1966, Ed. Columbus, 0., Aug. 22-24, 1966, Ed. J. T. Tou, pp. 57-89 W. P. Jaspert, pp. 2-11 (The Institute of Printing, London, (Academic Press, New York, 1967). 1967). Tou, J. T. and R. H. Wilcox, Eds., Computer and Information Van Heerden, P. J., Computers of the Future, Proc. IRE 50, Sciences, collected papers on learning, and adaptation and No. 5. 621 (May 1962). control in information systems, 544 p. (Spartan Books, Wash- Veaner, A. B., Developments in Copying Methods and Graphic ington, D.C., 1964). Communication 1965, Lib. Res. & Tech. Serv. 10, 199-210 Trabka, E. A. and P. G. Roetling, Image Transformations for (Spring 1966). PatternRecognitionUsing IncoherentIlluminationand Vernon, H. C. and R. C. Walsh, Character Recognition Method Bipolar Aperture Masks, J. Opt. Soc. Amer. 54, 1241-1252 and Apparatus, U.S. Patent No. 3.058,093, issued Oct. 9, 1962. (Oct. 1964). Vitale, H. S., A Study of Character Recognition Parameter Transparent-Mode Control Procedures for Data Communication, Analysis, RADC-TR-65-282, Final Rept. on Contract AF Using the American Standard Code for Information Inter- 30(602)3488, 70 p. (Computer Lab., General Electric Co., change-A Tutorial, Commun. ACM 8, 203-214 (April 1965). Sunnyvale, Calif., Sept. 1965). Vleduts, G. E., Ts. D. Osipenko and I. Ya. Pappe, Automating Uffelman, M. R., CONFLEX I-A Conditioned Reflex System, the Compilation of Formula Indicators of Chemical Compounds, 1962 IRE Int. Coin. Rec., Pt. 4, pp. 132-142. Nauch.-Tekh. Inform. No. 6, 13-17 (1963). Uffelman, M. R., W. Holford, J. Glenn, 0. Turner, [Scope, Inc./9 Voelker, C. S., Future Requirements for Automated Fingerprint Effects of Noise in Character Recognition Systems, Rept. Identification, in Law Enforcement Science and Technology, No. RADC-TR-66-720, 80 p. (Rome Air Development Center, Vol. 1, Proc. First National Symp. on Law Enforcement Science Griffiss AFB, N.Y., Jan. 1967). and Technology. Chicago, Ill., Mar. 1967, Ed. S. A. Yefsky, Uhr, L. and C. Vossler, A Pattern Recognition Program That pp. 445-447 (Thompson Book Co., Washington, D.C., 1967). Generates, Evaluates, and Adjusts Its Own Operators, Proc. Western Joint Computer Conf., Vol. 19, Los Angeles, Calif., Voltin, I. V., Hardware Considerations For Informations Systems May 9-11, 1961, pp. 555-569 (Western Joint Computer Conf., and Operations Analysis, in Information System Science and 1961). Technology, papers prepared for the Third Cong., scheduled Uhr, L. and C. Vossler, The Search to Recognize, in Optical for Nov. 21-22, 1966, Ed. D. E. Walker, pp. 7-9 (Thompson Character Recognition, Ed. G. L. Fischer, Jr., et al., pp. 319- Book Co., Wa,hington, D.C., 1967). 329 (Spartan Books, Baltimore, Md., 1962). Von Keller, T., Die Kennzeichnungvon Sprachlauten lurch Ullmann, J. R., Associating Parts of Patterns, Inf. & Control 9, Sprktrum, Autokorrelationsfunktion und Nulldurghgangsab- 583-601 (Dec. 1966). stHnde, Zur Erlangung des Akademischen Grades Cines Ullmann, J. R., Algebraic Inference of Pattern Similarity, The Doktor-Ingenieurs von der Fakultat rtir Elektrotechnik der Computer J. 10, 256-264 (Nov. 1967). Technischen Hochschule Karlsruhe genehmigte Dissertation, Ungeheuer, G., Ein Einfaches Varfahren zur Akustichen Klassi- Karlsruhe, June 1966, 85 p. fikationvon Sprechern, Paper A-17, FifthInternational Vossler, C. M. and N. M. Branston, The Use of Context for Cor- Congress on Acoustics, Liege, Belgium, Sept. 7-14, 1965 3 p. recting Garbled English Text, Proc. 19th National Conf. AFM,

164 Philadelphia, Pa., Aug. 25-27, 1964, pp. D2.4-1 to D2.4-12 Wells, M. B., Recent Improvements in MADCAP, Commun. (Assoc. for Computing Machinery, New York. 1964). ACM 6, N. 11, 674-678 (Nov. 1963). Wagner, F. V. and J. Granholm, Design of a General-Purpose Wells, R. S., Mapping the Moon by Radar, Data Proc. Mag. 9, Scientific Computing Facility, in Information Processing 1965, 26-30, 32 (Nov. 1%7). Proc. IFIP Congress 65, Vol. 1, New York, N.Y., May 24-29, Westin, A. F., Ed., Privacy and Freedom, 487 p. (Atheneum, 1965, Ed. W. A. Kalenich, pp. 283-289 (Spartan Books, N.Y., 1967). Washington, D.C., 1965). Whipple, G. H., J. F. Dickson, H. Hirobe and L. Stark, Remote, Waldo, W. H. and M. De Backer, Printing Chemical Structures Online, Real-Time Computer Diagnosis of the Clinical Electro- Electronically: Encoded Compounds Searched Generically cardiogram, Commun. ACM 8, 49-52 (Jan. 1965). with IBM-702, in National Academy of Sciences National White, H. E.,Printed English Compression by Dictionary Research Council, Proc. Int. Conf. on Scientific Information, Encoding, Proc. IEEE 55, 390-396 (Mar. 1%7). Vol.1, Washington, D.C., Nov. 16-21, 1958, pp. 711-730 White, H. S., DAPRDigital Automatic Pattern Recognition (NASNRC, Washington, D.C., 1959). for Bubble Chambers, in Pictorial Pattern Recognition, Proc. Walker, D. E., Ed., Information System Science and Technology Symp. on Automatic Photointerpretation, Washington, D.C., papers prepared for the Third Cong., scheduled for Nov. 21-22, May 31June 2, 1967, Ed. G. C. Cheng et al., pp. 175-198 1966, 406 p. (Thompson Book Co., Washington, D.C., 1967). (Thompson Book Co., Washington, D.C., 1%8). Walter, W. G., The Living Brain, 311 p. (W. W. Norton, New White, L. M. and R. W. Jones, Infrared Fault-Finder Pinpoints York, 1963). Defects in Multilayer Boards, Electronics 38, No. 24, 96-98 Ward, J.E., Systems Engineering Problems in Computer- (1965). Driven CRT Displays for Man-Machine Communication, IEEE Wholey, J. S., The Coding of Pictorial Data. IRE Trans. Inf. Trans. Systems Sci. & Cybernetics SSC-3, 47-54 (June 1967). Theory IT-7, 99-104 (Apr. 1%1). Ward, J. H., Jr., Hierarchical Grouping to Optimize an Objective Wier, J. M., Communications and Information Retrieval, in Function, J. Amer. Stat. Assoc. 58, No. 301 (Mar. 1963). Toward aNationalInformation System, Second Annual Ward, J. H., Jr. and M. E. Hook, Application of an Hierarchical National Colloquium on Information Retrieval, Philadelphia, Grouping Procedure to a Problem of Grouping Profiles, Educ. Pa., April 23-24, 1%5, Ed. M. Rubinoff, pp. 105-109 (Spartan and Psych. Measurement 23, No. 1, 69-82 (Spring 1963). Books, Washington, D.C., 1965). Ware, W. H., Impressions of Interdata 65, Commun. ACM 8, Wiesner, J. B., Communication Sciences in a University Environ- No. 7, 472-474 (July 1965). ment, IBM J. Res. & Dec. 2, 268-275 (Oct. 1958). Watanabe, S., P. F. Lambert, C. A. Kulikowski, J. L. Buxton Wigington, R. L., Graphics as Computer Input and Output, and R. Walker, Evaluation and Selection of Variables in 1966 IEEE Int. Cony. Rec., Vol. 14, Pt. 3, Computers, IEEE Pattern Recognition, in Computer and Information Sciences Int. Cony., New York, N.Y., Mar. 21-25, 1966, pp. 86-90 IL. Proc. 2nd Symp. on Computer and Information Sciences, (The Inst. of Electrical and Electronics Engineers, New York, Columbus, 0., Aug. 22-24, 1966, Ed. J. T. Tou, pp. 91-122 1966). (Academic Press, New York, 1%7). Winder, R. 0., Threshold Logic in Artificial Intelligence, in Watts, T. L, Scanning and Measuring Photographs of Bubble Artificial Intelligence, Preprint of papers presented at Artificial Chamber Tracks Using a Computer Controlled Line Segment Intelligence Sessions, IEEE Winter General Meeting, New (`PEPR'), in Pictorial Pattern Recognition, Proc. Symp. on York, N.Y., Jan. 27Feb. 1, 1963, pp. 108-128 (Pub. by IEEE, Automatic Photointerpretation, Washington, D.C., May 31 New York, Jan. 1963). June 2, 1967, Ed. G. C. Cheng et al., pp. 207-220 (Thompson Wishner, R., The Role of Paper Tape and Optical Scanning Book Co., Washington, D.C., 1968). Computer Input in Textual Data Processing, Proc. 1965 Weaver, W., Science and Complexity, Amer. Scientist 36, Congress F.I.D., 31st Meeting and Congress, Vol. II, Wash- 536-544 (Oct. 1948). ington, D.C., Oct. 7-16, 1965, pp. 235-240 (Spartan Books, Wee, W. G. and K. S. Fu, An Adaptive Procedure for Multiclass Washington, D.C., 1966). Pattern Classification, IEEE Trans. Computers C-17, 178-182 Woo, P. W., A Proposal for Input of Hand-Drawn Information (Feb. 1968). to a Digital System, IEEE Trans. Electron. Computers EC-13, Weeks, R. W., Rotating Raster Character Recognition System, 609-611 (Oct. 1964). Presented at AIEE Pacific General Meeting, San Diego, Calif., Wooster, H., Long Range Research in the Information Sciences, Aug. 8-12, 1960, Conf. Paper No. CP 60-1005, 10 p. Rept. No. AFOSR-1571, Presentation at the Science and Wegstein, J. H. and J. F. Rafferty, Machine Oriented Fingerprint Engineering Symp., San Francisco, Calif., Oct. 3-4, 1961, Classification System, in Law Enforcement Science and 24 p. (U.S. Air Force, Office of Scientific Research, 1961). Technology, Vol. 1, Proc. First National Symp. on Law En- forcement Science and Technology, Chicago, Ill., Mar. 1967, Wooster, H., Long Range Research in Information Retrieval, Ed. S. A. Yefsky, pp. 459-465 (Thompson Book Co., Washing- in Information Retrieval Today, papers presented at the ton, D.C., 1967). Institute conducted by the Library School and the Center for Wegstein, J. H., J. F. Rafferty and W. J. Peneak, Matching Continuation Study, Univ. of Minnesota, Sept. 19-22, 1962, Fingerprints by Computer, NBS Tech. Note 466, 16 p. (U.S. Ed. W. Simonton, pp. 149-164 (Minnesota Univ., 1963). Govt. Printing Office, Washington, D.C., July 1968). Wyllys, R. E., Research in Techniques for Improving Automatic Weihrer, A. L., J. R. Whiteman and C. A. Caceres, The Com- Abstracting Procedures, Tech. Memo. No. TM-1087/000/01, puterized Electrocardiogram: A Model for Medical Signal 30 p. (System Development Corp., Santa Monica, Calif., Analysis, Proc. 22nd National Conf., ACM, Washington, D.C., Apr. 19, 1963). Aug. 29-31, 1967, pp. 289-295 (Thompson Book Co., Washing- ton, D.C., 1967). Yefsky, S. A., Ed., Law Enforcement Science and Technology, Welch, A. J., J. L. Devine and R. G. Loudon, Computer Appli- Vol. 1, Proc. First National Symp. on Law Enforcement Science cations in Biomedical Electronics Pattern Recognition Studies, and Technology, Chicago, Ill., Mar. 1967, 985 p. (Thompson AFIPS Proc. Spring Joint Computer Conf., Vol. 30, Atlantic Book Co., Washington, D.C., 1967). City, N.J., April 18-20, 1967, pp. 257-263 (Thompson Books, Washington, D.C., 1967). Zworykin, V. K. and L. E. Flory, assignors to Radio Corporation Wells, M. B., MADCAP: A Scientific Compiler for a Displayed of America, Apparatus for Indicia Recognition, U.S. Patent Formula Textbook Language, Commun. ACM 4, No. 1, 31-36 2,616,983; patented Nov. 4, 1952; filed Jan. 3, 1949, Class (Jan. 1961). 179-100.3.

165

U.S. GOVERNMENT PRINTING OFFICE : 1070 OL-344-107