Extraction of Typ ographic Elements from Outline
Representations of Fonts
Ariel Shamir and Ari Rapp op ort
Institute of Computer Science The Hebrew University of Jerusalem
Jerusalem Israel farik arirg cs huji ac il
Abstract
Digital typefaces for computer graphics and multimedia applications must be capable of supporting operations
such as font variations transformations deformations and blending A powerful implementation of such op
erations must rely on the inherent typographic attributes of the typeface However even today s most advanced
typeface representations support only geometric outline representations and basic font variations
In this paper we discuss high level typeface representations which we term Parametric Typ ographic Represen
tations PTRs We present an algorithm for automatical ly extracting typographic elements of typefaces from
their outline representation which is an essential initial step in converting typefaces from outline representa
tions to PTRs The extracted typographic elements include serifs bars stems slants bows arcs curve stems
and curve bars Most notable is the treatment of serifs which are represented by nite automata The algorithm
only needs to learn a serif type once and is then capable of automatical ly recognizing it in di erent typefaces
We show an application of a PTR for automatic high quality hinting of fonts which is one of the most
important stages in digital font production Our system was used to generate hints for dozens of thousands of
Kanji Roman and Hebrew characters
Keywords Digital typ ography outline fonts typ ographic elements hinting parametric typ ographic repre
sentations
Intro duction
Printed pages have constituted a ma jor way of communication b etween literate p eople ever since the invention
of print by Guten b erg in the th century The existence of electronic media has mo di ed this situation
Typsetting page layout typ efaces and fonts need to change their nature and b ecome more exible A wide
range of applications such as multimedia publishin g computer animation and even mechanical or electronic
computer aided design systems must incorp orate a diverse set of text manipulation functions from simple
a ne transformations to advanced deformations The basic technology which supp orts this functionality is
encapsulated in the representation used for digital typ ographic typ efaces or digital fonts Hence the degree
of sophisticati on of digital typ eface representations is a crucial factor in the expressive p ower and usability of
mo dern visual communication
In the design and analysis of typ efaces one uses terms such as stems bars serifs style size weight and
width Lawson Rubinstein Bringhurst Bauermeister see the gures in this pap er for examples
However even the advanced typ eface representations in use to day e g QuickDraw GX Apple TrueTyp e
and TrueTyp e Op en Microsoft Typ e Adob e Adob e concentrate only on representing the geometry
of a font and on the provision of relatively simple font variations There is a large gap b etween the high level
terms used by designers and b etween current representations A representation which supp orts high level
parameterization based on typ ographic terminology is needed The general term which we use for such a
representation is a Parametric Typographic Representation PTR Extraction of typ ographic elements from
existing font representations is essential in order to convert to day s outline representations into any parametric
typ ographic representation
Previous Work
Some previous work of identifying typ ographic elements has b een done for the sake of automating the pro cess
of hinting an outline font hints are needed in order to rasterize scaled fonts correctly whether on a screen
or in print Basic hints can b e added to the outline description automatically by recognizing horizontal and
vertical bars and curvilinear shap e extrema Andler or even some serif parts Karow
Hersch and B etrisey have presented an elab orate automatic hinting metho d Hersch a which requires a
top ological font indep endent mo del for the description of each Roman letter shap e These mo dels include
sp eci c hints connected to various typ ographic elements In order to add hints to a concrete letter shap e a
matching algorithm b etween the real shap e and the mo dels is p erformed After a match has b een found the
hints contained in the matched mo del are adapted to the real letter shap e by identifying similar p oints in
b oth the mo del and the actual letter shap e This scheme involves the de nition of a mo del for each top ology
of a letter shap e a pro cess which is b oth di cult and time and space consuming for example in ideographic
Chinese Japanese and Korean scripts
Several other techniques for automatic hinting of outlined characters exist inside the font development
systems of commercial companies such as Bitstream Adob e and Apple These systems are considered a
commercial secret but by examining their output e g TrueTyp e or Typ e fonts one can see that a large
amount of typ ographic information is still not identi ed by the automatic hinting pro cess and therefore a lot
of manual pro o ng and hinting is taking place in order to achieve the desired quality
Some research has b een done on the extraction of strokes from the outline description for applications such
as display of Kanji characters Chialing conversion to a di erent representation Durst a Durst b and
optical character recognition Feng The general spirit of these pap ers is opp osite to ours since they are
interested in removing the subtle typ ographic details of typ efaces in order to create a skeleton representation
of characters which is simpler and easier to recognize
A font represented in metafont Knuth is in principle parametric since metafont is a pro cedural
programming language However all the parameterization is done manually by the programmer Conversion
of PostScript fonts to metafont has b een describ ed in Haralamb ous but the metho d is sp eci c to these
two representations and it do es not treat automatic extraction of typ ographic elements
Contribution
In this pap er we present an algorithm to extract typ ographic elements of typ efaces from an outline repre
sentation of a font and show how to use its results for automatic hinting The conversion algorithm extracts
basic typ ographic elements from the outline description of a character Figure and can gather information
regarding the relationshi ps b etween them b oth inside each character and across the font This information
can then b e used to assign parametric attributes to the typ ographic elements With this capability the algo
rithm can b e viewed as an essential initial step in converting a typ eface from an outline representation into a
parametric typ ographic representation
Our conversion metho d involves two main stages The rst stage is the designation and classi cation of
characteristic p oints along the glyph outline which is similar in spirit to some previous work Hersch a
but is given here in a more detailed manner The second stage uses data from the rst stage for the actual
extraction of the typ ographic features from the outline Among the basic features recognized are stems and
bars b ows and arcs curve stems and curve bars slants extrema and most notably serifs of all typ es These
typ ographic features are in turn gathered to create higher degree elements such as groups of bars inside a
glyph or a Kanji stroke like element Zhang
Sp ecial treatment has b een given to the extraction of serifs Serifs play a crucially imp ortant role in the
design of a typ eface and there is an enormous diversity of serif designs We create a nite automaton Lewis
that de nes the sequence or sequences of p oints which characterize each typ e of serif Once this automaton
has b een de ned serifs of this typ e can b e extracted and recognized in any outlined font input to the system
If a new typ eface containing new typ es of serifs is intro duced to the system only a few new nite automata
need to b e de ned in order to extract the new typ eface features
Along with a raster to outline mo dule our system can convert a typ eface in any lower level description
technique to a higher level parametric typ ographic representation We do not require a mo del for each glyph
since the information regarding typ ographic elements is stored in a higher representational level that applies
to any glyph The algorithm is time and space e cient if n is the numb er of p oints de ning the outline of
a glyph and k is the numb er of features extracted from it where k n the pro cess for recognizing basic
2
elements for one glyph takes O n log n k time and O n k space
Our metho d needs much less manual manipulati on and pro o ng than previous work it is multi ling ual in
nature not using language dep endent letter mo dels and it pro duces higher level typ ographic details suitable
to mo dern visual communication applications
The problem dealt with in this pap er is very similar to the problem of feature recognition extensively
researched in geometric and solid mo deling Wo o dwark In b oth cases we desire to convert a b oundary
representation Brep which is a relatively low level representation to a higher level one supp orting terms
taken from the applicatio n domain
In Section we describ e the state of the art in digital typ ography to day and the hierarchy of digital typ eface
representations Section describ es the classi cation of the p oints describing the outline Section describ es
the actual extraction of the typ ographic elements Finally in Section we discuss an imp ortant application
of the resulting parametric typ ographic representation namely automatic hinting of fonts
Hierarchy of Digital Typ eface Representations
Examination of the di erent representations available to day in digital typ ography reveals a hierarchy in the
degree of abstraction concerning the freedom to mo dify the font at the level of the typ ographic design axes
Rubinstein Karow a Karow b Southall
The hierarchy b egins with bitmap fonts A bitmap font is a collection of matrices of dots in a xed size
whose app earance corresp onds to the original typ eface design Bitmaps are used in order to create the image
of a character in most output systems to day including phototyp esetters printers and CRT screens Therefore
every higher level representation should supp ort its eventual conversion to a bitmap Bitmaps are also the
most constrained representation having a xed co ordinate in all axes of the typ ographic taxonomy and almost
no p ossibili ty for an appropriate variation in any dimension The Macintosh and Windows op erating systems
have algorithms for slanting condensing b olding etc of bitmap fonts but the outcome is very p o or in quality
Digital outline fonts were created mainly to solve the xed size constraint in bitmap fonts Figure They
hold the ability to create bitmaps in any size using some sp ecial pro cess of scaling grid tting rasterizing and
lling Rubinstein Karow Microsoft Adob e Southall However many have found that simply
using an outline description of a typ eface do es not su ce in order to p erform the task of bitmap creation
Rubinstein Karow Hersch B etrisey Hersch a Hersch b Karow Additional information in
the form of hints has b een added to the font Hints are sp ecial constraints that help the grid tted outline
and thus the bitmap created from it retain some typ ographic attributes e g equality of stems height of
character symmetry of serif Microsoft Adob e Hersch a Karow Changyuan
Figure Outline of characters from left to right Roman Kanji and Hebrew fonts including their de ning
control points
Outlines can also b e used successfully in order to rotate text or to create outlined or slanted text sub ject to
sp ecial distortion algorithms which consider some more typ ographic constraints and probably using hints This
suggests that outline fonts could carry the ability to vary along any dimension in the typ ographic taxonomy
Attempts have b een made to use several outline representation instances of one typ eface in order to create a
representation with the ability to vary along one or more dimensions using interp olation b etween outlines or
extrap olation from one outline Apple Adob e The conclusion of these attempts has b een that in order
to succeed some more information concerning the typ ographic attributes of the typ eface should b e added to
the outline representation and later on used in the deformation algorithms
Several other description techniques for digital fonts have b een de ned with the intention of remaining as
much as p ossible in the typ ographic domain and saving as much information as p ossible while converting the
designer images to the actual description of the font Character parts in these representations are linked to
their typ ographic de nitions such as a stem an arc a dot or a serif They include typ ographic attributes
such as width angle sp ecial optical correction or o set values which in turn could b e parametrized creating
variation instances of the typ eface Knuth Adams Stamm a Stamm b McQueen We use the
term Parametric Typ ographic Representation PTR when referring to this class of digital typ eface
representations
A ma jor problem of previous work on PTR techniques is that it deals mainly with the creation or de nition
of new fonts disregarding the huge wealth of existing fonts already de ned or converted to the outline form
of representation Conversion techniques such as Haralamb ous rely on manually recognizing and inserting
parametric values to the outline which is very time consuming and requires sp ecial skills
Classi cation of Points
The rst stage in the conversion from an outline to a PTR consists of geometric and top ological analysis of
the outline Although there are several di erent metho ds for representing an outline of a glyph almost all of
them share some basic characteristics A glyph outline is a closed path consisting of consecutive segments of
two typ es straight line segments and curve segments see Figure lines segments are b ounded by squares
and curve segments are b ounded in b oth sides by circles Each segment is uniquely de ned by its start and
end p oints which lie on the outline of the glyph The p oints are numb ered consecutively and the direction of
the outline at each p oint is the direction of travel from lower indexed p oint to higher indexed p oints
Curve segments include some additional representation dep en dent control p oints These p oints lie outside
the outline and do not include any sp ecial information regarding the top ological nature of the outline therefore
we do not include them in our classi catio n These p oints are imp ortant in the pro cess of converting an outline
representation from one form to another e g cubic B ezier in Typ e fonts to Quadratic B Splines in TrueTyp e
fonts This involves converting only curve segments de ned by a set of some control p oints to the same or
approximated segments de ned by a set of other control p oints
In order to p erform the typ ographic decomp osition of the outline prop erly certain p oints have to b e included
in the glyph s outline Such p oints are lo cal maxima or minima in the X or Y direction in ection p oints
p oints where a discontinuity of derivative o ccurs or tangent p oints see de nitions b elow Often these p oints
exist in the outline and are used to de ne its segments However in order to make sure that no p oint is
missing and in order to correct misplaced p oints the outline description of glyphs go through a well known
preparation stage used often after digitizati on Karow Hence the rst degree classi cation of the outline
p oints b orrows its de nitions from the digitizatio n pro cess Figure
A corner p oint denoted as C p oint is a p oint where two outline segments meet with discontinuity of the
derivative of the outline
A curve p oint denoted as CT p oint is a p oint where two curve segments meet with continuity of the
derivative
A tangent p oint denoted as T p oint is a p oint where a curve segment and a line segment having the
direction of the derivative of the curve meet
The only neglected pair in these de nitions is the case of two line segments meeting with continuous
derivative It is safe to say that for almost all cases this situation is undesirable b ecause it carries redundant
information and the preparation stage of the outline would recognize such cases and replaces the two line
segments by a single one A case where this information should not b e discarded is for example when two
master glyphs are b eing matched in order to create interp olated instances b etween them It could b e that in
Figure Further classi cation of points H
Figure Classi cation of points corners points points dark in facing triangles V points light
squares tangents points circles curve point in facing triangles CH points dark out facing
triangle triangles CV points light out facing triangles
VR points rhombus
one of them this p oint is in fact a corner p oint discontinui ty and in the other the derivative of the segments
are continuous Discarding this p oint in one of the glyphs would make it hard to match these glyphs and to
create the interp olatio ns
Second degree classi cation is further applied to the p oint in order to extract geometric prop erties Figure
A curve V extremum V p oint is a curve p oint which is a lo cal extremum in the x direction
A curve H extremum H p oint is a curve p oint which is a lo cal extremum in the y direction
An in ection p oint I p oint is a curve p oint which is an in ection p oint
A corner V extremum CV p oint is a corner p oint which is a lo cal extremum in the x direction
A corner H extremum CH p oint is a corner p oint which is a lo cal extremum in the y direction
A spike S p oint is a p oint which is b oth a x extreme and a y extreme
A V corner VR p oint is a corner with one of its neighb oring segments de ned as a vertical line segment
A H corner HR p oint is a corner with one of its neighb oring segments de ned as a horizontal line
segment
An L corner LR p oint is a corner which is b oth a V corner and an H corner
One imp ortant distinction should b e made in these de nitions b etween real geometric prop erties and
de ned geometric prop erties The di erence lies in the notion of tolerance For example the real de nition
of vertical segments would b e
De nition Segment s is vertical if it is a line segment and its start and end p oints have the same x
co ordinate
When we refer to de ned vertical segments we generalize the previous de nition to b e
De nition Segment s is considered vertical if it is a line segment and for some prede ned epsilon the x
co ordinate of its start and end p oints do not di er by more than
For certain typ e of typ efaces where straight line segments are rare Figure a further generalization is
needed
De nition Segment s is considered vertical if for some prede ned epsilon the width xmax xmin of its
b ounding b ox is not greater than or equal to
We will call these kinds of generalizatio ns the epsilon rule Its implementation in geometric de nition
is crucial to the success of feature extraction b oth from non regularized typ efaces and from subtly designed
typ efaces
Beside the typ e of a p oint the additional information classi ed for each p oint includes its co ordinates the
typ e of segments meeting at the p oint and the direction of these segments
Figure Part of the outline of a character from the Macintosh Sai Mincho font Notice that the stem is
made of a left vertical line and a right vertical curve
Extraction of Typ ographic Elements
The second stage of conversion to a PTR is the actual extraction of typ ographic elements This stage uses all the
data acquired in the rst stage when the top ological and geometric analysis of the outline was p erformed Every
typ ographic element is de ned in terms of the typ es of p oints and segments buildin g it and the relationship s
b etween them This section presents these de nitions for each typ ographic element and discusses the algorithm
for recognizing them in the outline
All up coming de nitions are sub ject to the epsilon rule stated earlier Not only that the de nitions of
vertical or horizontal segments the de nition of equality of angles or sizes the de nition of the signi cant size
of overlaps etc are all within a certain parameter epsilon in fact the de nition of a segment may also vary
in turn and consist of several continuous or even broken segments with the same direction
Color Figure exempli es the terms in this section
Stems Bars and Slants
A stem consists of two vertical segments with opp osite directions overlapping each other vertically with
the space b etween the overlaps considered internal to the outline
A bar consists of two horizontal segments with opp osite directions overlapping each other horizontall y with
the space b etween the overlaps considered internal to the outline
A slant consists of two slanted segments having the same angle with opp osite directions overlapping each
other and the space b etween the overlaps is considered internal to the outline
The algorithm for nding such elements is quite straightforward we rst nd the segments and then match
opp osite direction segments For example the stems and bars extraction was de ned for Japanese characters
in Chialing
Slants might present some complication b ecause of the variety of angles We used a data structure sorted by
angle to insert non vertical or horizontal segments and match groups of lines created within a certain range
of angles
Bows and Arcs
A b ow consists of two V extreme p oints having the same y co ordinate and the same extreme typ e min or
max the outline passing through them having opp osite directions and the line b etween them considered
internal to the outline
An arc consists of two H extreme p oints having the same x co ordinate and the same extreme typ e min or
max the outline passing through them having opp osite directions and the line b etween them considered
internal to the outline
Finding b ows and arcs is done by nding all x or y extremum p oints sorting them by y or x co ordinate
and matching them to each other within the epsilon ranges
Curve Stems and Curve Bars
A curve stem is a V extreme p oint and a vertical segment at the side of the extreme on the left if the
V extreme p oint is a minimum and on the right if it is a maximum overlapping in some y co ordinate the
outline passing through them having opp osite directions and the line b etween them considered internal to
the outline
A curve bar is a H extreme p oint and a horizontal segment at the side of the extreme lower if minimum
and higher if maximum overlapping in some x co ordinate the outline passing through them having opp osite
directions and the line b etween them considered internal to the outline
The same sorted data structure having x or y extreme p oints and vertical or horizontal segments can
b e used to match extreme p oints to straight segments and extract curve stems and curve bars
Actually the three algorithms for nding stems bars arcs b ows curve stems and curve bars are executed
all in one pass First all the basic features like extreme p oints and vertical or horizontal segments or multi
segments are inserted to a data structure sorted by x or y co ordinate and then the matching is done to nd
the typ ographic elements The complexity of the whole pro cess is therefore O n log n where n is the numb er
of p oints in the outline The dominant part is the sorting stage
Serifs
Extraction of serif parts is a p owerful feature of our algorithm Serifs are one of the most imp ortant elements
in typ ographic design previous works have usually not addressed this issue completely
All other typ ographic elements e g bars stems arcs could b e de ned geometrically with acceptable
variations incorp orated in the epsilon rule Serifs could carry almost any geometric shap e and the rules whether
a serif is included or not in certain parts of the character may vary dramatically from one language to another
For example all bars in Japanese Mincho style characters p ossess a serif at the right end On the other hand
Hebrew serifs are lo cated only at the left end of the topmost bar in a character Therefore there is no way to
distinguish and recognize a serif without prior knowledge of the design of a font
We de ne a serif as a sequence or several p ossible sequences of p oints having certain prede ned typ es
according to the design of the serif For example a Mincho style bar serif and a Roman style stem serif could
b e de ned as seen in Figures and
The identi cation of a certain serif is done by using a nite automaton Lewis built to recognize the
sequence of p oints de ning the serif The outline is traversed and the typ es of the p oints encountered are fed
by order of app earance to the di erent nite automata for di erent serif typ es Whenever an automaton has
reached a nal accepting state it means that its serif has b een identi ed All nite automata are then reset to
the start state and the outline is further traversed In order not to dep end on the selection of the rst outline
p oint the outline is traversed once again or until all automata have b een reset
Figure Roman serif de ned by the sequence
Figure Kanji serif de ned by the sequence
fT point T point LC point LC point LC point
fHR point CH point T point V point T pointg
LC point T point T pointg
Whenever a new typ e of serif is encountered within a new typ eface only the sequence s of p oints de ning
it should b e inserted to the system It is given a name e g MinchoBarCap or RomanRoundHalfSeri f and a
nite automaton recognizing the sequence is created and added to a p o ol of automates Automates from this
p o ol can then b e turned on or o while converting a certain typ eface dep ending on the serif typ es exp ected
to b e encountered
There are cases when one serif s de ning sequence is a sub sequence in some other serif de ning sequence
Several metho ds could b e used to overcome this situation The simplest is turning o identi catio n for the
incorrect one This might not always b e p ossible since one typ eface could include b oth typ es of serifs A b etter
metho d would b e to further classify the p oint typ es For example an L corner could b e sub classi ed to four
di erent corner typ es according to orientation This metho d can also b e seen as adding geometric constraints
to the transition rules in the nite automata
When the metho ds ab ove fail we sp ecify certain automata as b eing sub automata of others This means
that when the sub automaton has reached its nal state the includin g automaton is not reset Both serifs
are identi ed and the correct one is chosen In our exp erience the combination of these metho ds solves most
ambiguities and clearly identi es all serifs
Automatic Hinting
The pro cess of hinting an outline font must recognize and use typ ographic features of the glyph and font in
order to create the hint rules or constraints In Hersch a a top ological mo del holding hinting information
of each character and variants of the same character is built A matching algorithm b etween real shap es of
characters and the mo dels is p erformed allowing the transformation of hints from the mo del to the real shap e
We have already mentioned the di culty of creating such mo dels for all huge variety of shap es representing
Roman letters moreover the task seems close to imp ossible when lo oking at ideographic scripts
Examining the di erent hint situations Karow reveals that most of them are actually lo cal in nature
meaning you must only examine a small p ortion of the outline of a glyph includin g several neighb oring p oints
neighb oring not only in the sense of consecutive p oints lying on the same contour but also close in space in
order to decide whether a hint should b e applied or not This means there is no need for a top ological mo del
for the whole glyph but rather a mo del for each hint situation This is exactly the typ e of mo del a PTR is
see Figure Once an outline font has b een converted to a PTR it already includes all information needed
for creating grid tting hints to the outline of its glyphs and so automatic hinting can easily b e p erformed
We have used our feature extraction technique on several di erent typ efaces having widely di erent designs
converting them to a PTR A large p ortion of the design di erence b etween typ efaces lies in the di erent
typ es of serif design or stroke endings in sans serif typ efaces This di erence can b e handled by de ning a
numb er of typical nite automata to describ e each typ eface s serifs Using this metho d we have managed to
extract automaticall y all hinting situations from the outline of glyphs see Figure and have actually used
this information to create constraints for Typ e fonts and instructions for TrueTyp e fonts
Most typ ographic features are also cross lingual in nature Examining distinct typ es of scripts and writings
one can see that a ma jor part of the di erences b etween letter shap es lies in the di erent usage and likeliho o d
of the same typ e of features This of course is the outcome of all of them sharing the common source pro cess
of writing in fact a few ma jor di erences arise from di erences in the writing techniques for example using
a p en in Roman scripts vs a brush in Chinese and Japanese Again di erent serif designs are a ma jor
distinction b etween languages Extracting typ ographic features should therefore b e cross lingual which it is
in our algorithm Figure
Hinting situations can therefore b e de ned almost exactly the same in all languages and so we have also
used a PTR for Japanese Kanji and PTR Hebrew typ efaces in order to automatically hint and create digital
fonts Some di erence in hinting di erent languages is of course inevitable but can b e overcome e ectively
using a PTR As an example Figure shows the outline of the letter m This letter exempli es one of the
most di cult rasterizing problems of outline fonts which even caused a sp ecial hint referred to as counter to
b e de ned The typ ographic constraints for rasterizing this glyph are b oth black and white in nature The
three stems of the letter should retain their equal widths at all sizes black but also the two gaps b etween
the stems should retain their equal widths at all sizes white
This task is not so simple to achieve when imp osing a pixel grid of certain size on the glyph s outline as can
b e seen in Figure The question whether to make the leftmost and rightmost stems one or two pixels wide
could b e answered only by lo oking at the middle stem Furthermore which pixel column should b e chosen to
represent the stems could b e answered only by lo oking at the gaps b etween the stems
In ideographic languages this problem is further enhanced b ecause of the quantity and complexity of bars
and stems in a glyph Figure and due to the nature of reading in these languages which uses mainly
recognition of bars and stems structures in order to discrimina te b etween symb ols
Our answer to this problem using a PTR is to de ne a higher level typ ographic element using lower level
bars and stems The sp ecial case where several stems are equal in width and length and are placed in equal
distances one after the other all within the tolerance rule is called a stem ladder or a bar ladder in the
horizontal case see the X symb ol in Figure This typ ographic element is then used as a constraint while
grid tting or deforming the letter and can b e further converted to rules or hints in a rasterizing language
such as TrueTyp e or Typ e Adob e Typ e technology supp orts only triple ladders
A PTR generated by our feature extraction algorithm is therefore a general and robust solution to the
problem of automatic hinting of digital outline typ efaces which is one of the most imp ortant stages of digital
font pro duction to day
Conclusion
We have describ ed a complete and e cient algorithm for extracting typ ographic elements from outline repre
sentations of typ efaces The algorithm is essential for converting existing typ efaces into the mo dern family of
representations which we call parametric typ ographic representations PTR
The algorithm was implemented as a part of a commercial system for conversion b etween di erent outline
representations and for automatic hinting and regularization of typ efaces The system has b een used on dozens
of thousands of Roman Kanji and Hebrew characters and it is b eing successfully used on a daily basis The
fonts generated are consistently of higher hinting quality than most commercially available fonts
A PTR is capable of providing answers to well known and di cult problems in font rasterization and
digitizatio n such as hinting and regularization In addition it enables typ efaces to take a more active part
in electronic media using deformations variations and animations while retaining their original typ ographic
nature and design Lack of space prevents us from describing these PTR applications in the present pap er
they are more fully describ ed in the technical rep ort Shamir and will b e presented in future pap ers
Additional future work should concentrate on the issues of parametrization of the extracted typ ographic
elements and on the de nition of typ ographic parts for extensively cursive and illustrated typ efaces
References
Adams Adams D Ab cdefg a b etter constraint driven environment for font generation in Andr e Hersch Eds
Raster Imaging and Digital Typ ography Cambridge University Press pp
Adob e Adob e Develop er Supp ort Adob e Typ e Font Format Multiple Master Extensions
Adob e Adob e Systems Inc The Typ e Format Sp eci cation Addison Wesley
Andler Andler S Automatic generation of grid tting hints for rasterization of outline fonts or graphics in R
Furuta Ed EP Pro ceedings of the International Conference on Electronic Publishing Do cument Manipulation
Typ ography Cambridge University Press Sep pp
Apple Apple Computer QuickDraw GX Typ ography Font File Formats
Bauermeister Bauermeister B A Manual of Comparative Typ ography the PANOSE system Van Nostrand Rein
hold New York
B etrisey B etrisey C and Hersch Roger D Flexible application of outline grid constraints in Andr e Hersch Eds
Raster Imaging and Digital Typ ography Cambridge University Press pp
Bringhurst Bringhurst Rob ert The Elements of Typ ographic Style Hartley Marks
Changyuan Changyuan Hu and Fuyan Zhang Automatic hinting of Chinese outline font based on stroke sepa
rating metho d Pro ceedings of the First Paci c Conference on Computer Graphics and Applications Paci c Graphics
Vol pp
Chialing Chialing Ou and Yoshio Ohno Font generation algorithms for Kanji characters in Andr e Hersch Eds
Raster Imaging and Digital Typ ography Cambridge University Press pp
Durst a Durst Martin J Co ordinate indep endent font description using Kanji as an example Electronic Publishing Origination Dissemination and Design John Wiley Sons
Durst b Durst Martin J Structured character description for font design a preliminary approach based on Prolog
Pro ceedings of the First Paci c Conference on Computer Graphics and Applications Paci c Graphics Vol
pp
Feng Feng H F and Pavlidis T Decomp osition of p olygons into simpler comp onents feature generation for syn
tactic pattern recognition IEEE Transactions or Computers vol June pp
Haralamb ous Haralamb ous Yannis Parametrization of PostScript fonts through metafont an alternative to
Adob e Multiple Master fonts Electronic Publishing Origination Dissemination and Design John
Wiley Sons
Hersch a Hersch Roger D and B etrisey C Mo del based matching and hinting of fonts Pro ceedings SIGGRAPH
ACM Computer Graphics Vol July pp
Hersch b Hersch Roger D and B etrisey C Advanced grid constraints p erformances and limitations in Morris
Andr e Eds Raster Imaging and Digital Typ ography Cambridge University Press pp
Hersch Hersch Roger D Character generation under grid constraints Pro ceedings SIGGRAPH ACM Com
puter Graphics Vol July pp
Karow a Karow Peter Digital Typ efaces Description and Formats Springer Verlag Berlin
Karow b Karow Peter Font Technology Description and To ols Springer Verlag Berlin
Karow Karow Peter Automatic hinting for intelligent font scaling in Andr e Hersch Eds Raster Imaging and
Digital Typ ography Cambridge University Press
Karow Karow Peter Intelligent FontScaling Technical rep ort URW Unternehmensb erstung Hamburg Germany
Knuth Knuth Donald E The metafont Bo ok Addison Wesley
Lawson Lawson Alexander Anatomy of a Typ eface Hamish Hamilton
Lewis Lewis H R and Papadimitriou C H Elements of The Theory of Computation Prentice Hall
McQueen McQueen Clyde D and Beausoleil Raymond G In nifont a parametric font generation system Elec
tronic Publishing Origination Dissemination and Design John Wiley Sons
Microsoft Microsoft Corp oration The TrueTyp e Font Format Sp eci cation
Rubinstein Rubinstein Richard Digital Typ ography An Intro duction to Typ e and Comp osition for Computer
System Design Addison Wesley
Shamir Shamir A and Rapp op ort A Parametric typ ographic representation of typ efaces automatic conversion
from outline fonts and applications Technical Rep ort TR Institute of Computer Science The Hebrew Univer
sity
Southall Southall Richard Character Description Techniques in Typ e Manufacture in Morris Andr e Eds Raster
Imaging and Digital Typ ography Cambridge University Press pp
Stamm a Stamm Beat Dynamic regularization of intelligent outline fonts Electronic Publishing Origination Dis
semination and Design John Wiley Sons
Stamm b Stamm Beat Ob ject oriented and extensibility in a font scalar Electronic Publishing Origination Dis
semination and Design John Wiley Sons
Wo o dwark Wo o dwark J R Some sp eculations on feature recognition in Geometric Reasoning Kapur and Mundy
Eds
Zhang Zhang Fuyan and Ge Weimua An automatic stroke separating metho d Pro ceedings of the third international conference on Chinese information pro cessing ICCCIP Beijing
Figure The lower case letters of the Bookman Light typeface with the typographic features extracted marked
in di erent colors Vertical elements such as stems and bows are shown green horizontal elements such as bars
and arcs are shown blue slants are shown yel low and serifs are shown red
Figure Typographic elements stems green bars blue bows red arcs violet curve stems orange
notice the extremum triangle
Figure Similar features extracted from di erent type of G glyphs
Figure Enhanced use of bars and stems in Japanese
Figure The multiple stems rasterizing problem