Europäisches Patentamt

(19) European Patent Office Office européen des brevets (11) EP 0 551 941 B1

(12) EUROPEAN PATENT SPECIFICATION

(45) Date of publication and mention (51) Int. Cl.7: G06K 9/46, G07C 9/00, of the grant of the patent: A61B 5/117 12.07.2000 Bulletin 2000/28

(21) Application number: 93200057.3

(22) Date of filing: 11.01.1993

(54) Classifying faces Gesichtsklassifizierung Classification de faces

(84) Designated Contracting States: (56) References cited: DE FR GB US-A- 4 975 969

(30) Priority: 17.01.1992 GB 9201006 • IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS vol. 35, no. 6 , June 1988 , USA pages (43) Date of publication of application: 636 - 647 PITAS I., VENETSANOPOULOS A. 'A 21.07.1993 Bulletin 1993/29 New Filter Structure for the Implementation of Certain Classes of Image Processing (73) Proprietors: Operations' • PHILIPS ELECTRONICS UK LIMITED • IEEE PROCEEDINGS OF THE 9TH Croydon CR9 3QR (GB) INTERNATIONAL CONFERENCE ON PATTERN Designated Contracting States: RECOGNITION, IEEE CAT. NR. 88CH2614-6, vol. GB 2 , 17 November 1988 , ROME, ITALY pages 833 - • Koninklijke Philips Electronics N.V. 835 ZHU Q. AND POH L. 'A Transformation- 5621 BA Eindhoven (NL) Invariant Recursive Subdivision Method for Designated Contracting States: Shape Analysis' DE FR • JOURNAL OF THE OPTICAL SOCIETY OF AMERICA (OPTICS AND IMAGE SCIENCE) vol. 3, Inventors: (72) no. 6 , June 1986 , NEW YORK US pages 771 - • Trew, Timothy Ian Paterson 776 SHENG Y. AND ARSENAULT H. Redhill, Surrey RH1 5HA (GB) 'Experiments on pattern recognition using • Gallery, Richard David invariant Fourier-Mellin descriptors' Redhill, Surrey RH1 5HA (GB) • NTZ ARCHIV vol. 8, no. 10 , October 1986 , BERLIN DE pages 245 - 256 BUHR R. 'Analyse (74) Representative: und Klassifikation von Gesichtsbildern' Andrews, Arthur Stanley et al • PATENT ABSTRACTS OF JAPAN vol. 8, no. 101 Philips Electronics UK Limited (P-273)12 May 1989 & JP-A-59 011 471 (SONY Patents and Trade Marks Department KK) 21 January 1984 Cross Oak Lane • PROCEEDINGS OF THE SECOND EUROPEAN Redhill, Surrey RH1 5HA (GB) CONFERENCE ON COMPUTER VISION, ECCV'92, SPRINGER VERLAG, 23 May 1992 , SANRA MARGHERITA LIGURE, ITALY pages 92 - 96 CRAW I. ET AL 'Finding Face Features'

Note: Within nine months from the publication of the mention of the grant of the European patent, any person may give notice to the European Patent Office of opposition to the European patent granted. Notice of opposition shall be filed in a written reasoned statement. It shall not be deemed to have been filed until the opposition fee has been paid. (Art. 99(1) European Patent Convention). EP 0 551 941 B1 Printed by Xerox (UK) Business Services 2.16.7 (HRS)/3.6 12EP 0 551 941 B1

Description vision Method for Shape Analysis". [0007] Fourier-Mellin transformations are known [0001] The invention relates to a face classification from Journal of the Optical Society of America (Optics system. The invention further relates to a security sys- and image science) vol. 3, no. 6, June 1986, New York tem capable of identifying persons using face classifica- 5US pages 771 - 776 Sheng Y. and Arsenault H. "Exper- tion. iments on pattern recognition using invariant Fourier- [0002] Any method of face classification divides Mellin descriptors". naturally into two stages; a) Face location and b) Clas- [0008] A departure from the either wholly statistical sification. or wholly structural approaches is disclosed by I. Craw [0003] A survey of the use of face recognition for 10 & P. Cameron, "Parameterising Images for Recognition security applications by M. Nixon, "Automated Facial and Reconstruction" BMVC 91, pp 367-370, Springer- Recognition and its Potential for Security" IEE Colo- Verlag (1991) which uses a hybrid structural/statistical quium on "MMI in Computer Security", (Digest No. 80) approach, in which a large number of feature points are (1986), classifies face recognition techniques as either utilised to normalise the shape of the face, and then statistical or structural in side view of front view. For 15 principal component analysis is used to obtain a lower front view the structural techniques are further classified dimensional representation of the face. This is com- in terms of feature measurements and angular meas- pared with a database of similarly encoded faces for urements. recognition purposes. Such a hybrid approach offers [0004] Any system which is to operate in an uncon- the advantage of not having to constrain the face unduly strained environment, must use a structural approach 20 (structural), and also retains the significant advantages which incorporates some knowledge of the structure of of statistical methods. a face so that features may still be located under varying [0009] US4975969 discloses methods, apparatus lighting conditions, background and pose. Metrics are and security systems for uniquely identifying individuals constructed based upon the relationships between by their particular physical characteristics and, in partic- located facial features, and are then used to classify the 25 ular, distances and / or ratios of distances between face. One problem with this approach is choosing identifiable points on the human face. Also disclosed is appropriate metrics. For example, R. Buhr, "Front Face a security system comprising a video camera and a Analysis and Classification (Analyse und Klassification security system for a computer terminal. von Geischtsbildern)", ntz Archiv 8, No. 10, pp 245-256 [0010] The invention provides a face classification (1986) proposed using 45 measures. Additionally, for 30 system comprising first means for locating dimensional such an approach it is necessary to locate the facial fea- scene, second means for locating the face in the repre- tures with high precision. Other examples of this sentation of the scene, third means for forming a rota- approach are exemplified by R. J. Baron, "Mechanism tion, scaling, translation, and grey level intensity of Human Facial Recognition", Int. J. Man-Machine invariant representation of the face and producing a fea- Studies, 15, pp 137-178 (1981) and T. Sakai, M. Nagao 35 ture vector therefrom, and fourth means for comparing & M. Kanade, "Computer Analysis and Classification of the feature vector of the presently located face with the Photographs of Human Faces", Proc. 1st USA-Japan feature vector of a previously located face to determine Computer Conf. AFIPS Press, New Jersey, pp 55-62 whether the presently located face matches the previ- (1972). ously located face characterised in that the third means [0005] The attraction of the statistical approach is 40 comprises means for fitting an outline to the face, the possibility that simple methods may be employed to means for locating the mid-point between the eyes, and extract feature vectors. I. Aleksander, W.V. Thomas & means for performing a Fourier-Mellin transformation on P.A. Bowden have disclosed in "A Step Forward in the face referenced to the mid-point between the eyes to Image Recognition", Sensor Review, July, ppl20-124 produce a feature vector of the face. (1984) a statistical face recognition system, using 45 [0011] By applying a rotation, scaling, translation, WISARD, which is able to recognise a pattern within and grey level intensity invariant representation to the one frame period. The main shortcoming of this system face constraints on the position and orientation of the is that it is specific to a particular position and orienta- face and on the scene lighting can be relaxed enabling tion, so the individual's characteristics must be learnt for the face classification to be carried out unobtrusively as a series of spatial displacements, reducing the reliability 50 far as the person being classified is concerned. How- of the identification and reducing the storage capacity of ever, whilst the Fourier-Mellin transformation is rela- the system. tively insensitive to noise, on a global picture is not [0006] A transformation invariant method of recur- translation invariant. By locating the mid-point between sive subdivision for shape analysis is known from IEEE the eyes and referencing the transformation to that point Proceedings of the 9th international conference on pat- 55 an effectively translation invariant transformation can be tern recognition, IEEE CAT. NR. 88CH2614-6, vol. 2, 17 obtained. November 1988, Rome, Italy pages 833 to 835 Zhu Q [0012] The invention further provides a face classi- and Pol L. "A transformation-Invariant Recursive Subdi- fication system comprising first means for locating a

2 34EP 0 551 941 B1 face in a two dimensional representation of a three authorised to use the card onto the card's memory, dimensional scene, second means for locating the face which could then be compared with the face of the indi- in the representation of the scene, third means for form- vidual attempting to use the card during transactions. Of ing a rotation, scaling, translation, and grey level inten- critical importance in this application is the length of the sity invariant representation of the face and producing a 5feature vector, which must fit within the memory sup- feature vector therefrom and fourth means for compar- ported by the card. Thus in such an application, the ing the feature vector of the presently located face with facial feature extraction method used must also com- the feature vector of a previously located face to deter- press the facial data presented to it. Use of the recursive mine whether the presently located face matches the second order sub-division of moments is one way of ful- previously located face characterised in that the third 10 filling this requirement. means comprises means for filling an outline to the face, [0020] It is also important to minimise the amount of means for locating the eyes and nose in the face, means space the stored feature vectors occupy in central mem- for sub-dividing the face by two lines, one joining the ory for the automatic access to secure areas if the data eyes and the other perpendicular thereto and through base of authorised persons is extensive. While it would the nose, and means for performing the recursive sec- 15 be possible where access to a secure area is concerned ond order sub-division of moments on the sub-divided to compare the face presented at the entrance with all areas of the face. authorised persons stored in the data base it may be [0013] This produces a short feature vector which preferable to have a personal identification number allo- enables minimisation of the storage required for the rep- cated to each authorised person and to use this number resentation of a face and also has the advantage of 20 to identify the stored feature vector of the face of the using information from the whole face and not just the person seeking entry for comparison with that produced boundary or other edges. by an entry camera. [0014] Preferably, the first sub-division takes place [0021] An alternative procedure with credit cards on the line joining the eyes and the second sub-division would be to have a central data base of feature vectors takes place on the perpendicular line. 25 of all card holders and for the transaction terminal to [0015] The recursive second order sub-division of view and form a feature vector of the card user. This fea- moments is purely statistical in its operation and is noise ture vector would then be transmitted to the central data sensitive in that a small perturbation in the location of base where the comparison would be made, optionally the centre of gravity will change the subsequent subdi- using credit card data to address the appropriate fea- visions substantially. By including structural information, 30 ture vector in the data base. Subsequently an authorisa- that is by constraining the initial sub-division to take tion signal would be relayed to the transaction terminal. place through the eyes such that two regions are formed This removes the need to store the feature vector on the and then perpendicular to the previous subdivision card and hence removes the constraint on the length of through the nose so that each region is further divided the feature vector caused by the storage capacity of the in two, the transformation is made more robust and less 35 card. It is still, however, desirable to minimise the length sensitive to noise. of the feature vector to reduce the required capacity of [0016] A known homomorphic filter may be pro- the central data base and also the transmission time vided for producing a grey level intensity invariant repre- required for transmission of the feature vector from the sentation. transaction terminal to the central data base. [0017] This is a convenient way of minimising the 40 [0022] By transmitting the feature vector to a central effects of lighting changes in the scene. data base the complexity of the transaction terminal can [0018] The invention yet further provides a security be minimised in that the comparison means for the fea- system comprising a video camera for producing a pic- ture vectors is not required to be present in the transac- ture frame having a face located therein, means for tion terminal. There will, of course, be many transaction determining whether the present face matches at least 45 terminals connected to the card verification data base one previously located face by a face classification sys- and minimising their cost is important in the provision of tem as described above, and means for initiating a a comprehensive system. security measure if the two faces are not matched. [0023] Such a security system for use with a com- [0019] There are many potential areas of applica- puter terminal may periodically monitor the user's face tion for security systems including an automatic face 50 and determine whether the present face matches the identification system. Automatic access to secure areas user's face at logon. is one example, in which the face of the individual seek- [0024] As the use of workstations in financial, and ing to enter might be compared with a data base of other commercial environments, becomes more wide- faces of individuals allowed access, optionally with addi- spread, workstation security is becoming of paramount tional verification by personal identification number. 55 importance. A major problem in maintaining the integrity Another potential area of application is in preventing of workstation security is that of authorised users leav- credit card fraud. This might involve the encoding of a ing a terminal unattended without logging out or putting feature vector, representing the face of the individual it into pause mode. Unauthorised users may then gain

3 56EP 0 551 941 B1 access to the system. One existing solution to this prob- and from detecting areas of highest movement. Thus lem is password verification, either periodically, or if the the eyes and mouth are frequently changing in shape keyboard is unused for some time, which is of course a and are at the corners of a triangle. relatively obtrusive method. Another, biometric, tech- [0030] A recursive second order sub-division of nique used is based upon the analysis of the keystroke 5moments is then applied 7 to the filtered image data rhythm of the authorised user. However such tech- from the homomorphic filter 2. This transform is niques encounter difficulty in coping with individuals restricted to the face region using the model of the out- whose typing style changes when performing different line of the face produced by the face extraction and tasks. location means 3 and is constrained by first a sub-divi- [0025] By continuously monitoring the facial fea- 10 sion on a line joining the two eyes and second a sub- tures of the user it is possible to ensure that no one else division on a line perpendicular thereto passing through has taken over use of the terminal since logon. the nose. By constraining the sub-divisions the effects [0026] The above and other features and advan- of noise can be reduced. The means 7 for applying the tages of the invention will be apparent from the exem- recursive second order sub-division of moments pro- plary embodiments of the invention which will now be 15 duces a feature vector characteristic of a given face. described with reference to the accompanying draw- This is fed to a classification stage 8 which compares ings, in which:- the feature vector just obtained with a stored feature vector to determine whether it represents the same Figure 1 shows in block schematic form one face. The classification stage 8 may be a multi-layer per- embodiment of a face classification system accord- 20 ception which can be trained in known manner to clas- ing to the invention, sify the faces. For suitable training methods reference Figure 2 shows in block schematic form one could be made to the textbook "Neural Computing - embodiment of a computer workstation in which a Theory and Practice" by Philips D. Wasserman pub- face classification system according to the inven- lished by Van Nostrand Reinhold, New York. tion is incorporated. 25 [0031] Figure 3 shows examples of the feature vec- Figure 3 illustrates the feature vectors produced by tors produced by the second order sub-division of the second order moment sub-division faces, and moments constrained by a first sub-division on a line Figure 4 is a flow diagram illustrating the operation joining the eyes and a second sub-division on a line per- of a security system for a computer workstation. pendicular thereto through the nose for two instances of 30 the faces of two individuals. [0027] As shown in figure 1 image data from a video [0032] Figure 2 shows in block schematic form a camera (not shown) is applied via an input 1 to a homo- computer workstation which comprises a central morphic filter 2 to remove the effects of lighting changes processing unit 10 having connected to it a keyboard on the representation of the scene. The image data is 11, a video display unit (VDU) 12, and memory 13. A also fed to face location and extraction means 3. In this 35 video camera 14 is mounted to view a user of the com- particular embodiment the face location and extraction puter workstation and provides an output which is fed to means 3 comprises an edge detector 4 and Hough a frame grabber 15 and Codec 16. The output of the transform means 5 but it could take other forms, for frame grabber 15 is fed to the central processing unit example that disclosed by E. Badiqué in a paper entitled 10. "knowledge-Based Facial Area Recognition and 40 [0033] The flow diagram shown in Figure 4 illus- Improved Coding in a CCITT-Compatible Low-bitrate trates the operation of the security system incorporated Video-Codec" presented at the Picture Coding Sympo- in the workstation shown in Figure 3. At the commence- sium at Cambridge, Massachusetts on 26-28th March ment ST (start), i.e. logon of the operator, the video 1990. camera frames are grabbed by the frame grabber 15 [0028] A simple model of the outline of a face is an 45 and the central processing unit 10 performs the homo- ellipse and an architecture for fitting ellipses to faces is morphic filtering of the image data and the face location disclosed by S. A. Rajala and A. M. Alattair in "A Paral- as illustrated by box LF (locate face). A decision is then lel/Pipeline Structure for the Primitive-Based Image taken IFAC? (is face aligned correctly?) as to whether Codec for Coding Head and Shoulders Images" PCS90, the face is correctly aligned, for example whether both pp 9.16-9.17 (1990). More accurate modelling of faces 50 eyes are visible to the camera. If the camera is mounted in which simulated annealing is used to fit a head outline adjacent to the VDU this will usually be the case but if to a face is described by A. Bennet and I. Craw in "Find- the decision is N (no) then loop 100 is followed and fur- ing Image Features Using Deformable Templates and ther frames are grabbed until the face is correctly Detailed Prior Statistical knowledge". BMVC 91, pp aligned. Once a correctly aligned face has been 233-239, Springer-Verlap (1991). 55 detected it is extracted from the image EFFI (extract [0029] The location of the eyes and mouth within face from image) and then a model, for example an the face is determined by means 6. This may be ellipse, is fitted to its boundary FMFB (fit model to facial achieved from a knowledge of the geometry of the face boundary). The calculation of the recursive second

4 78EP 0 551 941 B1 order sub-division of moments is then commenced con- to a restricted area a person may have to stand in front strained by the initial sub-division through the eyes CIS of a camera in a defined position. Then either a search (constrain initial sub-division) and the feature vectors through a data base of faces of authorised persons is calculated CFV (calculate feature vector). A decision is performed or by entering a code either by means of a then made as to whether this is the system initialisation 5card or a keyboard a comparison is made with the sin- run SIN? (is this the system initialisation run?) and if Y gle stored face identified with that code. (yes) then the feature vector is stored SFV (store feature [0038] For a credit card verification system a similar vector) for comparison with feature vectors calculated arrangement can be used in that a central data base of on later images. There is then a pause for a predeter- feature vectors of all authorised card users can be mined time WFNC (wait for next check) until the next 10 accessed by means of a transmission link from a trans- check is to be done when the face location check LF is action terminal. The terminal will include a video cam- made. era and means for extracting the face from the image [0034] If it is not the system initialisation run then produced by the camera and for generating the feature the feature vector obtained is compared with the stored vector. This feature vector is then transmitted to the cen- feature vector and a difference vector is generated GDV 15 tral data processing establishment where the feature (generate difference vector). A decision is then made as vectors of all card holders are stored. The card number to whether the same face has been located by thresh- can be used to select the feature vector of the author- olding the difference vectors SF? (same face?) Since ised user for the card from the store and the classifier some components of the vectors are more sensitive used to determined whether the face captured by the than others the components should be individually 20 transaction terminal is that of the authorised user. An weighted. This is implemented by taking pairs of faces, authorisation or alarm signal as appropriate can then be which might be of the same or different individuals and transmitted to the transaction terminal to either give or training a multi-layer perception, which forms the classi- withhold an authorisation for the card to be used. fier, on the difference between the vectors to recognise [0039] An alternative arrangement is to have the these two classes enabling the classifier to decide 25 feature vector of the authorised user stored on the card. whether or not the vectors are sufficiently similar to In this case it is either necessary that the classifier is each other to be from the same face. If the answer is Y incorporated in the transaction terminal or that both the that is the classifier believes it is the same face then the stored feature vector from the card and the feature vec- wait until next check TFNC loop is followed and in due tor generated by the transaction terminal are transmit- course the face of the user will again be checked to see 30 ted to a classifier at the central data processing that the same person is still at the workstation. In the establishment for the decision as to whether or not the event that a different person is detected at the worksta- person presenting the card is the authorised user to be tion then the answer to SF? is N and further security made. measures are initiated IFSM (initiate further security [0040] From reading the present disclosure, other measures). These could take various forms, for exam- 35 modifications will be apparent to persons skilled in the ple audible or visible alarms, disabling the workstation, art. Such modifications may involve other features requiring the entry of a password, etc. which are already known in the design, manufacture [0035] The sequence of operation of the security and use of face classification and security systems and system for a computer workstation can be summarised devices and component parts therefor and which may as follows (neglecting the initial stage in which it grabs 40 be used instead of or in addition to features already the image of the face of the user at logon). described herein. Although claims have been formu- lated in this application to particular combinations of 1) Locate user face in image, features, it should be understood that the scope of the 2) grab image of face, disclosure of the present application also includes any 3) extract face from background, 45 novel feature or any novel combination of features dis- 4) extract features from face, closed herein either explicitly or implicitly or any gener- 5) compare current user face with logon user face, alisation thereof, whether or not it relates to the same 6) if same user at 5) repeat 1) to 5) after given time invention as presently claimed in any claim and whether interval, or not it mitigates any or all of the same technical prob- if different user at 5) initiate further security action. 50 lems as does the present invention. The applicants hereby give notice that new claims may be formulated to [0036] The initial feature vector at logon is acquired such features and/or combinations of such features dur- using steps 1) to 4) above and the feature vector thus ing the prosecution of the present application or of any acquired is stored to enable the comparison step 5) to further application derived therefrom. be carried out. 55 [0037] For other security applications modifications Claims to this procedure may be required and corresponding system differences may occur. For example for access 1. A face classification system comprising first means

5 91EP 0 551 941 B1 0

for locating a face in a two dimensional representa- for a computer terminal (10, 11, 12, 13) wherein the tion of a three dimensional scene, second means user's face is periodically monitored and the at least for locating the face in the representation of the one previously located face comprises the user's scene, third means for forming a rotation, scaling, face at logon. translation, and grey level intensity invariant repre- 5 sentation of the face and producing a feature vector Patentansprüche therefrom and fourth means for comparing the fea- ture vector of the presently located face with the 1. Gesichtsklassifikationssystem mit ersten Mitteln feature vector of a previously located face to deter- zum Orten eines Gesichtes in einer zweidimensio- mine whether the presently located face matches 10 nalen Darstellung einer dreidimensionalen Szene, the previously located face characterised in that the mit zweiten Mitteln zum Orten des Gesichtes in der third means comprises means for fitting an outline Szene, mit dritten Mitteln zum Bilden einer dre- to the face, means for locating the mid-point hungs-, skalierungs-, translations- und graupegel- between the eyes, and means for performing a Fou- intensitätsinvarianten Darstellung des Gesichtes rier-Mellin transformation on the face referenced to 15 und zum Erzeugen eines Kennzeichenvektors dar- the mid-point between the eyes to produce a fea- aus, und mit vierten Mitteln zum Vergleichen des ture vector of the face. Kennzeichenvektors des soeben georteten Gesich- tes mit dem Kennzeichenvektor eines vorher geor- 2. A face classification system comprising first means teten Gesichtes um zu bestimmen, ob das soeben for locating a face in a two dimensional representa- 20 geortete Gesicht mit dem vorher georteten Gesicht tion of a three dimensional scene, second means übereinstimmt, dadurch gekennzeichnet, dass die for locating the face in the representation of the dritten Mittel Mittel aufweisen zum Passen einer scene, third means for forming a rotation, scaling, Kontur auf das Gesicht, und Mittel zum Durchfüh- translation, and grey level intensity invariant repre- ren einer Fourier-Mellin-Transformation an dem sentation of the face and producing a feature vector 25 Gesicht, bezogen auf den Mittelpunkt zwischen den therefrom and fourth means for comparing the fea- Augen zum Erzeugen eines Kennzeichenvektors ture vector of the presently located face with the des Gesichtes. feature vector of a previously located face to deter- mine whether the presently located face matches 2. Gesichtsklassifizierungssystem mit ersten Mitteln the previously located face characterised in that the 30 zum Orten eines Gesichtes in einer zweidimensio- third means comprises means for fitting an outline nalen Darstellung einer dreidimensionalen Szene, to the face, means for locating the eyes and nose in mit zweiten Mitteln zum Orten des Gesichtes in der the face, means for sub-dividing the face by two Darstellung der Szene, mit dritten Mitteln zum Bil- lines, one joining the eyes and the other perpendic- den einer drehungs-, skalierungs-, translations- ular thereto and through the nose, and means for 35 und graupegelintensitätsinvarianten Darstellung performing the recursive second order sub-division des Gesichtes und zum Erzeugen eines Kennzei- of moments on the sub-divided areas of the face. chenvektors daraus, und mit vierten Mitteln zum Vergleichen des Kennzeichenvektors des soeben 3. A face classification system according to claim 2 georteten Gesichtes mit dem Kennzeichenvektor wherein the first sub-division takes place on the line 40 eines vorher georteten Gesichtes um zu bestim- joining the eyes and the second sub-division takes men, ob das soeben geortete Gesicht mit dem vor- place on the perpendicular line. her georteten Gesicht übereinstimmt, dadurch gekennzeichnet, dass die dritten Mittel Mittel auf- 4. A face classification system as claimed in any pre- weisen zum Passen einer Kontur auf das Gesicht, ceding claim in which a homomorphic filter (2) is 45 Mittel zum Orten der Augen und der Nase in dem provided for producing a grey level intensity invari- Gesicht, Mittel zum Unterteilen des Gesichtes ant representation. durch zwei Linien, wobei die eine Linie die Augen miteinander verbinden und die andere Linie senk- 5. A security system (14, 15, 16) comprising a video recht darauf steht und durch die Nase geht, und camera 14 for producing a picture frame having a 50 Mittel zum Durchführen der rekursiven Unterteilung face located therein, means for determining von Momenten zweiter Ordnung an den unterteilten whether the present face matches at least one pre- Gebieten des Gesichtes. viously located face by a face classification system according to any preceding claim, and means for 3. Gesichtsklassifikationssystem nach Anspruch 2, initiating a security measure if the two faces are not 55 wobei die erste Unterteilung durch die Linie stattfin- matched. det, welche die Augen miteinander verbindet und die zweite Unterteilung stattfindet durch die senk- 6. A security system (14, 15, 16) as claimed in claim 5 recht darauf stehende Linie.

6 11 EP 0 551 941 B1 12

4. Gesichtsklassifizierungssystem nach einem der localisé correspond au visage précédemment loca- vorstehenden Ansprüche, wobei ein homomorphes lisé, caractérisé en ce que les troisièmes moyens Filter (2) vorgesehen ist zum Erzeugen einer grau- comprennent des moyens pour doter le visage d'un pegelintensitätsinvarianten Darstellung. contour, des moyens pour localiser les yeux et le 5 nez sur le visage, des moyens pour subdiviser le 5. Sicherheitssystem (14, 15, 16) mit einer Videoka- visage en deux lignes, l'une joignant les yeux et mera (14) zum Erzeugen eines Bildes mit einem l'autre étant perpendiculaire à ceux-ci et passant Gesicht, mit Mitteln um zu bestimmen, ob das aktu- par le nez, et des moyens pour effectuer la subdivi- elle Gesicht mit wenigstens einem vorher von sion récurrente du second ordre de moments sur einem Gesichtsklassifizierungssystem nach einem 10 les zones subdivisées du visage. der vorstehenden Ansprüche georteten Gesicht übereinstimmt, und mit Mitteln zum Auslösen einer 3. Système de classification de visages suivant la Sicherheitsmaßnahme, falls die zwei Gesichter revendication 2, dans lequel la première subdivi- nicht miteinander übereinstimmen. sion a lieu sur la ligne joignant les yeux et la 15 deuxième subdivision a lieu sur la ligne perpendi- 6. Sicherheitssystem (14, 15, 16) nach Anspruch 5 für culaire. einen Computerterminal (10, 11, 12, 13), wobei das Gesicht des Benutzers periodisch überwacht wird 4. Système de classification de visages suivant une und das wenigstens eine vorher geortete Gesicht quelconque revendication précédente, dans lequel das Gesicht des eingeloggten Benutzers umfasst. 20 un filtre homomorphique (2) est prévu pour pro- duire une représentation invariante de l'intensité du Revendications niveau de gris.

1. Système de classification de visages comprenant 5. Système de sécurité (14, 15, 16) comprenant une des premiers moyens pour localiser un visage dans 25 caméra vidéo 14 pour produire une image de pho- une représentation bidimensionnelle d'un décor tri- tographie contenant un visage y étant localisé, des dimensionnel, des deuxièmes moyens pour locali- moyens pour déterminer si le présent visage cor- ser le visage dans la représentation du décor, des respond à au moins un visage précédemment loca- troisièmes moyens pour former une rotation, une lisé par un système de classification de visages mise à l'échelle, une translation, et une représenta- 30 suivant une quelconque revendication précédente, tion invariante de l'intensité du niveau de gris du et des moyens pour initier une mesure de sécurité visage et la production d'un vecteur de traits à partir si les deux visages ne correspondent pas. de là et des quatrièmes moyens pour comparer le vecteur de traits du visage présentement localisé 6. Système de sécurité (14, 15, 16) suivant la revendi- au vecteur de traits d'un visage localisé précédem- 35 cation 5 pour un terminal d'ordinateur (10, 11, 12, ment pour déterminer si le visage présentement 13) dans lequel le visage de l'utilisateur est sur- localisé correspond au visage précédemment loca- veillé périodiquement et le au moins un visage pré- lisé, caractérisé en ce que les troisièmes moyens cédemment localisé comprend le visage de comprennent des moyens pour doter le visage d'un l'utilisateur à l'entrée en communication. contour, des moyens pour localiser le point central 40 entre les yeux, et des moyens pour effectuer la transformation de Fourier-Mellin sur le visage réfé- rencé au point central entre les yeux pour produire un vecteur de traits du visage. 45 2. Système de classification de visages comprenant des premiers moyens pour localiser un visage dans une représentation bidimensionnelle d'un décor tri- dimensionnel, des deuxièmes moyens pour locali- ser le visage dans la représentation du décor, des 50 troisièmes moyens pour former une rotation, une mise à l'échelle, une translation, et une représenta- tion invariante de l'intensité du niveau de gris du visage et la production d'un vecteur de traits à partir de là et des quatrièmes moyens pour comparer le 55 vecteur de traits du visage présentement localisé au vecteur de traits d'un visage localisé précédem- ment pour déterminer si le visage présentement

7 EP 0 551 941 B1

8 EP 0 551 941 B1

9 EP 0 551 941 B1

10