Robust Face Recognton through Local Graph Matchng Ehsan Fazl-Ers, John S. Zele and John K. Tsotsos, Department of Computer Scence and Engneerng, Yor Unversty, Toronto, Canada E-Mal: [efazl, tsotsos]@cse.yoru.ca Department of Systems Desgn Engneerng, Unversty of Waterloo, Waterloo, Canada E-Mal: zele@uwaterloo.ca Abstract A novel face recognton method s proposed, n whch face mages are represented by a set of local labeled graphs, each contanng nformaton about the appearance and geometry of a -tuple of face feature ponts, extracted usng Local Feature Analyss (LFA technque. Our method automatcally learns a model set and bulds a graph space for each ndvdual. A two-stage method for optmal matchng between the graphs extracted from a probe mage and the traned model graphs s proposed. The recognton of each probe face mage s performed by assgnng t to the traned ndvdual wth the maxmum number of references. Our approach acheves perfect result on the ORL face set and an accuracy rate of 98.4% on the FERET face set, whch shows the superorty of our method over all consdered state-of-the-art methods. Index Terms Local Feature Analyss (LFA, Gabor wavelet, Prncpal Component Analyss (PCA, Gaussan Mxture Models (GMM, ORL database, FERET database. Ths wor s the extenson of the paper ttled Local Graph Matchng for Face Recognton, by E. Fazl-Ers and J. Zele, whch appeared n the proceedngs of the Eghth IEEE Worshop on Applcatons of Computer Vson 007, Austn, Texas, USA. 007 IEEE. I. INTRODUCTION In recent years face recognton has receved substantal attenton from both research communtes and the maret, but has stll remaned very challengng n real applcatons. A large number of face recognton algorthms, along wth ther modfcatons, have been developed durng the past decades whch can be generally classfed nto two categores: holstc approaches and local feature based approaches. The maor holstc approaches developed for face recognton are Prncpal Component Analyss (PCA, combned Prncpal Component Analyss and Lnear Dscrmnant Analyss (PCA+LDA, and Bayesan Intra-personal/Extra-personal Classfer (BIC. PCA [] computes a reduced set of orthogonal bass vectors, called egenfaces, from the tranng face mages. A new face mage can be approxmated by a weghted sum of these egenfaces. PCA+LDA [] provdes a lnear transformaton on PCAproected feature vectors, by maxmzng the betweenclass varance and mnmzng the wthn-class varance. The BIC algorthm [] proects the feature vector onto extra-personal and ntra-personal subspaces and computes the probablty that each feature vector came from one or the other subspace. In the local feature based approaches developed for face recognton, one wdely nfluental wor s that of Wsott et al. [4], called Elastc Bunch Graph Matchng (EBGM. By tang advantage of the fact that all human faces share a smlar topologcal structure, EBGM represents faces as graphs, wth the nodes postoned at fducal ponts (e.g., eyes, nose and the edges labeled wth the dstances between the nodes. Each node contans a set of 40 complex Gabor wavelet coeffcents at dfferent scales and orentatons, whch are called a Gabor Jet. The dentfcaton of a new face conssts of determnng among the constructed graphs, the one whch maxmzes the graph smlarty functon. In contrast to EBGM, most of the avalable feature based approaches perform sngle feature matchng for recognton (e.g., [5]. In tranng, a large set of features are extracted from the tranng mages of each ndvdual, and then n recognton a nearest neghbourhood classfer s used to assgn a tranng feature to each test feature. Each of the tranng features belong to a certan ndvdual and therefore, the probe mage s assgned to the most referenced traned ndvdual. Motvated n part by the wor of Wsott et al. [4], n ths paper we propose a novel technque for face recognton whch taes advantage of the fact that a sngle feature can be confused wth other features at a local scale; however, the ambguty s less lely f we consder groups of features. Le the wor of Wsott s group, our approach compares faces usng a combnaton of local features. However, unle that approach, we do not use a pre-defned set of features and a complex graph matchng process for locatng the features. In our technque, face mages are represented by a set of -node labeled graphs, each contanng nformaton on the appearance and geometry of a -tuple of face feature ponts, where feature ponts are extracted usng the LFA technque, and each extracted feature pont s descrbed by a Gabor Jet.
Our method automatcally learns a model set and bulds a graph space for each ndvdual. A two-stage method for fast matchng s developed, where n the frst stage a Bayesan classfer based on PCA factorzaton s used to effcently prune the search space and select very few canddate model sets, and n the second stage a nearest neghborhood classfer s used to fnd the closest model graphs to the query mage graphs. Each matched mage graph votes for the possble dentty of the probe face mage and the recognton s performed based on the number of votes each ndvdual obtans durng the matchng. The remander of ths paper s structured as follow: n Secton II, we ntroduce the -node labeled graphs and the way we extract them from face mages; the learnng and recognton phases of our method are descrbed n detal n Sectons III and IV, respectvely; n Secton V several expermental results on the ORL and FERET face datasets are reported, and fnally Secton VI concludes the paper. II. IMAGE GRAPHS Faces frequently dstngush themselves not by the propertes of ndvdual features, but by the contextual relatve locaton and comparatve appearance of these features. A tractable and effcent way for modellng ths s to employ mage graph models. Graphcal models have been successfully used n pattern recognton and computer vson as a powerful and flexble representaton mechansm (e.g., [4], [6]. In our approach, we represent face mages usng a set of local graphs wth nodes and edges, where nodes are dstnctve feature ponts of the face mage, labelled wth ther descrpton vectors, and edges (lnes connectng the nodes are labelled wth dstances between ther end nodes. In our system feature ponts are extracted usng the LFA technque, and each extracted feature pont s descrbed by a Gabor Jet. In the followng sub-sectons, we brefly descrbe the feature extracton and descrpton technques used n our system, and then dscuss the graph propertes and the way we extract the graphs from tranng and probe face mages. A. Local Feature Analyss (LFA The statstcal Local Feature Analyss (LFA technque s used n our method to extract a set of feature ponts from each face mage, at locatons wth hghest devatons from the statstcal expected face. LFA defnes a set of topographc, local ernels that are optmally matched to the second-order statstcs of the nput ensemble [4]. Gven the zero-mean matrx X of n vectorzed face mages wth normalzed energy, the egenvalues of the covarance matrx XX T are calculated and the frst largest egenvalues, λ λ, and ther assocated All face mages used to derve the LFA ernels, were frst rectfed usng eye coordnates, and cropped wth a sem-ellptcal mas to exclude non-face area. Furthermore, the grey hstograms over the face area n each face mage were equalzed. Ths preprocessng procedure s appled to all face mages (gallery/probe used n our experments (presented n Secton IV. egenvectors, e e are selected. Penev and Atc [4] defned a set of ernel, K as: K( x, y = r= er ( x er ( y λ The rows of K contan the LFA ernels, whch have spatally local propertes and are topographc n the sense that they are ndexed by spatal locaton (see Fg.. The ernel matrx K transforms X to the LFA output O = KX T, whch nherts the same topography as the nput space. LFA produces an n dmensonal representaton, where n s the number of pxels n the mage. Snce the n outputs are descrbed by <<n lnearly ndependent varables, there are resdual correlatons n the output. Penev and Atc [4] showed that the resdual correlaton of the outputs can be obtaned usng: r= r ( P( x, y = e ( x e ( y ( r Each pont x m n the output array O(x s correlated wth other outputs va P(x,x m, so t can predct the other outputs to some extent, usng the followng equaton [4]: P( x, xm O pred ( x = O( x ( P( xm, xm Penev and Atc [4] used ths property and proposed an teratve sparsfcaton algorthm for reducng the dmensonalty of the representaton by choosng a subset M of outputs that were as decorrelated as possble. At each tme step, the output pont that s predcted most poorly by multple lnear regressons on the ponts n M s added to M. The reconstructon (predcton of O(x at m th step s acheved by: wth: a O rec M m= r ( x = a ( x O( x (4 m M m ( x = P( x, xl P ( xl, xm (5 l= In our system, for an arbtrary face mage, the ponts selected through the sparsfcaton of ts LFA outputs are consdered as the most dstnctve features of that face mage, and are used for learnng and recognzng the face. The result of applyng the sparsfcaton algorthm to a sample mage s shown n Fg.. The locatons and the order of the frst 0 ponts show that the outputs wth the largest devatons from the expectaton (e.g., the most unusual features of the face are selected frst. B. Gabor Jet Each feature pont n our system s descrbed by a Gabor Jet,.e., a set of convoluton coeffcents for Gabor wavelet ernels of dfferent orentatons and frequences at the locaton of the feature n the mage. Gabor wavelets m
are bologcally motvated convoluton ernels n the shape of plane waves, restrcted by a Gaussan envelope functon [4]. The general form for a D Gabor wavelet s: 0 7 65 8 6 9 0 r,, r r s o s o = r σ ψ s, o ( exp exp( exp( s, o (6 σ σ where σ s a parameter to control the scale of the Gaussan (n our experments σ = π and s,o s a D wave vector whose magntude and angle determne respectvely the frequency, s, and the orentaton, o, of the Gabor ernel. Smlar to [4], n our method wavelet responses at 5 frequences (n whch s,o = {π/, π/ 8, π/4, π/, π/8} and 8 orentatons (varyng n ncrements of π/8 from 0 to 7π/8 are used for descrpton, resultng n 40-element descrpton vectors. C. Graph Propertes The most mportant and dstnctve property of each graph s ts appearance, whch s computed from the descrpton vectors of the graph nodes. To compute ths, one opton s to concatenate the descrpton vectors (Gabor Jet of the graph nodes nto a sngle descrpton vector, representng the appearance of the graph. 9 8 (a (b (c (d (e (f Fgure. (a shows the orgnal mage, (b llustrates the frst 0 selected ponts overlad on the orgnal mage and numbered sequentally, (c llustrates the frst 50 selected ponts, (d shows the LFA output of the orgnal mage,, (e shows the reconstructed output, and (f shows the reconstructon error scaled from 0 to 55, respectvely. 5 4 7 4 (a (b (c (d Fgure. K(x,y and P(x,y derved from a set of 96 face mages. The average of the face mages (the statstcal expected face s shown n (a, mared wth four postons -4, (b shows the frst four derved PCA ernels, (c and (d show K(x,y and P(x,y respectvely, at the four mared postons. 4 Although ths technque s smple and straghtforward, the hgh dmensonalty of the resultng appearances (x40 restrcts ts applcablty n the matchng, partcularly when the dataset of model graphs, created from the tranng mages, s very large. A very powerful alternatve to ths s to model the ont dstrbuton of the nodes appearances through a probablstc framewor. Usng ths technque, not only s the graph descrbed based on the descrpton vectors of ts ndvdual nodes, the spatal relatonshps between the appearances of the graph nodes and the nteracton between them are also taen nto account. Besdes the power of ths appearance modellng scheme, a very fast matchng s also nherently acheved. We wll dscuss further the modellng of the appearance of the graphs n more detal n Secton III. Another graph property whch s used n our system as a descrptor for the graphs s graph geometry. Graph geometry,.e., the way the three nodes n a graph are arranged spatally, could play an mportant role n dscrmnatng graphs, partcularly when the use of appearance alone causes some ambgutes. In our approach, the lengths of graph edges are used to descrbe the geometrcal propertes of each graph. The lengths of graph edges, e, are smply computed by measurng the Eucldean dstances between the locatons of each two end nodes. D. Image Graph Extracton Consderng that we extract around 50 feature ponts from each face mage, approxmately.e+6 (50x49x48 -node graphs could be generated for each mage. Evaluatng ths number of graphs for each probe mage would be very computatonally expensve. On the other hand, even n the learnng phase, where the computatonal tme s not usually a crucal ssue, ths
could be a problem for the modellng of the appearance of the graphs (Secton III.A. Ths necesstates the selecton of only a subset of all possble graphs for the subsequent processng. One potental way s to defne a graph radal threshold, R th, and then consder only those graphs that the dstance between each node of the graph wth respect to the graph centre s lower than R th (n our system R th s set to a fxed value of 0, where the sze of the face mage s x9. Ths method s effcent, snce t selects only those graphs n whch the nodes are suffcently close to each other, and also t averagely selects around 0.6% of the graphs, whch s a very good reducton rate. However, ths technque s reasonable to be used for graph extracton n the tranng stage only, not n the run tme (recognton; snce computng the graph centre and edge dstances for all graphs of a probe mage, s stll a tme consumng process. Therefore we need to fnd another graph selecton method to be appled to the probe mages. Snce our graphs have three nodes, the most straghtforward way for dong ths s to apply a trangulaton technque. Ths has the advantage of selectng very small number of graphs whch can be used as the representatves of the probe mages. III. LEARNING In ths secton, we develop and dscuss our approach for learnng model graphs, extracted from tranng face mages. In practce, gven M tranng face mages for N ndvduals, our learnng algorthm wors as follow:. Intalze N empty model sets.. For each tranng mage I : a. Extract and descrbe ts feature ponts (see Secton II.A and II.B. b. Extract the mage graphs (see Secton II.D. c. Place the extracted graphs n one of the N model sets, accordng to the dentty of the tranng face mage I.. Construct an appearance model for each created model set from ts model graphs. All steps of the above algorthm have been dscussed before, except the last step whch s the constructng of an appearance model for each created model set. In the followng we explan ths step n detal. A. Constructng Appearance Models Gven G = {( x, x, x }, representng the th -node graph extracted from a probe face mage, where x, x A trangulaton of a dscrete set of ponts P s a subdvson of the convex hull of the ponts nto smplces such that any two smplces ntersect n a common face or not at all and the set of ponts that are vertces of the subdvdng smplces concdes wth P. Trangulaton could not be a good choce for extractng graphs n the tranng stage, because then the number of extracted graphs from tranng mages would not be suffcent for reasonably estmatng the graph spaces (dscussed n Secton. and x are the descrpton vectors of the graph nodes 4 ; we are nterested n obtanng the probablty of G belongng to each model set (ndvdual to assgn t to the one that maxmzes the posteror probablty: R MAP = argmax P({( x, x, x } C n n Ths s called the Maxmum a Posteror (MAP soluton 5. In the above equaton, C n s the appearance model or graph space for the n th ndvdual. The am of ths secton s to develop a method for estmatng C n for each ndvdual, usng the model graphs extracted from ts tranng mages. Modellng the total ont lelhood of all -node graphs of a model set, that s, to construct the graph space (appearance model, becomes a (xd-dmensonal dstrbuton, where D s the length of the descrpton vector of each node, whch s 40 n our approach (see Secton II.B. To mae the appearance model estmaton more accurate and tractable, we need to apply a dmensonalty reducton technque over the set of all descrpton vectors, extracted from tranng mages of all ndvduals. Prncpal Component Analyss (PCA s a standard technque for dmensonalty reducton and has been appled to a broad class of computer vson and machne learnng problems. Whle PCA suffers from a number of shortcomngs such as ts mplct assumpton of Gaussan dstrbutons and ts restrcton to orthogonal lnear combnatons, t remans a popular method for dmensonalty reducton due to ts smplcty and low computatonal tme. Gven the set of all descrpton vectors, X, PCA calculates the egenvalues of the covarance matrx XX T, and selects the frst largest egenvalues and ther assocated egenvectors to form the PCA proecton matrx W. Usng the PCA, the orgnal 40-dmensonal descrpton vectors X are factorzed nto -dmensonal vectors S, where S=W*X. In our system we expermentally chose to be 8, therefore the ont lelhood of the -node graphs becomes (x8- dmensonal rather than (x40-dmensonal, whch s more manageable and tractable. Gven s the factorzed descrpton vector of x, we now estmate the P({ s s s }, rather than estmatng,, P({ x, x, x }. Ths can be modelled for each model set by usng parametrc or non-parametrc dstrbuton estmaton technques. In our approach, we estmate the ont dstrbutons of graph appearances for each ndvdual by usng a Gaussan Mxture Model (GMM as a parametrc approxmaton technque. In a Gaussan Mxture Model [0], dfferent Gaussan dstrbutons represent dfferent domans of the data, and have dfferent output characterstcs; GMMs try to descrbe a complex system usng combnaton of all the Gaussan clusters, 4 Generally G should be represented by ts all propertes, however snce n ths secton we are only worng wth the descrpton vectors of ts nodes, we represent G as: G = {( x, x, x }. 5 Gven that the classes Cn have equal pror probablty. (7
nstead of usng a sngle Model. In mathematcal terms, a GMM can be defned as: where, K P( x θ = w. N( x μ, σ (8 = ( x μ σ ( x μ, σ =. e N σ π are the components of the mxture, K s the number of components (Gaussans n the mxture model, K θ =, μ, σ } conssts of the means and varances { w = of the Gaussans, and w s the weght of Gaussan [0]. Snce the famly of mxtures of Gaussans s parametrc, the densty estmaton problem can be defned more specfcally as the problem of fndng the parameter vector θ that specfes the models from whch the data are most lely to be drawn (the mxng weght vector w should also be estmated. Among the avalable technques for estmatng the parameters of Gaussans, we choose to use the standard Expectaton-Maxmzaton (EM technque []. The EM algorthm s an effcent teratve procedure to compute the Maxmum Lelhood (ML estmate n the presence of mssng or hdden data. In ML estmaton, we wsh to estmate the model parameters (Ө for whch the observed data are the most lely: * θ = arg max P( x θ (9 θ Each teraton of the EM algorthm conssts of two processes: The E-step (or expectaton step, and the M- step (or maxmzaton step. In the E-step, gven the observed data and current estmate of the model parameters, the mssng data are estmated. Ths s acheved usng the condtonal expectaton. In the M-step, the lelhood functon s maxmzed under the assumpton that the mssng data are nown (For more detals about the EM algorthm see []. Once all the graph spaces for each ndvdual s model set are estmated through Gaussan mxture models and Expectaton-maxmzaton technque, the learnng s done and matchng can be performed based on the constructed appearance models. IV. MATCHING AND RECOGNITION In ths secton, a two-stage method for optmal matchng between the graphs extracted from a probe mage and the traned model graphs s developed, where n the frst stage a MAP soluton s used to effcently select the most lely ndvduals model sets based on the appearances of the graphs, and n the second stage, a nearest neghbourhood classfer s used to enable correspondence wth learned model graphs of the selected ndvduals model sets, by ncorporatng the geometry of the graphs. In practce, gven N model sets and ther correspondng appearance models, where N s the number of tranng ndvduals, the followng algorthm s appled on the probe mage for matchng and recognton:. Extract and descrbe the mage s feature ponts (see secton II.A and II. B.. Calculate the factorzed descrpton vectors for the extracted feature ponts by usng the PCA proecton matrx, W, learned durng the tranng (see secton III.A.. Extract the mage graphs (see secton II.B. 4. For each graph G, = N G, (where N G s the total number of extracted graphs from the probe mage: a. Obtan the probablty of G belongng to each appearance model, C n, P(G C n, and select the r model sets wth hghest P(G C n - r n our system s 5% of N. b. Incorporate the geometrcal propertes of G and search the learned nstances (model graphs of the selected model sets, pcng the model graph wth hghest smlarty to the test graph (Graph smlarty functon n terms of appearance and geometry s descrbed later n ths secton. c. Vote for the dentty of the model set that one of ts model graphs s matched to the consdered test graph. 5. Select the ndvdual wth the maxmum number of votes. All steps of the above algorthm have been dscussed before, except step 4.b, whch requres a smlarty functon based on appearance and geometry of the graphs, to enable correspondence wth learned nstances of the selected model sets. Gven G = {( x, x, } x, a test graph, and G = {( x, x, } x, a model graph, the smlarty between the appearances of the graphs can be smply calculated by averagng the smlarty between the correspondng nodes descrpton vectors ( x wth x for =. By employng the cosne smlarty measure 6, the appearance smlarty functon can be formulated as: Y ( G, G = x. x = ( x. x.( x. x (0 As dscussed n Secton II.C, besdes the appearance, the geometry can be also used as a descrptor for a graph. In our system, the lengths of graph edges are used to descrbe the geometrcal propertes of each graph. Now, gven G = {( e, e, } e, a test graph, and G = {( e, e, } e, a model graph, the dssmlarty between the geometry of the graphs can be measured by: ( e = e Z( G, G ( e 6 x. x S( x, x = ( x. x.( x. x (
By combnng Equatons 6 and 7, the fnal graph smlarty functon taes the form: S G, G = Y( G, G αz( G, G ( ( where α determnes the relatve mportance of appearance and geometry smlartes (n our system α =0.. V. EXPERIMENTS In order to valdate the robustness of our technque for face recognton, several experments were performed on two of the publcly avalable and wdely used face datasets, the ORL [] and the FERET [6]. Our frst experment was performed on the ORL face set. ORL contans 400 mages from 40 ndvduals (0 for each, wth varatons n pose, facal expresson and a certan amount of scale and vewpont (see Fg. for an example. The frst fve mages for each ndvdual have been used for tranng and the other fve mages for testng. PCA was appled to a dataset of 0,000 feature descrpton vectors extracted from tranng mages (50 features form each of the 40x5 tranng mages, and the dmensonalty of feature vectors was reduced from 40 to 8 (wth 7.% of non-zero egenvalues retaned. In graph space estmaton, mxture models wth 0 Gaussans were used for each ndvdual, to model the ont dstrbutons of the nodes factorzed descrpton vectors. Our system acheved an accuracy rate of 00% n ths experment on the ORL face dataset (see [7] for a comparatve study of several face recognton technques on the ORL dataset. Although the ORL dataset s one of the most challengng publcly avalable face datasets, the small number of contaned ndvduals (40 cannot provde a good estmate of the ablty of a face recognton system n worng wth datasets wth large number of ndvduals (e.g., greater than 000. To ths am, we have tested our face recognton algorthm on the FERET test set. Two FERET mage sets were used n our experments: FA, whch contans frontal mages of 96 subects (one mage per person; and FB, whch contans frontal mages of 95 of the subects avalable n FA, wth an alternatve facal expresson than n FA photographs. We detected the faces manually from all mages and reszed them to x9 pxels. The face mages n FA were used for tranng and the FB mages for testng the system. The performance of our technque n ths experment s shown n Table I 7, n comparson wth the results of fve state-ofthe-art methods: Elastc Bunch Graph Matchng (EBGM, LDA+PCA, Bayesan Intra-personal/Extra-personal Classfer (BIC, Boosted Haar Classfer [], and the wor of Tmo et al. [5] based on the LBP (Local Bnary Patterns texture analyss. The comparsons ndcate that our method acheved better results than the other evaluated methods. 7 When comparng the results n Table I, note that almost smlar face normalzaton algorthms were used n all technques, except [4]. Fgure. Images of an ndvdual from ORL. The frst 5 mages are used for tranng and the other 5 for testng. TABLE I. COMPARISON OF OUR RESULT ON THE FERET DATASET WITH THE RESULTS OF SEVERAL STATE-OF-THE-ART FACE RECOGNITION METHODS Methods Recognton Rates (EBGM - Wssott et al. [4] 95.5% (LDA + PCA - Etemad et al. [] 96.% (BIC Moghaddam et al. [] 94.8% (Boosted Haar Jones et al. [] 94.0% (LBP - Tmo et al. [5] 97% Our method 98.4% The computatonal tme of our method drectly depends on the number of features, extracted from each face mage, as the most tme consumng process of our method s feature extracton. The computatonal tme for extractng 50 features from a face mage taes about.4 second on a. GHz CPU. However, n contrast to most of the avalable face recognton technques, the overall computatonal tme of our recognton method does not depend very much on the number of ndvduals learned, whch s an advantage of our method. VI. CONCLUSIONS Face recognton, because of ts many applcatons n automated survellance and securty, has garnered a great deal of attenton. Whle there have been many papers publshed n ths area, much of the debate has now moved outsde of the academc area. Snce detals of many of the best commercal algorthms are not publcly avalable, t can be dffcult to compare results or gauge progress. In ths paper we presented a novel technque for face recognton whch represent face mages by a set of - node labeled graphs, each contanng nformaton on the appearance and geometry of a -tuple of face feature ponts. Our method automatcally learns a model set and bulds a graph space for each ndvdual. A two-stage method for fast matchng s used n recognton, where n the frst stage a MAP soluton based on PCA factorzaton s used to effcently prune the search space and select very few canddate model sets, and n the second stage a nearest neghbourhood classfer s used to fnd the closest model graphs to the query mage graphs. Our proposed technque acheves perfect results on the ORL face set and an accuracy rate of 98.4% on the FERET face set, whch shows the superorty of the proposed technque over all consdered state-of-the-art methods.
ACKNOWLEDGMENT The authors would le to acnowledge the Ontaro Centre of Excellence (OCE and the Natonal Scence and Engneerng Research of Canada (NSERC for ther partal support. REFERENCES [] M. Tur and A. Pentland. Egenfaces for Recognton, Journal of Cogntve Neuroscence, 99, Vol., pp. 7-86. [] K. Etemad and R. Chellappa, Dscrmnant Analyss for Recognton of Human Face Images, Journal of the Optcal Socety of Amerca, 997, Vol. 4, pp. 74-7. [] B. Moghaddam, C. Nastar and A. Pentland, Bayesan Face Recognton usng Deformable Intensty Surfaces, In Proceedngs Computer Vson and Pattern Recognton 96, 996, pp. 68-645. [4] L. Wsott, J.-M. Fellous, N. Kruger and C. Von Der Malsburg, Face Recognton by Elastc Bunch Graph Matchng, IEEE Trans. on Pattern Analyss and Machne Intellgence, 997, Vol. 9, No. 7, pp.775-779. [5] E. Fazl Ers, J.S. Zele, Local Feature Matchng for Face Recognton, In Proceedngs of the rd Canadan Conference on Computer and Robot Vson (CRV'06, 006. [6] P. Felzenszwalb and D. Huttenlocher, Pctoral Structures for Obect Recognton, In Proceedngs of the IEEE Conference on Computer Vson and Pattern Recognton, pp. 066-07, 000. [7] C. Harrs and M. Stephens, A Combned Corner and Edge Detector, n Alvey Vson Conference, pp. 47-5, 998. [8] B. Trggs, Detectng Keyponts wth Stable Poston, Orentaton and Scale under Illumnaton Changes, n Egth European Conference on Computer Vson, 004. [9] L. Gubas, J. Stolf, Prmtves for the Manpulaton of General Subdvsons and the Computaton of Vorono, ACM Transactons on Graphcs, vol. 4, no., pp.74-, 985. [0] J. McLachlan and D. Peel, Fnte Mxture Models New Yor, John Wley & Sons Ltd., 000. [] R. A. Redner and H. F. Waler, Mxture Denstes, Maxmum Lelhood and the EM Algorthm, SIAM Revew, vol. 6, no., pp. 95-9, 984. [] F. Samara and A. Harter. Parametersaton of a Stochastc Model for Human Face Identfcaton, Proceedngs of nd IEEE Worshop on Applcatons of Computer Vson, 994. [] M. J. Jones and P. Vola, "Face Recognton usng Boosted Local Features", In Proceedngs of Internatonal Conference on Computer Vson, 00. [4] P.S. Penev et al., Local Feature Analyss: a General Statstcal Theory for Obect Representaton. Networ: Computaton n Neural Systems, 996, 7(: 477-500. [5] A. Tmo, H. Abdenour, and P. Matt, Face Recognton wth Local Bnary Patterns, Proceedngs of the ECCV, pp. 469-48, 004. [6] P. J. Phllps et al., The FERET Database and Valuaton Procedure for Face Recognton algorthms, Image and Vson Computng, 998, 6(5: 95-06. [7] L et al., A Novel Face Recognton Method wth Feature Combnaton, Journal of Zheang Unversty: Scence, 005, 6(5. Ehsan Fazl-Ers receved an undergraduate degree n Computer Engneerng from the Azad Unversty of Mashad, Mashad, Iran, n 004, and a Master s degree n Systems Desgn Engneerng from the Unversty of Waterloo, Waterloo, Canada n 006. He s now pursung a Ph.D. degree n Computer Scence at the Yor Unversty, Toronto, Canada. Hs research nterests nclude computer and robot vson, partcularly, obect detecton and recognton, and moble robot localzaton and navgaton. John S. Zele s an Assocate Professor n the Systems Desgn Engneerng department at the Unversty of Waterloo. He s also co-drector of the Intellgent Human-Machne Interface lab. Hs research nterests can be best summarzed as beng n the area of ntellgent Mechatronc control systems that nterface wth humans; specfcally, the areas are ( wearable sensory substtuton and assstve devces; ( probablstc vsual and tactle percepton; ( wearable haptc devces ncludng ther desgn, synthess and analyss; and (4 humanrobot nteracton. He s a co-founder of Tactle Sght Inc. He was awarded wth the 004 Young Investgator Award by the Canadan Image Processng & Pattern Recognton socety for hs wor n robotc vson. He has also been awarded on numerous occasons wth the best paper n varous conferences. Hs wor has also been featured n over 0 televson, newspaper and rado stores. John K. Tsotsos receved an honours undergraduate degree n Engneerng Scence n 974 from the Unversty of Toronto and contnued at the Unversty of Toronto to complete a Master's degree n 976 and a Ph.D. n 980 both n Computer Scence. In 980 he oned the faculty n Computer Scence at the Unversty of Toronto where he founded the computer vson research group n the department. He served two 5-year terms as Fellow of the Canadan Insttute for Advanced Research. Hs research focuses on the computatonal complexty of bologcal vsual percepton and the development of an accompanyng theory of vsual attenton that maes strong predctons (and now wth extensve supportng expermental evdence about the psychophyscs and neurobology of human and prmate percepton. Another research thrust s the practcal applcaton of ths wor n the development of PLAYBOT, a vsually-guded robot to assst physcally dsabled chldren n play. He moved to Yor Unversty n Toronto n January 000 where he was apponted Professor n the Department of Computer Scence and Engneerng. He was also named Drector of Yor's Centre for Vson Research, a poston he held untl November 006. He currently holds an NSERC Ter I Canada Research Char n Computatonal Vson and s also an Adunct Professor n Ophthalmology and n Computer Scence the Unversty of Toronto. He has served on numerous conference commttees and on the edtoral boards of Image and Vson Computng Journal, Computer Vson and Image Understandng, Computatonal Intellgence and Artfcal Intellgence and Medcne. He served as the General Char for the IEEE Internatonal Conference on Computer Vson 999. Recent edtoral actvtes nclude a Specal Issue on Attenton and Performance for Computer Vson and Image Understandng (wth L. Paletta, R. Fsher and G. Humphreys, and Neurobology of Attenton for Elsever Press (wth L. Itt and G. Rees.