2D face recognition survey.pdf

8/9/2019 2D face recognition survey.pdf

http://slidepdf.com/reader/full/2d-face-recognition-surveypdf 1/10

International Journal of Machine Intelligence, ISSN: 0975–2927, Volume 2, Issue 1, 2010, pp-74-83

Copyright © 2010, Bioinfo Publications, International Journal of Machine Intelligence, ISSN: 0975–2927, Volume 2, Issue 1, 2010

2D Face Recognition Techniques: A Survey

Patil A.M.1, Kolhe S.R.

2 and Patil P.M.

3

1Department of Electronics and Telecommunication Engineering, J.T.Mahajan College of Engineering,

Faizpur, India, E-mail: [email protected] of Computer Science, North Maharashtra University, Jalgaon, India, [email protected]

3Department of Electronics Engineering, Vishwakarma Institute of Technology, Pune,

[email protected]

Abstract- Numerous recent events, such as radical attacks, exposed stern weakness in most sophisticatedsecurity systems. Various supervision agencies are now more motivated to improve security data based onbody or behavioral distinctiveness, often called biometrics. Pin or password-based authenticationprocedures are too easy to lacerate. Face recognition is non-intrusive since it is based on images based onrecorded by a distant camera, and can be very effective even if the user is not aware of the existence of theface recognition system. The human face is undoubtedly the most common characteristic used by humansto recognize other people. Face recognition presents a challenging problem in the field of PatternRecognition. It has wide range of applications such as; law enforcement, banking, logical access control,immigration, national identity. In this paper, an overview of some of the well- known methods in variouscategories are discussed.Keywords: Biometric, Face recognition, PCA, LDA, ICA, Neural Networks, Pose, Illumination.

Introduction:Face recognition is one of the most importantbiometric which seems to be a good compromisebetween actuality and social reception andbalances security and privacy well. Also it has avariety of potential applications in informationsecurity, law enforcement, access controls[1].Face recognition system fall into twocategories: verification and identification. Faceverification is a 1:1 match that compares a faceimages against a template face images, whoseidentity being claimed .On the contrary, faceidentification is a 1: N problem that compares aquery face image against all image templates in aface database [2]. In last decade, major

advances occurred in face recognition, with manysystems capable of achieving recognition ratesgreater than 90%. However real-world scenariosremain a challenge, because face acquisitionprocess can undergo to a wide range ofvariations. There are five key factors that cansignificantly affect system face recognitionperformances:• Illumination variations due to skin reflectanceproperties and due to the internal camera control.Several 2D methods do well in recognition tasksonly under moderate illumination variation, whileperformances noticeably drop when bothillumination and pose changes occur.• Pose changes affect the authentication process,

because they introduce projective deformationsand self-occlusion. Even if methods dealing withup to 320 head rotation exists, they do not solvethe problem considering that security camerascan create viewing angles that are outside of thisrange when positioned. On the contrary, withexception of extreme expressions such asscream, the algorithms are relatively robust tofacial expression.• Another important factor is the time delay,because the face changes over time, in a

nonlinear way over long periods. In general thisproblem is harder to solve with respect to theothers and not much has been done especiallyfor age variations.• At last, occlusions can dramatically affect facerecognition performances, in particular if theylocated on the upper-side of the face, asdocumented in literature.

Categorization of methods for facerecognition Face recognition algorithms are divided by [3]into three categories, as follows

1. Holistic methods:- these methods identify aface using the whole face images as a input themain challenge faced by these methods is how toaddress the extremely small size problem.2. Feature – based methods:- these methodsused the local facial features for recognition.Care should be taken when deciding how toincorporate global configurationally informationinto local face methods3. Hybrid methods:- these methods used bothfeature-based and holistic features to recognize aface. These methods have the potential to offerbetter performance than individual holistic orfeature based method [1]In order to evaluate how well proposed method

works, several databases of face images havebeen built. A table 1 gives some details about thedatabases.

Automatic face recognition(Linear, nonlinear projection)Automatic face recognition “Fig (1)” can be seenas pattern recognition problem which is hard tosolve due to its nonlinearity. We can think of it asa template matching problem, where recognition




International Journal of Machine Intelligence, ISSN: 0975–2927, Volume 2, Issue 1, 2010 75

has to be performing in a high dimensionalityspace. Since higher the dimension of space is,more the computations are needed to find thematch. A dimensionality reduction technique isused to project the problem in a lowerdimensionality space. Eigen faces [4] can beconsidered as one of the first approaches in this

sense Kirby and Sirovich (1990) adopted thePCA. Nevertheless, one important date for facerecognition was beginning of the 90’s when Turkand Pentland implemented the Eigenfacesapproach [5, 6], which is surely the most popularface recognition method. This was the beginningof the appearance based methods for facerecognition [6, 7, 8, 9, 10, 11, 12]. AfterEigenfaces, different statistical approaches haveappeared that improve the results of Eigenfacesunder certain constraints. Many holistic methodsare based on Eigen face decomposition. Hereface images are represented as vectors byconcatenating the pixels of the image line-by-line.Then the average vector is computed that

represents a mean face. Also, a difference vectoris computed for each user to qualify thedifferences to the mean face. Then thecovariance matrix of the difference vectors iscomputed. Finally, principal axes can be obtainedby eigen decomposition of covariance matrix.The first N eigenvectors presenting the highesteigen values will be retained and represents themost significant features of faces. Finally, eachuser model is represented as a linearcombination (weighted sum) of coefficientscorresponding to each eigenface [13]. As thePCA “Fig (2)” is performed only for training thesystem, this method results to be very fast, whentesting new face images.

The LDA (Linear Discriminant Analysis [14, 15]has been proposed as a better alternative to thePCA. It expressly provides discrimination amongthe classes, while the PCA deals with the inputdata in their entirety, without paying any attentionfor the underlying structure. Indeed the main aimof the LDA consists in finding a base of vectorsproviding the best discrimination among the

classes, trying to maximize the between-classdifferences, minimizing the within-class ones.

The between- and within-class differences arerepresented by the corresponding scattermatrices Sb and Sw, while the ratio

detjSbj/detjSwj has to be maximized. Even if theLDA is often considered to outperform the PCA,an important qualification has to be done. Indeedthe LDA provides better classificationperformances only when a wide training set isavailable, and some results discussed by[15],confirm this thesis. Besides recent studies alsostrengthen this argument expressly tackling thisproblem referred to as the SSS (Small SampleSize) problem. In some approaches, such as theFisherfaces [16], the PCA is considered as apreliminary step in order to reduce thedimensionality of the input space, and then theLDA “Fig (3)” is applied to the resulting space, inorder to perform the real classification. However

it has been demonstrated in recent works [17, 18]that, combining in this way PCA and LDA;discriminant information together with redundantone is discarded. Thus, in some cases the LDA isapplied directly on the input space, as in [17, 18].Lu et al[14] proposed an hybrid between theDLDA (Direct LDA) and the F-LDA (FractionalLDA), a variant of the LDA, in which weighedfunctions are used to avoid that output classes,which are too close, can induce amisclassification of the input. The maindisadvantage of the PCA, LDA, and Fisherfacesis their linearity. Particularly the PCA extracts alow-dimensional representation of the input dataonly exploiting the covariance matrix, so that nomore than first- and second order statistics areused. Independent component analysis (ICA) iscurrently popular in the field of signal processing;it has been developed recently as an effectivefeature extraction technique and has beenapplied to image discrimination.[19,20] proposedusing ICA for face representation and found thatit was better than PCA when cosine was used asthe similarity measure [19,20,21,22]. Petridis andPerantonis revealed the relationship between ICA



Patil AM, Kolhe SR and Patil PM

Copyright © 2010, Bioinfo Publications, International Journal of Machine Intelligence, ISSN: 0975–2927, Volume 2, Issue 1, 2010 76

and LDA from the viewpoint of mutual information[23]. In[19] show that first- and second orderstatistics hold information only about theamplitude spectrum of an image, discarding thephase-spectrum, while some experiments bringout that the human capability in recognizingobjects is mainly driven by the phase-spectrum.

This is the main reason for which in[19] the ICAare introduced as a more powerful classificationtool for the face recognitionproblem.

The ICA can be considered as a generalization ofthe PCA, but providing three main advantages:(1) It allows a better characterization of data in ann-dimensional space; (2) the vectors found by theICA are not necessarily orthogonal, so that theyalso reduce the reconstruction error; (3) theycapture discriminant features not only exploitingthe covariance matrix, but also considering thehigh-order statistics. Besides linear projectionanalysis technologies, non-linear projectionanalysis represented by both KPCA and KFDalso has aroused considerable interest in thefields of pattern recognition and machinelearning, and over the last few years have showngreat potential in biometrics applications. KPCAwas originally developed by chölkopf [24], whileKFD was first proposed by Mika [25,26].Subsequent research saw the development of aseries of KFD algorithms [27,28,29,30,31,32,33,34,35,36,37]. The KFD algorithmsdeveloped by Mika, Billings and Cawley areformulated for two classes, while those ofBaudat, Roth and Yangare formulated formultiple classes. Because of its ability to extractthe most discriminatory on-linear features, KFD

has been found very effective in many real-worldbiometrics applications. Yang, Liu, Yang, and Xuused KPCA (KFD) for face feature extraction andrecognition and showed that KPCA (KFD)outperforms the classical PCA (LDA) [38, 39, 40,41, 42, 43].

The neural networksA further nonlinear solution to the facerecognition problem is given by the neuralnetworks, largely used in many other patternrecognition problems, and readapted to cope thepeople authentication task. The advantage ofneural classifiers over linear ones is that they canreduce misclassifications among theneighborhood classes. The basic idea is toconsider a net with a neuron for every pixel in theimage. Nevertheless, because of the patterndimensions (an image has a dimension of about112 × 92 pixels) neural networks are not directlytrained with the input images, but they arepreceded by the application of such a

dimensionality reduction technique. A firstsolution to this problem has been given by[44],which introduced a second neural net, thatoperates in auto-association mode. At first, theface image, represented by a vector x, isapproximated by a new vector h with smallerdimensions by the first network (auto-association), and then h is finally used as inputfor the classification net. Cottrell and Fleming [44]also shown that this kind of neural network doesnot behave better than the Eigenfaces even if inoptimal circumstances. Other kind of neuralnetworks also have been tested in facerecognition, in order to exploit their particularproperties. For examples Self Organizing Map

(SOM) are invariant with respect to minorchanges in the image sample, while convolutionalnetworks provide a partial invariance with respectto rotations, translations and scaling. In general,the structure of the network is strongly dependenton its application field, so that different contextsresult in quite different networks. In a recentwork, [45] presented the Probabilistic DecisionBased Neural Network, which they modeled forthree different applications (a face detector, aneyes localizer and a face recognizer). Theflexibility of these networks is due to theirhierarchical structure with nonlinear basisfunctions and a competitive credit assignmentscheme, which shown the capability ofrecognizing up to 200 people. At last, [46]introduced a hybrid approach, in which, throughthe PCA, the most discriminating features areextracted and used as the input of a RBF neuralnetwork. The RBFs perform well for facerecognition problems, as they have a compacttopology and learning speed is fast. In their workthe authors also face the problem of the overfitting: the dimension of the network input iscomparable to the size of the training set; of the





overtraining: high dimension of the input resultsin slow convergence, small sample size: thesample size has to exponentially grow for havinga real estimate of the multivariate densities whenthe dimension increases; the singular problem: ifthe number of training patterns is less than thenumber of features, the covariance matrix is

singular. In general, neural networks basedapproaches encounter problems when thenumber of classes increases. Moreover, they arenot suitable for a single model image recognitiontask, because multiple model images per personare necessary in order for training the system to‘‘optimal’’ parameter setting.

Gabor filters and waveletsThe Gabor filters represent a powerful tool bothin image processing and image coding, thanks totheir capability to capture important visualfeatures, such as spatial localization, spatialfrequency and orientation selectivity. In the mostcases the Gabor filters are then used to extract

the main features from the face images. Indeed,in [47] they have been applied to specific areas ofthe face region, corresponding to nodes of a rigidgrid. In each node of the grid the Gaborcoefficients are extracted and combined in jets.The nodes are linked to form such a DynamicLink Architecture, so that the comparisons amongdifferent subjects can be made by means of agraph matching strategy. [48] Further expandedon DLA and developed a Gabor wavelet basedelastic bunch graph matching method (EBGM) tolabel and recognize human faces. Furthermore,comparisons are made in two consecutive steps:a rigid alignment of the grid only accounts forglobal transformations, such as translations and

scale, then the local misplacement of the gridnodes is evaluated by means of a GraphSimilarity Function. Generally, dynamic linkarchitecture is superior to other face recognitiontechniques, in terms of rotation invariant;however, the matching process iscomputationally expensive. Perronnin andDugelay [49] proposed a further deformablemodel, whose philosophy is similar to the EBGM.They introduced a novel probabilistic deformablemodel of face mapping, based on a bi-dimensional extension of the 1D-HMM (HiddenMarkov Model). Given a template face FT, aquery face FQ and a deformable model M, theproposed method try to maximize the likelihoodP(FTjFQ,M). There are two main differencesbetween this method and the original EGM. Firstof all the HMM is extended to the 2D case toestimate P (FTjFQ, M), automatically training allthe parameters of M, so taking into account forthe elastic properties of the different parts of theface. Secondly, the model M is shared among allfaces, so the approach works well also when littleenrolment data is available. On the contrary, aquite different approach has been proposed by

[49]. A mother wavelet is defined and forty Gaborfilters are derived, considering five scales andeight orientations. Each of these filters isconvolute with the input image, resulting in fortyfiltered copies of the face image. To encompassall the features produced by the different Gaborkernels, the resulting Gabor wavelet features are

concatenated to derive an augmented Gaborfeature vector. Then, in order to reduce thedimensionality of the feature vector, both thePCA and the Enhanced Fisher LinearDiscriminant Model (EFM) are investigated. Theuse of Gabor filters renders this method veryrobust to changes in expression and illumination;however they dramatically increase thecomputational cost of the method, requiring thateach kernel is convolved with the input image. Afaster wavelet based approach has beenproposed by Garcia et al. [50], which presented anovel method for recognition of frontal views offaces under roughly constant illumination. It isbased on the analysis of a wavelet packet

decomposition of the face images, because veryfast implementations of this procedure areavailable in hardware. Each face image is firstlocated and then described by a subset of bandfiltered images containing wavelet coefficients.From these wavelet coefficients, whichcharacterize the face texture, they build compactand meaningful feature vectors, using simplestatistical measures. Then, they show how anefficient and reliable probabilistic metric derivedfrom the Bhattacharrya distance can be used inorder to classify the face feature vectors intopersons classes, so that even very simplestatistical features can provide a good basis forface classification. Then, in order to reduce the

dimensionality of the feature vector, both thePCA and the Enhanced Fisher LinearDiscriminant Model (EFM) are investigated. Theuse of Gabor filters renders this method veryrobust to changes in expression and illumination;however they dramatically increase thecomputational cost of the method, requiring thateach kernel is convolved with the input image. Afaster wavelet based approach has beenproposed by [50]which presented a novel methodfor recognition of frontal views of faces underroughly constant illumination. It is based on theanalysis of a wavelet packet decomposition of theface images, because very fast implementationsof this procedure are available in hardware. Eachface image is first located and then described bya subset of band filtered images containingwavelet coefficients. From these waveletcoefficients, which characterize the face texture,they build compact and meaningful featurevectors, using simple statistical measures. Then,they show how an efficient and reliableprobabilistic metric derived from theBhattacharrya distance can be used in order toclassify the face feature vectors into person





classes, so that even very simple statisticalfeatures can provide a good basis for faceclassification.

Open questions in face recognitionThe Automatic Face Recognition (AFR) can bethought as a very complex object recognition

problem, where the object to be recognized is theface. This problem is even more difficult to solve,since the search is done among objectsbelonging to the same class. The sensibility ofthe classifiers to illumination and pose variationsare the main problems researchers have beenfacing until now, while a smaller effort has beenmade to cope with occlusions and age variationproblems. Therefore, recent works can beclassified depending on their main contribution inorder to address some of these problems.

The changes in illuminationAmbient lighting changes greatly within andbetween days and among indoor and outdoor

environments. Due to the 3D structure of theface, a direct lighting source can cast strongshadows that accentuate or diminish certainfacial features. It has been shown experimentallyand theoretically for systems based on PrincipalComponent Analysis that differences inappearance induced by illumination are largerthan differences between individuals. Sincedealing with illumination variation is a centraltopic in computer vision numerous approachesfor illumination invariant face recognition havebeen proposed. In [51], investigate the way, inwhich changes in illumination can affectperformances of some face recognition methods.They define three different classes in order to

grade the methods: the shape from shadingapproaches, which extract the shape informationof the face, from one or more of its views, therepresentation based methods, which try to get acharacterization of the face invariant toillumination changes and the generativemethods, which produce a wide set of syntheticimages containing as variations as possible. Theauthors deduced that none of the experimentedtechnique (edge map, 2D Gabor Filters, first andsecond derivatives of the gray level images) isable to solve the problem by itself and the resultsthey report seems to confirm this hypothesis.Notwithstanding this, several efforts have beenmade in order to achieve better performances inuncontrolled conditions. Indeed, [52] extendedthe edge map technique defining a newapproach, namely the Line Edge Map, in whichthe face contours are extracted and combined insegments, which are then organized in lines. TheHausdorff distance has also been modified inorder to manage these new feature vectors.Besides, they also describe a new prefilteringcriterion for screening the whole set of individualsbefore to perform the real testing operation. The

method has been tested on several conditions forpose and illumination and the results show thatthis approach outperforms other methods, suchas Linear Subspaces or Eigenfaces, presented in[16]However, the Fisherfaces remain superiorthanks to their capability to maximize thebetween-person variability, minimizing the within-

person differences. This suggests that combiningseveral linear methods, performances can befurther improved. Indeed, an in-depth study onthe performances of the linear methods whenchanges in illumination occur has beenconducted by [53]. The examined techniqueshave been compared with respect to bothrecognition rate and time/memory complexity.The authors observe that the LDA combined witha generalization of the SVD (Singular ValueDecomposition), outperforms all the othermethods. Nevertheless this hybrid is lessadaptable to general face recognition problems,owning to its computational cost. Therefore, theauthors suggest that to combine the LDA with the

QR decomposition could represent the optimalchoice in most cases, since it provides almostcomparable performances with the LDA/ SVDapproach with a lower cost. On the contrary, thePCA and the PCA + LDA (Fisherfaces) performworse of all the other methods. To overcomelimits introduced by the linearity of the abovementioned strategies, nonlinear methods, suchas the ICA, have been studied. One of the mostrecent work has been proposed by[54]. The faceis split in different regions that overlap. on theboundaries. For each class containing all theelements belonging to the same face region theresidual space (the space spanned by the PCAafter removing a few leading Eigenfaces) is

computed and the ICA is applied to. The resultsunderline that the PCA components in theresidual space are the same that in the normalspace, while ICA components are different, sothat performances improve. Moreover, to split theface in several regions simplifies the statisticalmodel of the illumination variations, making therecognition task more robust with respect tochanges. On the contrary, not much has beenmade yet on generative methods. One of the fewgenerative methods has been proposed by [55].The face shape and the albedo are extracted byfew subject images, by means of a shape fromshading algorithm.

The changes in poseIn many face recognition scenarios the pose ofthe probe and gallery images is different. Forexample, the gallery image might be a frontal‘‘mug-shot’’ and the probe image might be a 3/4view captured from a camera in the corner of aroom. Approaches addressing pose variation canbe classified into two main categories dependingon the type of gallery images they use. Multi-viewface recognition is a direct extension of frontal





face recognition in which the algorithms requiregallery images of every subject at every pose. Inface recognition across pose we are concernedwith the problem of building algorithms torecognize a face from a novel viewpoint, i.e. aviewpoint from which it has not previously beenseen. Linear subspaces have been extended in

order to deal also with the problem of posechanges. Indeed[56] present a framework forrecognizing faces with large 3D pose variations,by means of parametric linear subspace modelfor representing each known person in thegallery. The authors investigate two differentlinear models: (1) the LPCMAP model, that is aparametric linear subspace model, combining thelinear subspaces spanned by principalcomponents (PCs) of training samples and thelinear transfer matrices, which associateprojection coefficients of training samples ontothe subspaces and their corresponding 3D headangles; (2) the PPLS model, that extends theLPCMAP by using the piecewise linear approach,

that is a set of local linear models, each oneproviding continuous analysis and synthesismappings, enabling to generalize to unknownposes by interpolation. The experimental resultsshow that the recognition system is robustagainst large 3D head pose variations covering500 rotations along each axis. While significantlycompressing the data size, the PPLS systemperformed better than the LPCMAP system.However, the number of known people isrelatively small and the samples included someartificiality which might accidentally increase theperformance. Another drawback is that therecognition systems uses pixel-wise landmarklocations for representing facial shape and

deriving head pose information, but findinglandmark locations in static facial images witharbitrary head pose is an ill-posed problem.Then,[57] proposed to use the light-field toachieve a greater robustness and stability solvingthe problem of pose variation in face recognition.The light-field is a 5D function of position (3D)and orientation (2D), which specifies the radianceof light in free space. In particular, the authorsapply the PCA to a collection of light-fields offaces of different subjects, obtaining a set ofeigen light-fields, while the mean light-field couldalso be estimated and subtracted from all of thelight-fields. Since, any image of the objectcorresponds to a curve in the light-field

The occlusionOne of the main drawbacks of the appearance-based paradigm (e.g., PCA), is its failure torobustly recognize partially occluded objects.One way to deal with partially occluded objects(such as faces) is by using local approaches. Ingeneral, these techniques divide the face intodifferent parts and then use a voting space to findthe best match. However, a voting technique can

easily misclassify a test image because it doesnot take into account how good a local match is.In [58]in order to cope with this problem, eachface image is divided into k different local parts.Each of these k local parts is modeled by using aGaussian distribution (or, equivalently, with amixture of Gaussians), which accounts for the

localization error problem. Given that the meanfeature vector and the covariance matrix forevery local subspace are drawn out and theprobability of a given match can be directlyassociated with the sum of all k Mahalanobisdistances. This approach differs from previouslocal PCA methods in that it uses a probabilisticapproach rather than a voting space. In his workthe author investigates on the amount of theocclusion that can be handled by the proposedapproach, and the minimum number of localareas needed to successfully identify a partiallyoccluded face. Martinez demonstratedexperimentally that the suppression of 1/6 of theface does not decrease accuracy, while even for

those cases where 1/3 of the face is occluded,the identification results are very close to thoseobtained without occlusions. He also has shownthat worse results are obtained when the eyearea is occluded rather than the mouth area. Theprobabilistic approach proposed by Martinez isonly able to identify a partially occluded face,while [59] proposed a method that alsoreconstructs the occluded part of the face anddetects the occluded regions in the input image,by means of an auto-associative neural network.At first the network is trained on the non-occludedimages in normal conditions, while during thetesting the original face can be reconstructed byreplacing occluded regions with the recalled

pixels. The training data set consisted of ninetythree 18 × 25, 8-bits images, while the trainednetwork has been tested using three types of testdata: pixel-wise, rectangular, and sunglass. In theresults the authors claim that the classificationperformance is not decreased even if 20– 30% ofthe face images is occluded. On the other hand,this method suffers from two of the mainproblems of the NN based approaches: thesystem retraining in case of new enrolments andthe little availability of training samples.Moreover, a method, which is able to deal withboth occlusions and illumination changes, hasbeen proposed by [60].They presented acomplete scheme for face recognition based onsalient feature extraction

The ageMany of the considered techniques drop inperformances, when the time lapse between thetraining and testing images is not negligible. Thismakes clear that all the introduced methods donot take into account for problems due to the agevariations. Some strategies overcome thisproblem periodically upgrading the gallery or





retraining the system. Nevertheless this not verysuitable solution only applies to those systemsgranting services, which perform theauthentication, task frequently, while it isimpractical in other situations, such as lawenforcement. Alternatively the age of the subjectcould be simulated trying to make the system

more robust with respect to this kind of variation.Several techniques for the age simulation aregiven in literature: Coordinate Transformations,Facial Composites, Exaggeration of 3DDistinctive Characteristics, but none of thesemethods has been investigated in the facerecognition framework. In a recent work [61, 62]proposed a new method based on age functions.Every image in the face database is described bya set of parameters b, and for each subject thebest age function is drawn depending on his/herb. The greatest advantage of this approach isthat different subject-based age functions allowtaking into account for external factors whichcontribute towards the age variations. The

authors tested this approach on a database of 12people, with 80 images in the gallery and 85 inthe probe. They reported an improvement ofabout 4–8% and 12–15% swapping the probeand gallery set. In both the experiments the meanage of the subjects has been simulated, beforeperforming the recognition task. Notice that thenumber of the subject in the database is verysmall, emphasizing the absence of a standardFERET-like database, which systematicallymodels the age variations. However to improvethe robustness of the face recognition systemswith respect to changes in age is an interestingand still unexplored aspect in low enforcementapplications, mainly for the prediction of facial

appearance of wanted/missing persons.

There a more general way to state a techniquebetter than others!Methods presented in previous sections haveboth advantages and drawbacks. State whichone is the best is very difficult and stronglydepends on what is required the system to do.Moreover, most of these approaches have beentested on different datasets. One way to make amore general evaluation is to pick a set ofsignificant parameters, rather than consideringthe recognition rate only the parameter set mustincludes several aspects that need to be takeninto account when testing. Examples are numberand database characteristics, probe dimensionand gallery sets, input size and so on. It is quiteinteresting to analyze the way in which theseparameters can drive a more accuratecomparative study of face recognition algorithmsas shown in Table 2. Obviously, the greater thenumber of used databases is, the thorough theassessment of the performances can be. On thecontrary, the connection between the dimensionof the input and the effectiveness of the method

is less self-evident. In general, to speed uptraining/testing tasks, the higher thecomputational complexity is, the smaller thedimension of the input images can be. While it isclear that more information is carried by largerinput, some studies shows that recognition is stillpossible on 18 × 25 grayscale images [60].

However, high resolution images and videosmade possible by recent technologies andpresented in the upcoming FRVT2005 confirmthat the higher the resolution is, the betterperformances are. The probe and gallery set sizealso has to be taken into account mainly withrespect to the SSS problem. It is well known thatonly one image is available for training in mostreal situations, while the identification isperformed many times. This suggests that thesmaller the gallery set is, the higher the capabilityof extracting discriminant features is. This can befurther improved by a large probe set. It makessense then to minimize the ratio (gallerysize)/(probe size). Many research results show

that several approaches are more sensitive tochanges in high frequencies than to low ones[63]. This is not a desirable property, becauselow frequencies carry most of the invariantinformation about the identity of the subject, whilehigh frequencies are often affected by changes inenvironmental conditions. Therefore, theusefulness of a time lapse between sessionsproviding the images of the gallery and probe setbecomes apparent. Thus, the larger the numberof the addressed problems as shown in Table 3.is, the higher the adaptability to real-worldapplications can be esteemed. Finally, all themethods exposed so far require some kind ofinput preprocessing; and this could significantly

reduce the usefulness of a face recognitionalgorithm suggesting that the system flexibilityincreases when normalization on input data isreduced.Based on these considerations is then possible toinvestigate which techniques provide a betterapproximation of pinpointed parameters. ThePDBNN based algorithm seems to provide thebest experimentation. It addresses most of theproblems, while experiments conducted on threedifferent databases with a large number ofimages reported a high recognition rate. Asfurther example, the LEM approach can beconsidered. The recognition rate is lower thanother methods such as Th-Infrared [63] or [64],but it has been tested on more databases and itaddresses three different problems rather thanone.

ConclusionPCA gave average results in case of theillumination and pose variation databases. Incase of the dataset, where the training set wassmall and subject number large, PCA outrankedall the other algorithms. In terms of computational





[35] Gestel T. V., Suykens J. A. K., Lanckriet G.,Lambrechts A., De Moor B., andVanderwalle J. (2002) NeuralComputation ,15(5), 1115-1148.

[36] Cawley G. C. and Talbot N. L. C. (2003)Pattern Recognition , 36(11), 2585-2592.

[37] Lawrence N. D., and Schölkopf B. (2001)

Proceedings of the 18th InternationalConference on Machine Learning (pp.306-313).

[38] Yang M. H. (2002 ) Proceedings of the FifthIEEE International Conference onAutomatic Face and GestureRecognition (RGR’02), 215-220.

[39] Liu C. (2004) IEEE Transactions on PatternAnalysis and Machine Intelligence ,26(5),572-581.

[40] Yang J., Zhang D., Frangi A.F. and Yang J-Y. (2004) IEEE Transactions onPatternAnalysis and MachineIntelligence , 26(1), 131-137.

[41] Yang J., Yang J-Y. and Frangi A.F. (2003)

Image and Vision Computing , 21(12),1037-1044.

[42] Xu Y., Yang J.Y. and Yang J. (2004)Pattern Recognition , 37(6), 1299-1302.

[43] Yang J., Zhang D., Yang J-Y., Zhong J.,and Frangi A. F. (2005) IEEE Trans.Pattern Anal. Machine Intell . 27(2), 230-244.

[44] Cottrell G.W., Fleming M.K. (1990) Proc.Intell. Neural Network Conf., pp. 322– 325.

[45] Lin Shang-Hung, Kung Sun-Yuan, Lin Long-Ji (1997) IEEE Trans.Neural Networks 8(1), 114–132.

[46] Meng J.E., Shiqian W., Juwei L., Hock L.T.,

(2002) IEEE Trans. Neural Networks 13(3), 697–710.

[47] Lades M., Vorbruggen J.C., Buhmann J.,Lange J., van der Malsburg C., WurtzR.P., Konen W., (1993) IEEE Trans.Comput .42 (March), 300–311.

[48] Wiskott L., Fellous J.M., Kruger N., von derMalsburg C. (1997) IEEE Trans.PatternAnal. Machine Intell . 19 (July),775–779.

[49] Perronnin F., Dugelay J.-L. (2003) Proc.IMAGE’2003: Learning, Understanding,Information Retrieval, Medical, Cagliari,Italy, June.

[50] Garcia C., Zikos G., Tziritas G. (2000).Image Vision Comput . 18, 289–297.

[51] Adini Y., Moses Y., Ullman S. (1997 ) IEEETrans. Pattern Anal. Machine Intell .19 (7), 721–732, July.

[52] Gao Y., Leung M.K.H. (2002) IEEE Trans.Pattern Anal. Machine Intell . 24 (6),764–779.

[53] Li Q., Ye J., Kambhamettu C. (2004) Proc.

2004 IEEE Computer Society Conf. onComputer Vision and PatternRecognition (CVPR04).

[54] Kim Tae-Kyun, Kim Hyunwoo, HwangWonjun, Kee Seok-Cheol, Kittler J.,(2003) Proc. IEEE Computer SocietyConference on Computer Vision andPattern Recognition , Vol. 1 (June), pp.579–586.

[55] Georghiades A.S., Belhumeur P.N.,Kriegman D.J., (2001) IEEE Trans.Pattern Anal. Machine Intell . 23 (6),643– 660.

[56] Okada, K., von der Malsburg, C.,(2002)Fifth IEEE Internat. Conf. on Automatic

Face and Gesture Recognition, May, pp.64–69.

[57] Gross R., Matthews I., Baker S. (2002 ) In:Proc. Fifth IEEE Internat. Conf. onAutomatic Face and GestureRecognition , May, pp. 1–7.

[58] Martinez A.M. (2002) IEEE Trans. PatternAnal. Machine Intell. 24 (6), 748–763.

[59] Sahbi Hichem, Boujemaa Nozha (2002 ) In:Biometric Authentication, Internat.ECCV2002 Workshop Copenhagen ,June, pp. 121–132.

[60] Kurita T., Pic M., Takahashi T. (2003) IEEEConference on Advanced Video andSignal Based Surveillance , July, pp. 53–

58.[61] Lanitis A., Taylor J.C. (2000) In: 10th

Mediterranean ElectrotechnicalConference., MEleCon. Vol. 2, pp. 478– 481.

[62] Lanitis A., Taylor J.C., Timothy F.C. (2002)IEEE Trans. Pattern Anal.Machine Intell. 24 (4), 442–455.

[63] Sergent J. (1986) (H..D. Eills, M.A. Jeeves,F. Newcombe and A. Young, eds.),Dordrecht: Nijhoff.

[64] Buddharaju P., Pavlidis I., Kakadiaris I.,(2004) Computer Vision and PatternRecognition Workshop June, pp. 133– 138.





Table 1- Most important 2D Face Image Databases

Dataset Location Description

MITDatabase

ftp://whitechapel.media.mit.edu/pub/images/

Faces of 16 people, 27 of each under variousillumination conditions, scale and head orientation.

FERETDatabase

http://www.nist.gov/humanid/feret A large collection of male and female faces. Imageshave neutral and non neutral expressions.

UMISTDatabase

http://images.ee.umist.ac.uk/danny/database.html

564 images of 20 subjects. The users are capturedunder variety of poses.

Universityof BernDatabase

ftp://iamftp.unibe.ch/pub/Images/FaceImages/

300 frontal face images of 30 people and 150 profileface images.

YaleDatabase

http://cvc.yale.edu/projects/yalefaces /

Faces of 15users and total of 165 images. The lightingconditions are well controlled.

AT&T(Olivetti)Database

http://www.uk.research.att.com Contains 10 images for each of its 40 subjects, withvarious lighting conditions and facial expression.

AR faceDatabase

http://rvli.ecn.purdue.edu/~aleix/ 4000 color images corresponding to126 people’s faces(70 men and 56 women). Images feature frontal viewfaces with different facial expression, illuminationcondition, and occlusions

M2VTSDatabase http://Poseidon.csd.auth.gr/M2VTS/index.html A multimodal database containing various imagessequences

Table 2- List of face recognition methodsApproach Representative work

PCA 4,5,6,7,8,9,10,11,12,13

LDA 14,15,17,18

FLDA 16

CA 19,20,21,22

KPCA 24

KFD 25,26,27,28,29,30,31,32,33,34,35,36,37

KPCA 38,39,40,41,42,43

NN 44,45,46

GABOR FILTERS AND WAVELETS 47,48,49,50

Table 3- Problems addressed and workProblem addressed Work done

Changes in illumination 51,52,53,54,55

Changes in pose 56,57,58,59

2D face recognition survey.pdf

Documents

Transcript of 2D face recognition survey.pdf