Face Recognition Using Principle Component Analysis (PCA .

2y ago
31 Views
5 Downloads
411.39 KB
6 Pages
Last View : 14d ago
Last Download : 2m ago
Upload by : Abby Duckworth
Transcription

International Journal of Electrical & Computer Sciences IJECS-IJENS Vol:12 No:0550Face Recognition using Principle ComponentAnalysis (PCA) and Linear DiscriminantAnalysis (LDA)M.N.Shah Zainudin., Radi H.R., S.Muniroh Abdullah., Rosman Abd. Rahim., M.Muzafar Ismail.,M.Idzdihar Idris., H.A.Sulaiman., Jaafar A.Faculty of Electronic and Computer EngineeringUniversiti Teknikal Malaysia MelakaHang Tuah Jaya76100 Durian Tunggal, Melaka, Malaysia Abstract— Image Recognition is one of the computer visionapplications in recent years. Commercially, security and lawapplications require the use of face recognition technology.Human face can be regarded as the most obvious humanidentifier. Apparently the face is the most visible part of humananatomy and serves as the first distinguishing factor of a humanbeing. It helps a person to distinguish an individual from one toanother. Each individual has his own uniqueness and this couldbe one of the most transparent and unique feature of a humanbeing. Face recognition involves comparing an image with adatabase of stored faces in order to identify the individual in thatinput image. The images can be analyzed and faces can then beidentified, before they can be recognized. There are differentmethods of face recognition which involve a series of steps thatserve to capturing, analyzing and comparing a face to a databaseof stored images. This project covered comparative study ofimage recognition between Linear Discriminant analysis (LDA)and Principal Component Analysis (PCA). In this study, theresult of PCA and LDA will be analyzed in term of its accuracy,percentage of correct recognition, time execution and databaseused.Index Term— PCA, LDA, feature extraction.I. INTRODUCTIONNowadays, image recognition has become a popular topicamong the researchers because of its broad usage in manyapplications such as digital cameras, surveillance camera,image editing software, Facebook and many more. InFacebook, it implements facial recognition technology thatallows all users to semi-automating the photo-tagging process.In this comparative study, face recognition was chosenbecause it is the most significant human identifier. The face isthe most visible part of human anatomy and serves as the firstdistinguishing factor of a human being. It helps a person todistinguish an individual from one to another. Every individualhas his own uniqueness and thiscould be one of the most transparent and unique feature of ahuman being.Face recognition involves comparing an image with a databaseof stored faces in order to identify the individual of that inputimage. The image will first be analyzed and faces can then beidentified, before it can be recognized. While this process maybe a trivial task for the human brain, it has proved to beextremely difficult for the artificial technology to imitate. It iscommonly used in applications such as humanmachine interfaces and automatic access control systems.The need of face recognition application in many areas suchas in Law Enforcement help the government to stay one stepahead of the world's ever-advancing terrorists, in AirportSecurity where it is use to enhance security efforts that alreadyunderway at most airports and other major transportation hubs(seaports, train stations, etc.), in Access Control to enhancesecurity efforts considerably, in Driver's Licenses & Passportsthat can leverage the existing identification infrastructure,Homeland Defense, Customs & Immigration and SceneAnalysis. All of this makes face recognition become moreimportant nowadays. Other potential applications includeATM and cash-checking security. The software is able toquickly verify a customer's face. After a customer‟s consent,the ATM or check-cashing kiosk captures a digital image ofhim.II. LITERATURE REVIEWA. Face recognitionA facial recognition system is a computer application toautomatically identifying a person from a digital image or avideo frame. One way to achieve this is by comparing selectedfacial features from the image to a facial database [2]. It istypically used in security systems and can be compared toother biometrics such as fingerprint or human iris [1].Currently, developers came up with the design that is capableof extracting and picking up faces from the crowd and have itcompared to an image source - database. The software has theability to know how the basic human face looks like in order1214404-3737-IJECS-IJENS October 2012 IJENSIJENS

International Journal of Electrical & Computer Sciences IJECS-IJENS Vol:12 No:05for it to work accordingly. Thus, developers designed theseprograms (by storing commands) to pinpoint a face andmeasure its features.There are different methods of facial recognition whichinvolve a series of steps that serve to capturing, analyzing andcomparing a face to a database of stored images. Some relatedsoftware was designed to recognize similarities through patternrecognition. Pattern recognition is often used under the namesof diagnosis and clarifications. Each of this software varies onhow it is designed to work yet the function and concept is stillthe same that is identifying on facial landmarks. Because ofthese, facial recognition is hard to fool since it comparesspecific proportions and angles of the defined facial features.Facial recognition software falls into a larger group oftechnologies known as biometrics. Biometrics uses biologicalinformation to verify identity. The basic idea behindbiometrics is that our body contains unique properties that canbe used to distinguish us from other persons.Face recognition has a number of advantages over otherbiometrics. Firstly, it is non-intrusive. While many biometricsrequire the subject‟s co-operation and awareness in order toperform identification, such as looking into an eye scanner orplacing their hand on a fingerprint reader, face recognitioncould be performed even without the subject's knowledge.Secondly, the biometric data used to perform recognition is ina format that is readable and understood by humans. Thismeans that a potential face recognition system can always bebacked up and verified by a human. For example, supposing aperson was falsely denied access to a site by a face recognitionsystem. That decision could easily be corrected by a securityguard that would compare the subject's face with the storedimage, whereas this would not be possible with otherbiometrics such as iris. Other advantages are that there is noassociation with crime as with fingerprints (few people wouldobject to looking at a camera) and many existing systemsalready store face images (such as police mug shots).B. Principle Component AnalysisPrincipal component analysis (PCA) method used for globalfeature extraction is a powerful technique for extracting globalstructures from high-dimensional data set and has been widelyused to reduce dimensionality and extract abstract features offaces for face recognition (Turk and Pentland, 1991; Zhao etal., 2000). It can also be used to identify patterns in data, andexpressing the data in such a way as to highlight theirsimilarities and differences. This provides an effectivetechnique for dimensionality reduction.The PCA method has been extensively applied for the taskof face recognition. Approximate reconstruction of faces in theensemble was performed using a weighted combination ofeigenvectors (Eigen pictures), obtained from that ensemble(Sirovich and Kirby, 1987). The weights that characterize theexpansion of the given image in terms of Eigen pictures areseen as global facial features. In an extension of that work,Kirby and Sirovich (1990) included the inherent symmetry of51faces in the Eigen pictures. All images in face image indatabase are representing in matrix as a very long vector.There are five steps involved in the system developed byTurk and Pentland. First, the system needs to be initialized byfeeding it a set of training images of faces. This is used theseto define the face space which is set of images that are facelike. Next, when a face is encountered it calculates aneigenface for it. By comparing it with known faces and usingsome statistical analysis it can be determined whether theimage presented is a face or not a face at all. Then, if an imageis determined to be a face the system will determine whether itknows the identity of the face or not. The optional final step isthat if an unknown face is seen repeatedly, the system canlearn to recognize it.The eigenface technique is simple, efficient, and yieldsgenerally good results in controlled circumstances [1]. Thesystem was even tested to track faces on film. There are alsosome limitations of eigenfaces. There is limited robustness tochanges in lighting, angle, and distance [6]. 2D recognitionsystems do not capture the actual size of the face, which is afundamental problem [4]. These limits affected the technique‟sapplication in security cameras because frontal shots andconsistent lighting cannot be relied upon.C. Linear Discriminant AnalysisLinear Discriminant is a “classical” technique in patternrecognition [4], where it is used to find a linear combination offeatures which characterize or separate two or more classes ofobjects or events. The resulting combination may be used as alinear classifier or, more commonly, for dimensionalityreduction before it can be classified.In computerized face recognition, each face is representedby a large number of pixel values. Linear discriminant analysisis primarily used here to reduce the number of features to amore manageable number before classification. Each of thenew dimensions is a linear combination of pixel values, whichform a template. The linear combinations obtained usingFisher's linear discriminant are called Fisher faces, while thoseobtained using the related principal component analysis arecalled eigenfaces.Linear Discriminant Analysis easily handles the case wherethe within-class frequencies are unequal and their performancehas been examined on randomly generated test data. Thismethod maximizes the ratio between-class variance to thewithin-class variance in any particular data set therebyguaranteeing maximal separability. The prime differencebetween LDA and PCA is that PCA does more of featureclassification and LDA does data classification. In PCA, theshape and location of the original data sets changes whentransformed to a different space whereas LDA doesn‟t changethe location but only tries to provide more class separabilityand draw a decision region between the given classes [5].Data sets can be transformed and test vectors can beclassified in the transformed space by two differentapproaches.1214404-3737-IJECS-IJENS October 2012 IJENSIJENS

International Journal of Electrical & Computer Sciences IJECS-IJENS Vol:12 No:05(i) Class-dependent transformation: This type of approachinvolves maximizing the ratio of between class variance towithin class variance. The main objective is to maximize thisratio so that adequate class separability is obtained. The classspecific type approach involves using two optimizing criteriafor transforming the data sets independently.(ii) Class-independent transformation: This approachinvolves maximizing the ratio of overall variance to withinclass variance. This approach uses only one optimizingcriterion to transform the data sets and hence all data pointsirrespective of their class identity are transformed using thistransform. In this type of LDA, each class is considered as aseparate class against all other classes.The goal of the Linear Discriminant Analysis (LDA) is tofind an efficient way to represent the face vector space. PCAconstructs the face space using the whole face training data asa whole, and not using the face class information. On the otherhand, LDA uses class specific information which bestdiscriminates among classes. LDA produces an optimal lineardiscriminant function which maps the input into theclassification space in which the class identification of thissample is decided based on some metric such as Euclideandistance. LDA takes into account the different variables of anobject and works out which group the object most likelybelongs to [3].III.52Fig. 1. One of ATT face database with ten different expressions.Fig. 2. Images corresponding to one individual.B. ATT face databaseRESULTA. DatabaseThere are two types of database that has been used which isATT Face Database [7] and Indian Face Database (IFD) [8].The used of these two databases is for performancecomparison. The different between these two databases is theangle of which the image was captured where ATT databasehas a very small change of angle for every dataset but IFD hasa significance change of angle for every image. Thedimensionality of ATT database is 92x112 pixels while IFD is64 x48 pixels.ATT face database contains ten different images for 40distinct subjects. For some subjects, the images were taken atdifferent times, varying the lighting, facial expressions (open /closed eyes, smiling / not smiling) and facial details (glasses /no glasses). All the images were taken against a darkhomogeneous background with the subjects in an upright,frontal position (with tolerance for some side movement).IFD contains images of 40 distinct subjects with elevendifferent poses for each individual. All of the images have abright homogeneous background and the subjects are in anupright, frontal position. For each individual, the followingpose for the face is included: looking front, looking left,looking right, looking up, looking up towards left, looking uptowards right, looking down. In addition to the variation inpose, images with four emotions - neutral, smile, laughter,sad/disgust - are also included for every individual. Thesetwo databases provide a comprehensive dataset for testing theperformance of the algorithms chosen.Fig. 3. Accuracy PCA and LDA using ATT databaseBased on the fig. 3, we can see LDA is more accurate thanPCA. LDA accuracy is 93% while PCA is 89.5%. Thisaccuracy is based on separating 50% of the images in thedatabase to be train image and the remaining 50% to be testimage.Fig. 4. Training time taken in millisecond for each image(ATT)1214404-3737-IJECS-IJENS October 2012 IJENSIJENS

International Journal of Electrical & Computer Sciences IJECS-IJENS Vol:12 No:0553Fig. 5. Testing time taken in millisecond for each image(ATT)Fig. 8. Training time taken LDA and PCA (IFD)Fig. 6. Total execution time taken in millisecond for each image(ATT)Based on the result above, LDA is proved to be better thanPCA. The accuracy for LDA is more than PCA. Moreover,LDA is faster than PCA in testing and training the image andof course LDA beat PCA for total execution of time taken.Fig. 9. Testing time taken LDA and PCA (IFD)C. Indian Face Database (IFD)Fig. 10. Total execution time taken LDA and PCA (IFD)Fig. 7. Accuracy LDA and PCA (IFD)By using IFD, the accuracy for LDA is more than PCA.But the time taken for execution, train and test the image ofLDA is more than PCA for IFD dataset.1214404-3737-IJECS-IJENS October 2012 IJENSIJENS

International Journal of Electrical & Computer Sciences IJECS-IJENS Vol:12 No:0554recognition process is simplified with a larger trained weightset.V. CONCLUSIONFig. 11. Accuracy vs. %train datasetThe figure shows that overall result for two dataset that hadbeen used which has been testing for each percentage form30% to 70% on PCA and LDA technique. Based on the result,it clearly shows that LDA is a more accurate technique and itwill be even more accurate when it being tested using ATTdatabase.In this project two type of feature extraction for facerecognition algorithms, PCA and LDA were studied. The PCAand LDA were implemented using MATLAB and theperformance was determined in terms of the recognitionaccuracy and the execution time taken. Experiments wereperformed under different conditions; by varying the input faceimage dataset and also by varying the parameters of theindividual algorithms. The study showed that the LDAperforms better than PCA in terms of the accuracy ofrecognition. For the future enhancement, method of findingand identifying faces accurately can maximize control ofsubject‟s pose and maximize control of environment. Bycontrolling a person‟s facial expression, as well as his distancefrom the camera, camera angle, the scene‟s lighting, a posedimage can minimize the number of variables in photograph.This control allows the facial recognition software to operateunder near ideal conditions greatly enhancing its accuracy.[1]IV. DISCUSSION[2]The performance of the two algorithms for the twodatabases has been compared considerably which showed adistinctive performance. It is observed that the recognition rateis higher in the ATT database than the IFD. This observationis due to the nature of images encompassed in the IFD. TheIFD database has images where each subject is portrayed withhighly varying orientation angles. Also, the IFD has imageswith a larger background region that the ATT images.In terms of accuracy the LDA shows a higher recognitionrate. This is because of the use of discrete classes to group theimages and perform a covariance minimization within thesame class. The use of this distinct class information increasesthen feature space used for classification.In term of training set used, it's not always when the trainingset is small, PCA can outperform LDA, it's also depend on thedatabase which is type of image had been used. It is also thesame when the number of samples is large and representativefor each class, LDA outperforms PCA, because based from theresult the LDA is always outperform than PCA even thetraining set is small.Result shows a classic shortcoming of the recognitionfunction which is seen to be heavily dependent upon thenumber of images the algorithm is trained upon. Although thisproject considers only closed loop recognition where theimages to be tested are also from the same database the choiceof a higher number of faces for training can lead to a higheraccuracy.However, the improved performance of the class basedLDA is evident from the results. Another noticeable propertyis that the execution time varies inversely with the percentageof images the algorithms are trained upon. This is because 6][17]REFERENCES[Online]. pplications.html[Online]. Available:http://en.wikipedia.org/wiki/Facial recognition systemB. J. Oh, Face Recognition using Radial Basis Function Networkbased on LDA. World Academy of Science, Engineering andTechnology 7 2005R. Duda and P. Hart, Pattern Classification and Scene Analysis. NewYork: Wiley, 1973.S. Balakrishnama, A. Ganapathiraju,Linear Discriminant Analysis.K. Etemad and R. Chellappa, “Discriminant Analysis for Recognitionof Human Face Images”, Journal of the Optical Society of America,Vol 14, pp 1724-1733, 1997AT&T Laboratories archive:pub/data/att faces.tar.V. Jain and A. Mukherjee, The Indian Face Database. [Online].Available: http://vis-www.cs.umass.edu/ vidit/IndianFaceDatabase/ ,2002P. Rauss, P. J. Phillips, M. Hamilton, and A. T. DePersia, „„FERET(Face Recognition Technology) Program,‟‟ in 25th AIPR Workshop:Emerging Applications of Computer Vision, D. Schaefer and W.Williams, eds., Proc. SPIE 2962, 253–263 (1996).M. A. Turk and A. P. Pentland, „„Face recognition using Eigenfaces,‟‟in Proceedings of the IEEE Computer Society Conference on ComputerVision and Pattern Recognition (IEEE Computer Society, Los Alamitos,Calif., 1991), pp.586–591.L. Sirovich and M. Kirby, „„Low-dimensional procedure for thecharacterization of the human face,‟‟ J. Opt. Soc. Am. A4, 519–524(1987).X. Lu, Image Analysis for Face Recognition, May 2003A. M. Martinez, A. C. Kak, “PCA versus LDA,” IEEE Trans. on PatternAnalysis and Machine Intelligence, vol. 23, pp. 228-233, 200Y. Atilgan “Face Recognition” Project report, Cankaya University,Turk

B. Principle Component Analysis Principal component analysis (PCA) method used for global feature extraction is a powerful technique for extracting global structures from high-dimensional data set and has been widely used to reduce dimensionality and extract abstract features of faces for face recognition (Turk and Pentland, 1991; Zhao et

Related Documents:

Keywords: Independent component analysis, ICA, principal component analysis, PCA, face recognition. 1. INTRODUCTION Several advances in face recognition such as "H lons, " "Eigenfa es, " and "Local Feature Analysis4" have employed forms of principal component analysis, which addresses only second-order moments of the input. Principal component

Subspace methods have been applied successfully in numerous visual recognition tasks such as face localization, face recognition, 3D object recognition, andtracking. In particular, Principal Component Analysis (PCA) [20] [13] ,andFisher Linear Dis criminant (FLD) methods [6] have been applied to face recognition with impressive results.

2.1 Face Recognition Face recognition has been an active research topic since the 1970’s [Kan73]. Given an input image with multiple faces, face recognition systems typically first run face detection to isolate the faces. Each face is pre

The goal of this lab is to implement face recognition using Principal Component Analysis (PCA). One of the most important applications of the PCA is mapping a dataset into a new space with smaller dimension in . 3 Face Recognition In face recognition, we have access to a dataset called the training set

18-794 Pattern Recognition Theory! Speech recognition! Optical character recognition (OCR)! Fingerprint recognition! Face recognition! Automatic target recognition! Biomedical image analysis Objective: To provide the background and techniques needed for pattern classification For advanced UG and starting graduate students Example Applications:

the face for biometric recognition. Our previous experiments with ear and face recognition, using the standard principal component analysis approach, showed lower recognition performance using ear images. We report results of similar experiments on larger data sets that are more rigorously controlled for relative quality of face and ear images.

Principal component analysis (PCA) is a popular unsuper-vised statistical method to find useful image representations. . Face recognition performance was tested using the FERET database [52]. Face recognition performances using the ICA representations were benchmarked by comparing them to per-

2 Annual Book of ASTM Standards, Vol 06.03. 3 Annual Book of ASTM Standards, Vol 14.03. 4 Reagent Chemicals, American Chemical Society Specifications, American Chemical Society, Washington, DC. For suggestions on the testing of reagents not listed by the American Chemical Society, see Analar Standards for Laboratory Chemicals, BDH Ltd., Poole, Dorset, U.K., and the United States Pharmacopeia .