Aframeworkfor Face Classification under Pose Variations Alwin D. Anuse Jagdish P. Sarode Dept. of E & TC Engineering Maharashtra Institute of Technology Pune, India jagdish.sarode@mitpune.edu.in Dept. of E & TC Engineering Maharashtra Institute of Technology Pune, India alwin. anuse@mitpune. edu.in Ab stract- Automatically verifying a person from a video frame or a digital image using computer application is known as a Face Recognition system. With changes in facial pose, face appearance changes drastically . Recognition of faces under pose variations is much more challenging. Now a day's recognizing human faces in unconstrained or wild environment is emerging as a critically important and technically challenging computer vision problem. gradually Recently shifting unconstrained its setting. Database called as face focus A "My new recognition on much community more unconstrained is challenging human unconstrained Database" has face been developed in this paper. A model based approach is used and the Moment based feature extraction techniques (Hu's, Zernike and Legendre Moments) are implemented on three different face databases containing different poses of the faces. This paper proposes a modified method called "Genetic Algorithm based Transfer Vectors" for generation of features of a frontal face from the features of different poses of image. Next, the generated called "Genetic Algorithm based Transfer vectors" for generation of features of a frontal face from the features of a different poses of image. Then the generated frontal features are matched with the actual frontal features [1]. Also this paper introduces new unconstrained human face Database called as "My unconstrained Database" which is inspired from IMFDB [7]. For feature extraction Hu, Zernike and Legendre moments are used [2] [6]. Feature vectors are obtained for the images in frontal and a non-frontal poses. This paper is organized as follows, Section II mentions flowchart and overview of the Face recognition System, section III discuss about three different Face Databases used in this paper, section IV gives details about feature extraction methods and their algorithms, section V discuss about proposed Genetic Algorithm based Transfer Vectors method, section VI and VII gives results and conclusion respectively. frontal features are matched with the actual frontal features. II. Extracted feature are classified by three different methods: kNN classifier, LDA and Transform Vector with Distance Metric and finally Correct Recognition Rate is determined. Index Terms- Face recognition, Pose Variation, Hu's Moments, Zernike Moments, Legendre Moments, kNN classifier, LDA, Genetic Algorithm, Transfer vectors, unconstrained face database. I. INTRODUCTION Automatically verifying or identifying a person from a video frame or a digital image using computer application is known as A Face Recognition system. Face Recognition provides a more direct, friendly and convenient identification method and it is more acceptable to users as compared to other biometric methods. Because of variation in face pose angle, illumination, expression and occlusion there are many challenges in face recognition. Face appearance changes drastically with change in facial pose and hence recognition of faces under pose variations has proved to be a difficult problem. A pose tolerance technique is needed because most of the time, for matching purpose of each test face under recognition, there exists only a one corresponding image of frontal face in data base. Hence, generation of either the frontal face image or the frontal face features from the non-frontal face image is necessary. To generate front face features there is a need to obtain such a transfonn through learning and finally transfonnation is used. This paper proposes a modified method 978-1-4799-3080-7114/$31.00 ©2014 IEEE FACE RECOGNITION SYSTEM FLOW CHART Flow chart of face recognition system is given in figure 1. In the preprocessing stage RGB to grayscale conversion, image resizing, histogram equalization to improve contrast of the image and padding operations are done. For feature extraction Hu moments, 4th order Zernike moments and 4th order Legendre moments are used. Hu moments feature vector contents seven moments, 4th order Zernike moments gives feature vector of nine moments and 4th order Legendre moments gives a vector of fifteen moments. kNN classifier with k 1, Linear Discrimenant Analysis are used to classify the face images. Also proposed Genetic Algorithm based Transfer Vectors method is used to transfer non frontal face moments to frontal face moments and classification is done using minimum distance method. Performance of the system is calculated on the basis of, out of the total given faces how many faces are correctly recognized and its percentage is given in the result tables. = III. A. FACES DATABASES The FERET Database, USA The FERET Database was created between 1993 to 1996 [3] [4]. The Database contains 196 subjects that include 100 males and 96 females. The images were collected in a semi­ controlled environment. The size of each image is 256 X 386 pixels and each pixel is of 256 grey levels. The files are in jpg 1886 Authorized licensed use limited to: MIT-World Peace University. Downloaded on May 30,2022 at 06:09:52 UTC from IEEE Xplore. Restrictions apply. much more challenging unconstrained settings. My unconstrained Database (MUDB) consist 270 faces of eighteen well-known personalities in India collected from web. Fifteen different poses of each person are collected. Manual selection of faces from web resulted in high degree of variability (in pose, scale, illumination, expression, age, image size etc). In the laboratory setting it difficult to capture variety of natural expressions. There are six female subjects and twelve male subjects in this Database. All images are in . jpg format. This can help in designing face recognition algorithms that can robustly recognize an individual. Image Acauisition (RGB to GreyscaJe, Image Resize, Histogram equalization, Padding) I.mage Preprocessing Summary of Databases used in this paper is given in table 1. Fig. 2 Subject of My unconstrained Database (MUDB) Performance Analysis Fig. 3 Different poses of a particular Subject TABLE I: SUMMARY OF DATABASE USED FOR TESTING Fig.1 Face recognition System flow chart format. Each subject is having following nine different poses of face rotation 0°, + 60°, +40°, + 2So, +I So, I SO, 2So, 4 0°, 60° _ _ B. - . Indian Face Database (IIT Kanpur) This Database is created in campus of Indian Institute of Technology, Kanpur [S]. It Database is of S7 distinct subjects having eleven different poses for each person. Images of the subjects in database are in an upright, frontal position and captured on bright homogeneous background. For each individual subject, following looking poses are captured: front, left, right, up, up towards left, up towards right, down, laughter, neutral, sad/disgust, smile. The files are in . jpg format. C. FERET Database Indian Face Database (lIT Kanpur) My Unconstrained Database (MUDB) Total Persons 196 57 18 Poses/Person 9 II 15 Poses used for Training 7 8 12 Poses used for Testing 2 3 3 - My unconstrained Database (MUDB) This paper introduces new unconstrained human face Database called as "My unconstrained Database" which is inspired from IMFDB [7]. Recognizing human faces in the wild environment is emerging as critically important and technically challenging computer vision problem. With the introduction of Databases such as LFW, Pubfigs and IMFDB face recognition community is gradually shifting its focus on IV. FEATURE EXTRACTION Feature extraction is a dimensionality reduction technique. In this paper Hu's Moments, Zemike Moments and Legendre Moments are used for feature extraction. A. Hu 's Moments: Hu, in 1962 introduced 'Hu's moments'. These moments are translation, scale and rotation invariants. Hu's moments are a set of seven moments. These seven moments are nonlinear combinations of normalized central moments up to order three. 20I4International Conference on Advances in Computing, Communications and Informatics (ICACCl) Authorized licensed use limited to: MIT-World Peace University. Downloaded on May 30,2022 at 06:09:52 UTC from IEEE Xplore. Restrictions apply. 1887 A 19orithm for Feature Extraction by using Hu 's Moments: I. 2. 3. 4. Hu's moments of subjects of FERET database for frontal pose are given in table 2. Start Acquire the image Convert colored image to grayscale image Calculate Geometrical Central Moments B. . . the moment of order (P+Q),p,q = 0, 1,2, . . . Where " t,.q IS and MxN is size of image 5. Calculate Normalized Central Moments l flpq = E!'=l L�l (X - X),P (y- y),C4f(x, y) (2) Here x and j are the image centroid and are defined by: x = tll:t .o tllj).O Y= - and -' 6. ij - Mz M3 M4 M5 = = = = = 'Ill (I !ti) �ltl() + 2 (4) 1120 + 110z 2 (1120 -1]02) + 411il (lh o - 31112)Z + (31h1 - lJ03)2 2 (1130 -1]1 ) + (1121 - 1103)Z 2 z (1]30 - 31112)(1130 + 1112 )[(1130 + 1hz) 2 3(1121 + 1103) ] + (31]z1. - 1103)(11z1. + 1103)[3(1130 1]12)Z - (11z1 + 1103)2] M6 M7 7. 8. = 2 (1120 -1]02)[(1130 + 11'12) - (1lz1. + 1103)2 ] + 4711.1(1130 + 1112) (lJ21 + 1103) = 2 (31121 - 1103)(1130 + 111 ) [(1]30 + 1]I ) 2 Z 3(1121 + lJ03)2] - (1130 - 311'12)(1121 + lJ03)[3(11 30 + z 1112) - (11z1 + 1103)2] (5) Combine these Seven Invariants to form the feature vector of the image Stop 1888 I. Start 2. Define the order and repetition of the moment. 3. Acquire the image. 4. Convert colored image to grayscale image. 5. Resize image to 120 x 120 pixels 6. Converts the intensity image I to double precision. 7. Calculate the Radial Polynomial using "(6)" (3) Calculate Seven Moments Invariants using Normalized Central Moments given by Equation (5). Ml Algorithm for Feature Extraction by using Zernike Moments: m� l HiO , o Above central moments are origin independent and hence they are translation invariants. For scale invariance divide corresponding central moment by ooth moment i. e. scaled energy of the original moment as given in Equation (4) 11 Zernike Moments (ZM): Prof. Frits Zemike introduced a set of complex polynomials called the Zemike polynomials. These polynomials are defined on the interior of the unit disc, i. e. , x1\2+ y1\2=1 and form a complete orthogonal basis set ZM is geometrical orthogonal moments. Orthogonal moments are rotation as well as scale invariants and very robust in the presence of noise. (6) 8. Calculate Zernike Moments using "(7)" Ztun = un 1 E!:-l �=-lf(x, y) V* nm (X, y) (7) 9. Combine these invariants to form feature vector of the image. 10. Stop. Zemike moments of subjects of FERET database for frontal pose are given in table 3. C. Legendre Moments: These moments are based on the Legendre polynomials [7]. These are orthogonal Moments. The two dimensional Legendre Moments of order (m + n), are defined as "(8)" . L_ - pq = ( I ) I a. . f a. p+1) q+1) 4 1 -] ,. .... (1d. I"''I' )f(:A!, y) d� d}, (8) The image is scaled in the region [-I, 1]. Hence Xi and Yj can be normalized as "(9)" 2j 2; x· = 1 , Yj eN- I ) 1 (M-1) (9) and - --- = - In this paper up to 4th order Legendre moments are calculated. Legendre moments of subjects of FERET database for frontal pose are given in table 4. 2014International Conference on Advances in Computing, Communications and Informatics (ICACCl) Authorized licensed use limited to: MIT-World Peace University. Downloaded on May 30,2022 at 06:09:52 UTC from IEEE Xplore. Restrictions apply. TABLE 2 HU MOMENTS OF FERET DATABASE FOR FRONTAL POSE Moments of the subject Subject of Database M, M, M3 M. Ms M. M7 Subject 1 6.37421 14.43838 24.17941 23.12982 46.96479 30.61291 47.48578 Subject 2 6.40062 14.64574 23.12028 22.63719 45.57875 29.98992 46.68987 Subject 3 6.30489 14.52329 24.89738 24.91324 49.99588 32.24101 50.78289 Subject 4 6.33768 14.49911 23.90741 25.5165 51.49041 3 3.05424 50.36829 Subject 5 6.41553 14.69512 23.95389 23.99652 48.06056 31.39594 48.98018 TABLE 3 ZERNIKE MOMENTS OF FERET DATABASE FOR FRONTAL POSE Moments of the subject Subject of Database M, M, M3 M. Ms M. M7 Ms M. Subject 1 0.6005 0.0906 0.4134 0.0604 0.1096 0.0259 0.9430 0.0611 0.0385 Subject 2 0.5869 0.0963 0.4362 0.0245 0.0366 0.0231 0.9935 0.0349 0.0107 Subject 3 0.5150 0.0284 0.2833 0.0179 0.0762 0.0203 0.7370 0.0195 0.0429 Subject 4 0.6030 0.0229 0.3406 0.0338 0.0652 0.0349 0.8815 0.0084 0.0433 Subject 5 0.6097 0.0413 0.4293 0.0095 0.0433 0.0025 0.9600 0.0279 0.0116 TABLE 4 LEGENDRE MOMENTS OF FERET DATABASE FOR FRONTAL POSE Moments of the subject Subject of Database V. M, M, M3 M. Ms M. M7 Ms M. M,. Mll M12 Subject 1 0.1170 -0.0230 0.0234 0.0002 0.0187 0.0544 0.0094 -0.0305 0.0222 -0.0146 0.0175 -0.0104 Subject 2 0.1145 -0.0131 0.0385 -0.0065 0.0177 0.0371 0.0059 -0.0136 0.0046 -0.0134 0.0141 -0.0015 Subject 3 0.1006 -0.0057 -0.0049 -0.0025 0.0115 0.0114 0.0053 -0.0446 0.0000 -0.0283 0.0340 0.0047 Subject 4 0.1175 -0.0085 0.0211 -0.0122 0.0448 0.0610 0.0141 0.0417 -0.0055 0.0335 -0.0196 0.0210 Subject 5 0.1188 -0.0056 0.0260 0.0026 0.0269 0.0383 0.0109 0.0076 0.0062 0.0049 0.0037 0.0020 GENETIC ALGORITHM BASED TRANSFORM VECTORS This paper proposes new method to transform feature vector of non frontal face pose to the feature vector of frontal face pose called "Genetic Algorithm based Transform Vectors" [J]. Let, matrix Ap contains feature vectors of all non frontal poses of face and matrix Af contains frontal face feature vectors. Let U be the desired optimized transform multiplier to be learned. U is calculated such that erl"(H' = Af - Ap' * U will be minimum. The optimized transformation multiplier "Uop" is calculated iteratively by using Genetic Algorithm. Algorithm for the fitting function of Genetic Algorithm used is given below. 1. 2. Read frontal face vector to variable af Read feature vectors of all non frontal poses of face to variable ap(n) respectively where n J, 2, . . . . , No. of different poses of face. Find err(n) vector by using Equation err(n) (af(n) - (ap(n) * U»; At the first iteration store err_old any high value in Excel sheet. = Thus, Al Therefore = (10) , U - Ap U = ((ApT _ Ap) - 1 (ApT An ) 3. (lJ) 4. = = 2014International Conference on Advances in Computing, Communications and Informatics (ICACCl) Authorized licensed use limited to: MIT-World Peace University. Downloaded on May 30,2022 at 06:09:52 UTC from IEEE Xplore. Restrictions apply. 1889 5. 6. 7. 8. Find err_new by taking summation of all the elements of err(n) matrix. Read err old from excel sheet. If err_new < err_old, then err_new err_old, else keep err_old as it is. Send err_new to genetic algorithm. = After obtaining optimized U i. e. Uop, it is used to transform the vectors of non frontal pose of face to the corresponding vectors of frontal pose of the face. These transformed victors are used for feature matching. Euclidean distance is used for feature matching. (12) As given in "12", Euclidean distance between two vectors is used to give the matching score. TABLE 7 (%) RECOGNITION RATE OF MY UNCONSTRAINED DATABASE (MUDB) Database Classifier kNN Classifier My Database Discriminant Analysis Transfonn Vector with Distance Metric Moments (%) Recognition Rate Hu 7.97 Zemike 16.72 Legendre 23.05 Hu 9.17 Zemike 25.32 Legendre 25.76 Hu 6.55 Zemike 10.92 Legendre 15.05 Graphical representation of Results are shown in figure 4. A minimum value of the Score along with its corresponding frontal pose vector is stored. The frontal posed vector which gives minimum Score value is the closer match of two vectors and are said to be matched or of same class. VI. TABLE 5 (%) RECOGNITION RATE OF FERET DATABASE Database Classifier kNN Classifier FERET Discriminant Analysis Transfonn Vector with Distance Metric TABLE 6 Database (%) Recognition Rate Hu 22.45 Zemike 90.45 Legendre 91.66 Hu 52.47 Zemike 85.38 Legendre 91.81 Hu 12.58 Zemike 62.98 Legendre 38.03 Classifier Moments (%) Recognition Rate kNN Hu 21.21 Zemike 71.80 Legendre 72.28 Discriminant Analysis Transfonn Vector with Distance Metric 1890 Moments (%) RECOGNITION RATE OF lIT K DATABASE Classifier IIT K RESULTS Hu 36.36 Zemike 71.80 Legendre 80.69 Hu 12.25 Zemike 54.85 Legendre 49.99 Fig. 4 Graphical representation of results in Table 5, 6 and 7. VII. CONCLUSION This paper proposes a new method called "Genetic Algorithm based Transfer Vectors" for generation of features of a frontal face from the features of a different poses of image. Also this paper introduces new unconstrained human face database called as "My unconstrained Database" which is inspired from IMFDB. In this paper face recognition using moments has been presented. The proposed method along with two other standard methods is implemented by authors and tested on all the three different databases mentioned in this paper. Nine experiments were conducted. Features were extracted using three different moments (Hu moments, Zemike moments and Legendre moments). Two different standard classifiers (kNN Classifier, Linear Descriminant Classifier) were used for classification. "Genetic Algorithm based Transform Vectors" method is used to transform non frontal vectors to frontal vectors and matching of two vectors is done using Euclidean distance method. From the experimental results it is observed that, the Legendre moments are the best among three moments and it gives 91. 81% recognition rate with LDA classifier. Results of Zemike Moments give maximum 90.45% recognition rate. Zemike Moments are good and can be used for face recognition. Hu moment gives maximum 52.45% recognition 2014 International Conference on Advances in Computing, Communications and Informatics (ICACCl) Authorized licensed use limited to: MIT-World Peace University. Downloaded on May 30,2022 at 06:09:52 UTC from IEEE Xplore. Restrictions apply. rate. The perfonnances of Hu's moments are not satisfactory and can't be used for face recognition. Proposed "Genetic Algorithm based Transfer Vector" method perfonnance is lower than the other two standard methods and need to be improved further. Newly introduced "My unconstrained Database" is useful to further boost the face recognition systems under unconstrained pose variations as all three methods gives lower recognition rate as compared to other two face Databases. VIII. ACKNOWLEDGMENT Jagdish P. Sarode thanks the Management, the Director of MAEER's M.I.T. Pune for sponsorship of Master of Engineering (Electronics - Digital Systems) course and Prof. (Dr.) G. N. Mulay, for financial help. IX. REFERENCES [I] [2] [3 ] [4 ] Jeet Kumar, Aditya Nigam, Surya Prakash, Phalguni Gupta, "An efficient pose invariant face recognition system", International Conference on Soft Computing for Problem Solving (SocProS 2011) December 20-22, 2011. Vol. -2, pp 145-152, Online ISBN 978-81-3220491-6. Rajiv Kapoor, Pallavi Mathur, "Face recognition using moments and wavelets", International Journal of Engineering and Applications (UERA) Vol. 3, Issue 4, Jul - Aug 2013, pp. 82 - 95 (ISSN:2248 9622). http://www.itl.nist.govliadlhumanid/feret/feret_master.html http://www.itl.nist.govliadlhumanid/feret/ [5] Vidit Jain, Amitabha Mukherjee. The Indian Face Database. http://viswww.cs.umass.edul-viditiIndianFace Database. [6] Jan Flusser, Thomas Suk and Barbara Zitova, "Moments and Moments Invariants in Pattern Recognition", John Wiley & Sons, Ltd. ISBN: 9780-470-69987-4 [7] Shankar Setty, et al. "Indian movie face Database: a benchmark for face recognition under wide variations", 4th IEEE National Conference on Computer Vision, Pattern Recognition, Image Processing and Graphic (NCVPRIPG), 18 - 21 December, 2013. Jodhpur, ISBN: 978-1-47991586-6 [8] R. Rajalakshmi, M. K. Jeyakumar, "A review on classification used in face recognition methods under pose and illumination variation", International Journal of Computer Science Issues (UCS!), Vol 9, Issue 6, No 2, November 2012, (rSSN:1694-0814) [9 ] Shan Du, Rabab ward, "Face recognition under pose variations", ScienceDirect, Journal of the Franklin Institute 343 (2006) 596 613 [10] Simon Xinmeng Liao, "Image analysis by moments", the department of electrical and computer engineering, the University of Manitoba Winnipeg, Manitoba, Canada, 1993. 2014International Conference on Advances in Computing, Communications and Informatics (ICACCl) Authorized licensed use limited to: MIT-World Peace University. Downloaded on May 30,2022 at 06:09:52 UTC from IEEE Xplore. Restrictions apply. 189 1