Gender Classfcaton usng Interlaced Dervatve Patterns Author Shobernejad, Ameneh, Gao, Yongsheng Publshed 2 Conference Ttle Proceedngs of the 2th Internatonal Conference on Pattern Recognton (ICPR 2) DOI https://do.org/.9/icpr.2.8 Copyrght Statement 2 IEEE. Personal use of ths materal s permtted. However, permsson to reprnt/ republsh ths materal for advertsng or promotonal purposes or for creatng new collectve works for resale or redstrbuton to servers or lsts, or to reuse any copyrghted component of ths work n other works must be obtaned from the IEEE. Downloaded from http://hdl.handle.net/72/3724 Grffth Research Onlne https://research-repostory.grffth.edu.au
2 Internatonal Conference on Pattern Recognton Gender Classfcaton Usng Interlaced Dervatve Patterns Ameneh Shobernejad School of Engneerng Grffth Unversty Brsbane, Queensland, Australa a.shobernejad@grffth.edu.au Abstract Automated gender recognton has become an nterestng and challengng research problem n recent years wth ts potental applcatons n securty ndustry and humancomputer nteracton systems. In ths paper we present a novel feature representaton, namely Interlaced Dervatve Patterns (IDP), whch s a dervatve-based technque to extract dscrmnatve facal features for gender classfcaton. The proposed technque operates on a neghborhood around a pxel and concatenates the extracted regonal feature dstrbutons to form a feature vector. The expermental results demonstrate the effectveness of the IDP method for gender classfcaton, showng that the proposed approach acheves 29.6% relatve error reducton compared to Local Bnary Patterns (LBP), whle t performs over four tmes faster than Local Dervatve Patterns (LDP). Keywords gender recognton; performance evaluaton; local dervatve pattern; Interlaced Dervatve Pattern I. INTRODUCTION Automatc face recognton and analyss s a challengng feld n computer vson whch allows many nterestng applcatons n securty ndustry and psychology. Examples nclude face detecton, person dentfcaton, age estmaton and gender classfcaton. Gender classfcaton can sgnfcantly mprove human dentfcaton as t reduces the process of matchng the face n the databases. Ths s also useful n many applcatons that may be more specfc to a partcular gender. Gender classfcaton methods can be roughly dvded nto appearance-based and feature-based methods [7]. The very early research on automatc gender classfcaton goes back to the begnnng of the 99s. The frst attempts were reported by Cottrell and Metcalfe [3] and Golomb et al. [4]. Each of them used a mult-layer neural network approach to dentfy gender from face mages. Gutta et al. [5] presented a hybrd approach consstng of radal bass functon networks and nductve decson trees. Moghaddam et al. [8] expermented wth a support vector machne (SVM) and radal bass functon (RBF) kernel. The abovementoned technques are appearance-based methods; they dentfy gender by tranng mages wthout extractng any geometrcal features. Brunell and Poggo [2] used HyperBF networks to extract a set of 6 geometrc features from frontal face mages. Sun et al. [2] proposed that feature selecton s an mportant ssue for gender classfcaton and appled genetc algorthms to select a Yongsheng Gao School of Engneerng, Grffth Unversty Queensland Research Lab, Natonal ICT Australa Brsbane, Queensland, Australa yongsheng.gao@{grffth.edu.au, ncta.com.au subset of features from frontal mages. Saatc and Town [] used an SVM that was traned wth the features extracted by an actve appearance model (AAM). Lan and Lu [6] expermented wth local bnary pattern (LBP) and SVM, and acheved consderably hgh results. The LBP features were orgnally desgned for texture descrpton [9]. The technque has been successfully mplemented n other applcatons such as face recognton [] and gender classfcaton [6]. Recently a powerful operator, called Local Dervatve Pattern (LDP) [3], has surpassed LBP n face recognton tasks. LDP encodes drectonal pattern features based on local dervatve varatons, whle LBP can conceptually be consdered as a nondrectonal frst-order local pattern. The hgh-order LDP captures more detaled dscrmnatve nformaton whch exceeds the LBP features n terms of face recognton precson. In ths paper, we nvestgate the feasblty and effectveness of LDP on gender classfcaton and propose a novel representaton of facal mages, the Interlaced Dervatve Pattern (IDP), for gender recognton. We dscuss that by applyng more effectve technques to extract features for gender dentfcaton, hgher rates n gender classfcaton wll be acheved. The proposed method s evaluated on FRGC ver.2. database []. The encouragng results demonstrate that the IDP technque outperforms LBP and LDP n gender classfcaton task, whle t operates much faster than LDP. The rest of ths paper s organzed as follows. Secton 2 presents the proposed IDP method n detal. In secton 3, the expermental results are provded. The last secton concludes the paper. II. HIGH-ORDER DERIVATIVE PATTERN In ths secton, we propose a bref revew of the conventonal hgh-order local dervatve pattern (LDP), then ntroduce the Interlaced Dervatve Pattern (IDP) and dscuss the superorty and effcency of the new approach over LBP and LDP. A. Local Dervatve Pattern The LDP operator s manly based on capturng detaled relatonshps n a local neghborhood. Whle LBP encodes the bnary results of the frst-order dervatve n the local neghborhood, the n th -order LDP can capture the changes of dervatve drectons among local neghbors. Therefore, more 5-465/ $26. 2 IEEE DOI.9/ICPR.2.8 53 59
detaled dscrmnatve features from the mage wll be obtaned by LDP, whch cannot be captured by LBP. The orgnal LDP was proposed for person dentfcaton by Zhang et al. [3]. The LDP operator fnds the dervatves along four drectons:, 45, 9 and 35. The n th -order drectonal LDP s defned as 2 3 8 4 7 6 5 (a) n n n LDP ) = { f ( I ), I ), n n n n f ( I ), I 8) f ( I ), I ),..., () where Z s a pont n the mage I(Z) and Z, =,,8 s the neghborng pont n a 3 3 neghborhood around Z, I (n-) (Z ) s the (n-) th -order dervatve n drecton at Z=Z, and ƒ(,) defned n (2) encodes the (n-) th -order gradent transtons nto bnary patterns, provdng an extra order pattern nformaton on the local regon. n n f ( I ), I )) = n n, f I ) I ) > n n, =,2,...,8., f I ) I ) The n th -order LDP s defned as n { (2) n LDP ) = LDP ) =,45,9,35 (3) Therefore for each drecton, LDP apples the encodng functon on all neghbors, regardless the drecton between the neghbor and the center pxel. Hence, t produces a 32-bt long representaton for each pxel whch makes the computatonal process much slow. B. Interlaced Dervatve Pattern IDP s a fully drectonal dervatve pattern that takes the advantage of more detaled hgh-order dervatve descrptons and keeps the spatal relatonshps n local regons. In ths technque, an IDP mage s produced for the orgnal mage. The IDP mage s a four-channel dervatve mage, representng four drectonal n th -order dervatve channels n, 45, 9, and 35, respectvely. The order of dervatves s derved from the order of the IDP operator;.e., for an n th -order IDP operator, the IDP mage wth four (n-) th -order dervatve channels s produced. These dervatve channels present more detaled descrpton of the mage n all possble drectons (see Fg. ). A 3 3 neghborhood s selected around each pont n the orgnal mage and the pxel s located n the IDP mage. For each neghbor, the drecton between the center and the neghbor s computed and the IDP mage channel wth the same drecton s selected. Channel Channel 45 Channel 9 Channel 35 (b) Fgure. (a) A 3 3 neghborhood around a pxel. (b) Four drectonal dervatve channels n the IDP mage. The neghbor s thresholded wth the center pxel value n the selected IDP channel and the result s encoded as a bnary number. Ths thresholdng actually encodes the bnary result of the frst-order dervatve among local neghbors and produce an extra order for the IDP operator. The nth-order IDP operator s presented n (4). n n n n n = { 35 35 9 9 2 n n n n f ( I45 ( z ), I45 ( z3 ), f( I ( z ), I ( z4)), n n n n f ( I35 ( z ), I35 ( z5 )), f( I9 ( z ), I9 ( z6 )), n n n n f( I45 ( z ), I45 ( z7 )), f( I ( z ), I ( z8)) IDP ) f( I ( z ), I ( z )), f( I ( z ), I ( z )), where the functon ƒ s defned as, f ( x y) f( x, y) =, f ( x y) < Therefore n each drecton, only the dervatves for the center pont and ts neghbor pont n that partcular drecton wll be calculated. Ths wll dramatcally decrease the length of the pxel representng code produced by the proposed operator compared to the LDP operator. LDP keeps the extra nformaton n a local neghborhood, whle the new approach encodes the relatonshps n the partcular drectons. In ths way, IDP keeps only the more mportant nformaton and makes the process much faster. It produces an 8-bt representaton of each pxel, whch makes the operator four tmes faster than LDP wth a 32-bt representaton of pxels. Also compared to LBP, IDP contans more detaled descrpton by calculatng the hgh-order dervatve drectonal varatons, whle LBP provdes frst-order dervatve nformaton and s ncapable of descrbng more detaled nformaton. Fg. 2 llustrates the 2nd-order IDP operator and Fg. 3 shows the vsualzed results of the IDP operator on a sample mage. (4) (5) 54 5
the dstrbuton of Interlaced Dervatve Patterns. Takng the spatal hstograms of the subregons and concatenatng them nto an enhanced feature vector as the mage descrptor s more robust aganst pose and llumnaton varatons than the holstc methods []. III. The FRGC ver.2. database [] was used n our experments. We selected 457 ndvduals; 262 male and 95 female subjects. All mages were normalzed wth the two eyes and cropped to 6 6 mages, so that each mage contaned lttle or no har nformaton. In our experments, we appled LDP and evaluated ts feasblty on the gender classfcaton for the frst tme. We also mplemented the new IDP approach as an mprovement to LDP technque and compared the performance and effcency of IDP wth LBP and LDP n gender recognton task. All three technques were used to produce approprate features for gender classfcaton. In each case, the operator was appled on all mages. The mages were dvded nto subregons and the mage feature vector was produced by concatenatng all subregon hstograms. The recognton rate was estmated wth fve-fold cross valdaton. A par of mean male/female subjects was produced for each set of tranng subjects, and hstogram ntersecton n (6) was appled to measure the smlarty between the test subject and the mean male/female subjects. (a) (b) EXPERIMENTAL RESULTS AND DISCUSSION Bnary number: (c) Fgure 2. (a) 3 3 neghborhood n orgnal mage. (b) 4-channel IDP representaton. (c) IDP code for pont. S HI ( H, S ) = = mn( H, S ) B (6) where S HI ( H, S ) s the hstogram ntersecton statstc wth H = ( H,..., H 8 )T and S = ( S,..., S8 )T. Dfferent orders for LDP and IDP operators were tested. We found that hgher order of LDP s requred for gender classfcaton than for dentfcaton. The 4th-order LDP performed the best n gender recognton whle the 3th-order had the best results n person dentfcaton [3]. For the proposed approach, the 2nd-order IDP had the hghest performance and outperformed LBP and LDP. Also dfferent numbers of hstogram bns n each subregon were expermented, and the recognton error rate curves of all the operators remaned relatvely flat (see Fg. 4). Table I shows the error rates of the three operators for male and female subjects and demonstrates that IDP technque outperformed LBP and LDP technques wth the hghest recognton rate of 9.2%. The results show that the more detaled nformaton extracted by IDP and LDP s more effectve for gender classfcaton than the frst-order dervatve nformaton of LBP. Table II demonstrates the average computaton tme for code generaton and matchng process for a sample mage based on Pentum 4 (2.8GHz) mplemented n MATLAB R28a. As shown, the proposed approach performs over four tmes faster than LDP. Fgure 3. Vsualzed results of IDP code generaton process for a face mage To extract the dscrmnatve IDP features of the mage, the mage s dvded nto rectangular subregons represented by R,, RL, and the spatal hstograms are used to model 5 55
Error rate (%) 6 4 2 8 8 32 64 28 Number of hstogram bns REFERENCES [] T. Ahonen, A. Hadd, and M. Petkanen, "Face Descrpton wth Local Bnary Patterns: Applcaton to Face Recognton," IEEE Trans. on Pattern Analyss and Machne Intellgence, vol. 28(2), 26, pp. 237-24. [2] R. Brunell and T. Poggo, "Hyperbf Networks for Gender Classfcaton," DARPA Image Understandng Workshop, 992, pp. 3 34. [3] G. Cottrell and J. Metcalfe, "Empath: Face, Emoton, and Gender Recognton Usng Holons," Neural Informaton Processng Systems, vol. 3, 99, pp. 564-57. LBP 4th-order LDP 2nd-order IDP TABLE I. Fgure 4. Classfcaton error rates CLASSIFICATION ERROR RATES OF DIFFERENT TECHNIQUES WITH 64 HISTOGRAM BINS Method error rate Female Male Overall LBP.8% 3.7% 2.5% 4 th -order LDP 4.9% 5.7% 9.6% 2 nd -order IDP 3.8% 5% 8. 8 % TABLE II. AVERAGE COMPUTATION TIME LBP LDP 2 LDP 3 LDP 4 IDP Tme(sec).4 2. 2.2 2..4 For each drecton, LDP apples the encodng functon on all neghbors around a pxel, whle IDP encodes the relatonshp between the pxel and ts neghbor n that partcular drecton. Therefore compared to LDP, our approach contans less nose and redundancy and produces compact representatve features. On the other hand, although ncreasng the order of the operator mproves the accuracy by extractng more detaled nformaton, t amplfes the nose whch defects the results accuracy. Therefore n hgher orders, the recognton accuracy start to declne and a 'best-result' order s determned for the operators. IV. CONCLUSION In ths paper, we present a hgh-order drectonal texture representaton, the Interlaced Dervatve Pattern (IDP), for gender classfcaton. It contans more dscrmnatng nformaton than the frst-order features (LBP) and dramatcally reduces the computatonal complexty compared wth the conventonal hgh-order features (LDP). The feasblty and effcency of the proposed approach s evaluated on FRGC ver.2. database and compared aganst LBP and LDP methods. The promsng results of the proposed approach n gender classfcaton mply ts potental capablty for other face recognton tasks. Besdes, the best-result order drops down from the fourth n LDP to the second n IDP, whch s an nterestng ssue for lookng nto detal n further study. [4] B. Golomb, D. Lawrence, and T. Sejnowsk, "Sexnet: A Neural Network Identfes Sex from Human Faces," Advances n neural nformaton processng systems, vol. 3, 99, pp. 572-577. [5] S. Gutta and H. Wechsler, "Gender Classfcaton of Human Faces Usng Hybrd Classfer Systems," Internatonal Conference on Neural Networks, vol. 3, 997, pp. 353-358. [6] H.C. Lan and B.L. Lu, "Mult-Vew Gender Classfcaton Usng Mult-Resoluton Local Bnary Patterns and Support Vector Machnes," Internatonal Journal of Neural Systems, vol. 7(6), 27, pp. 479-487. [7] E. Maknen and R. Rasamo, "An Expermental Comparson of Gender Classfcaton Methods," Pattern Recognton Letters, vol. 29(), 28, pp. 544-556. [8] B. Moghaddam and Y. Mng-Hsuan, "Learnng Gender wth Support Faces," IEEE Trans. on Pattern Analyss and Machne Intellgence, vol. 24(5), 22, pp. 77-7. [9] T. Ojala, M. Petkanen, and T. Maenpaa, "Multresoluton Gray-Scale and Rotaton Invarant Texture Classfcaton wth Local Bnary Patterns," IEEE Trans. on Pattern Analyss and Machne Intellgence, vol. 24(7), 22, pp. 97-987. [] P.J. Phllps, P.J. Flynn, T. Scruggs, K.W. Bowyer, C. Jn, K. Hoffman, J. Marques, M. Jaesk, and W. Worek, "Overvew of the Face Recognton Grand Challenge," IEEE Computer Socety Conference on Computer Vson and Pattern Recognton, vol., 25, pp. 947-954. [] Y. Saatc and C. Town, "Cascaded Classfcaton of Gender and Facal Expresson Usng Actve Appearance Models," 7th Internatonal Conference on Automatc Face and Gesture Recognton, 26, pp. 393-398. [2] Z. Sun, G. Bebs, X. Yuan, and S.J. Lous, "Genetc Feature Subset Selecton for Gender Classfcaton: A Comparson Study," Sxth IEEE Workshop on Applcatons of Computer Vson, 22, pp. 65-7. [3] B. Zhang, Y. Gao, S. Zhao, and J. Lu, "Local Dervatve Pattern Versus Local Bnary Pattern: Face Recognton wth Hgh-Order Local Pattern Descrptor," IEEE Transactons on Image Processng, vol. 9(2), 2, pp. 533-544. 56 52