Ab pro hu pro dis usu suf sub Be sop the pre res bot usi run dom iris AT 88. atta pri tru pro [1] sca soc to Ind mu att sys im iris rec im sta dow and tho fro Engineerin www.etasr F Jawaharl su bstract—Iris re omising nonin man identific ogram in Indi stinctive identi ually captured ffers from poor bject's relativ sides the tech phisticated thr e integrity of eserved. This p sistant approac th general eye ing spatial gra n length mat main. We did s image datab TVS-FIr iris d .83% with IITD Keywords-iris acks; contact le Iris recogniti ivate and go ustworthiness. omising techn ]. Iris recogni ale biometric cial benefits, e provide depen dia [2]. As a r ust also ensu acks, which c stems. The p mages and eye s recognition s Spoofing at cognition syst mages printed ated that comm wn to a resolu d at least 75% ose detailed i om the usage ng, Technology r.com Feature P. V. L. Su lal Nehru Tech Kakin Andhra Prad uvarchala_pvl ecognition is nvasive biomet cation. Numer ia - Aadhar, ity identificatio d under non id r brightness, lo ve movement nical challeng reats like spoof f such large paper presents ch which explo e and localized ay level depend trix (GLRLM experiments o bases. Correct database was D spoof iris da s recognition; ens attacks I. IN ion is an exten overnment ag It has been nologies to pro tion is now a identification e.g. iris biomet ndable identifi result, the vera ure to avoid can cause vul possibility of images with system. ttack techniq tem by presen using quality mercial system ution of 1200 % of the iris be n [4] show m e of textured y & Applied Sci e Set F uvarchala hnological Un nada desh, India l@yahoo.com considered as tric systems p rous program include iris b on to citizens. deal imaging c ow contrast, bl and eyelid ges, iris recogn f attacks. Ther scale iris dep s the developm oits the statist d iris regions dence matrix (S M) and contou on publicly ava classification 100% while atabases respec liveness detect TRODUCTION nsively used b gencies for n treated as ovide reliable an indispensab n for financia trics is used in fication for mi acity of such a likely threats lnerability to using good q contact lens c ques try to nting synthetic printing devi ms can be foo dpi, an iris di ing visible. La more sophistic superficial l ience Research Fusion niversity m s one of the providing auto ms, like uniqu biometric to pr . The active a onditions. It u lur due to cam eyelash occlu nition started refore it is vita ployments mu ment of a new ical dependenc in textural d SGLDM), gray urlets in tran ailable fake an rate obtained it was 95.63% ctively. tion; spoofing; biometric syste its precision one of the human recog ble module of al transaction n the Aadhar p illions of citiz a large scale s s such as spo the iris recog quality printe can compromi deceive the cally generate ices. Author oled with a pr iameter of 75 atest attacks su cated threats lenses that p h V Suvarchala an n for S J most mated ue ID rovide area is usually mera or usions. facing al that ust be spoof cies of omain y level nsform nd lens d with % and ; print em by n and most gnition f large ns and project ens in system oofing gnition ed iris se the e iris ed iris in [3] rintout pixels uch as rising present cha pos Wh cos oph to reco sho pro iris step [6] text dete obf cas thre bein sam AT fak from sho tran ava com Fig. from feat stoc com dev Vol. 8, No. 2, 20 nd Srinivas Ku Spoof Jawaharlal Neh anged iris text ssibly compro hile colored o smetic reaso hthalmologists correct the e ognition algor own that the ominent and co patterns. Dete p toward impr applied Fouri tured lenses. The hypothes ection algorit fuscated patter es. In the pro ee databases, ng soft lens a mple images o TVS-FIr a pub ke and real ima m both eyes o ows sample i nsparent conta ailable databas mprising of 65 1. Row 1: E m ATVS-FIr Data The proposed tures such chasticity of t mputed from veloped appro 018, 2859-2863 umar: Feature Iris D S. Srinivas Departmen hru Technolog Andhra Prad samay_ssk2@ ture to the iri omise the ins or textured con ons, it is s to prescribe eye sight. He rithm to find th texture detail ourse and they ection of cont roving reliabil ier transform t sis generated thm has to rns and allow oposed metho one of them b and textured le of the real and blicly availabl ages of iris (8 of 50 subjects mages of the act lens) iris im se [9] acquired 70 iris images ye images with p abase d spoof iris de as intensity the iris texture the gradien oach also exp 3 Set Fusion for etectio s Kumar nt of ECE gical Universit desh, India @yahoo.com is acquisition stalled iris rec ntact lenses a becoming contact lens f ence it is un he presence of ls in near-inf y will obfuscate act lens presen lity of iris reco to detect the f and tested in first reject th w only the sof od we tested being print att ens cases. Fig d fake (printed le database [8 00 from each treated as 10 e real and fa mages of IIIT- d from 101 su s. print attack, Row etection approa distribution, e and edge inf nt of global ploits the textu 2859 Spoof Iris Dete on ty, Kakinada system which cognition sys are being wor a practice for medical rea navoidable fo f lens. In [5] i frared imaging e the real or na nce is an impo ognition. Auth fake patterns d [7] is that, th he instances ft lens and no the hypothes tack and the o gure 1 shows d) iris images 8]. There are category) cap 0 classes. Fig ake (cosmetic D Vista, a pub ubjects, 202 cl 2: Original eye i ach analyzes i , coarseness formation whi eye image. ural region o ection h can stems. rn for for asons r the it was g are atural ortant hor in due to he iris with o lens is on others some from 1600 ptured gure 2 c and blicly lasses images mage and ich is The of the ent pro app dev att of als tra fro cla sup fea par exp con Fig len A. pre the rel and Engineerin www.etasr tire eye imag ovide very us proach for sp veloped appro acks, while su soft (transpar so computed ansform. Conto om images assification pe pport vector m The rest of th ature set gen rtition and cl perimental re nclusions of th g. 2. Row 1: C s images, Row 3: Fig Literature Su A detailed esented in [1]. e fake iris patt liable. Authors d other quality ng, Technology r.com ge using SGL seful descript poof iris detec oach works o ustaining recog rent) lens. In transform do ourlets are su with smooth erformance of machines (SVM his paper is st neration is de lassification a esults are pre he proposed m Colored contact l Images with no l g. 3. Proposed urvey survey on i . Author in [6] terns due to th s in [10] prov y related meas y & Applied Sci LDM and G tors. The bloc ction is shown on spoofed im gnition capabi addition to te omain feature uitable for ext h edges an f the algorith Ms). tructured as fo escribed. In are presented. esented. In s method are give lens images, Row lens from IIITD d d method block d ris recognitio ] used Fourier he presence o ved that focus sures can help ience Research GLRLM which ck diagram o n in Figure 3 mages with p ility in the pre extural feature es using cont tracting inform nd contours. hm is tested ollows. In secti section III d In section IV section V the en. w 2: Transparent database. diagram on algorithms r transform to f lens which i s, motion, occ p fake iris dete V Suvarchala an h can of our 3. The printed esence es, we tourlet mation The using ion II, dataset V the e key contact s was detect is less lusion ection. The reco LB atta vec ima mea cha Mu algo in [ LB [7] lens entr wor usin A. extr in circ diff met from is n Hen and sys thro unw B. par orig rep tem glob 4. T calc gray ima crea from Vol. 8, No. 2, 20 nd Srinivas Ku e effect of both ognition was p P operator an acks in iris re ctor machines ages. In [13], asuring edge aracterization t ulti-scale loca orithm were u [15] applied G P for the detec tested the tran s images using ropy with LBP rks well on pr ng textural and II. IRIS PR Iris Segmenta In any biome raction of diss deciding the cular iris and ferential opera thod proposed m ATVS-Fir d not possible t nce they cont d GLRLM) an tem as ‘Fakes ough the seg wrapped and c SGLD Matrix There are thr rts represent ginal eye imag resents the co mplate. SGLD bal descriptor The texture of culating ,(ap ay levels a and age, creating ated, then stat m it and provi Energy:  ba p , ( Contrast:  ba, Correlation: Sum of squar 018, 2859-2863 umar: Feature S h transparent a presented in [ nd statistical f ecognition in] to train the G , authors det sharpness in through iris-te al binary pat used to classif Gaussian smoo ction of textur nsparent conta g modified LB P to detect live rint spoof iris i d contourlet fe REPROCESSING ation and Norm etric system, p similar feature recognition pupil region ator and the iri d in [17]. As so database in [8 to apply cont tain only textu nd subsequentl s’. Images with mentation ste contourlet featu x and GLRL M ree parts in th the SGLDM ge and the grad ontourlet trans features and s derived from f an image is d )b i.e., how fre d b in a stated an SGLD ma istical parame de information ba 2),(  bapba 2 ,()( yx ba bapba   , ),( es: Variance  a 3 Set Fusion for and colored co [4, 5, 7]. Auth features for id . Authors in GLCM feature tected textured n the iris im extons and co- tterns and A fy textured len othed as well red lens in iris act lens images BP. Authors in eness. The me images and iri eatures. G AND FEATUR malization preprocessing es from it play rate. In the ns are segmen is region is un ome of the pri 8] failed the se tourlet decom ural features ( ly the images h textured lens ep successfull ures are extrac Matrix he feature vec M and GLRL dient of eye im sform features GLRLM feat m the eye imag described by S equently pixel d spatial relati atrix. Once th eters proposed n about the im b) y yx   ba apa , 2 ()(  2860 Spoof Iris Dete ontact lenses o hors in [11] ap dentifying the [12] used su es of textured d contact len mage, iris te -occurrence m AdaBoost lea ns in [14]. Au as SIFT weig images. Autho s as well as co n [16] made u thod proposed is with contact RE EXTRACTION g of the image ys an essentia current study nted using int nwrapped usin int attacked im egmentation st mposition on t (gradient, SGL are rejected b s and soft lens ly which are cted from them ctor. The first LM parameter mage. The third of normalize tures are low ge shown in F SGLD function l pairs with spe ionship occur he SGLD matr in [18] are de mage texture. (1) (2) (3) ba ), (4) ection on iris pplied print upport d lens ns by exture matrix. arning uthors ghted ors in olored use of d here t lens N e and l role y, the tegro- ng the mages tep, it them. LDM by the s pass then m. t two rs of d part ed iris level Figure ns by ecific in an rix is erived C. hav of GL dir Te num - S - L - G - R - R wh of D. Af ver the ma Engineerin www.etasr Inverse differ Entropy: a  GLRL Matrix A GLRL m ving the same pixels in the LRL matrix, rections having Fig. 4. R exture measure Let k be the     g rN i N j ipk 1 1 ( where Ng is mber of differ Short Runs Em Long Runs Em Gray level non Run length non Run percentage here p is the n one. Gradient Me One of the m fter gradient im rtical direction e direction of ay be found i ng, Technology r.com rence moment log[),( , bap ba  x matrix is a set gray level. Th e run. Let (p specifying th g the gray leve Row 1: SGLDM ex es measured fr e number of ru ji ), number of gr rent run length mphasis: RF1 mphasis: 2RF -uniformity: R n-uniformity: e: /1(5 pRF  number of runs easure most common u mages have b ns, the pixels w f the gradient in the directio y & Applied Sci t:  ba p ba, 2)( 1 )],( bap of collinear, he length of th ),( ji be the ( he number of el i and length xample. Row 2: G rom the GLRL uns. Then ray levels in t hs.     g rN i N j j jip k 1 1 2 ),(1 ,( 1 1 21 ip N i N j j k g r     1 1 3     gN i N jk RF 1 1 4    rN j N ik RF ),( 1 1 ) jip N i N j p g r     s if all possibl uses of gradien been computed with the highes become edge on perpendicu ience Research bap ),( (5) (6) consecutive he run is the nu thji ),( entry i f primitives h .j GLRLM example LM are as follo (7) the image and (8) ), j (9) 2 1 ),(   rN jip (10 2 1 ),(   gN i jip (11 (12 le runs had a l nt is edge dete d in horizonta st gradient val e pixels, and ular to the gr V Suvarchala an pixels umber in the of all ows: d Nr is 0) ) 2) length ection. al and lues in edges radient dire lens ima sho the pre lim var Fig. imag Fig. Fir r E. effi as a the The pyr poin resp ban coin sign 40× leve com feat mad and sum dete Vol. 8, No. 2, 20 nd Srinivas Ku ection. There i s image from age texture is c own in Figure real image i sence of conta mbic and pupil riance and mor 5. Row 1: ges. Row 2: Eye i 6. Row 1: G real and fake eye Contourlet Fe Contourlets iciently with li a tool to comp geometrical e double filte ramid (LP) and nt discontinui pectively. Hen nk (PDFB). It ncide with lo nificant coeffi ×240 template el and 4 dire mputed for eac ture vector. In de up of SGL d contourlet co mmarizes the ection. 018, 2859-2863 umar: Feature S is almost no di m a normal ir course and ran 4. The iris stru is smooth but act lens. The f llary boundari re uniformity a Gradient images images with textu Gradient images o images eature Extract [21] represe ittle redundanc pensate the lim structures tha er bank struc d directional f ities and linki nce the name can be noticed ocation and d icients. Once t , it is then dec ections at II l ch direction a n the proposed LDM, GLRLM oefficients of n comparison 3 Set Fusion for ifference in th ris image but ndom, which d ucture in the s t the additiona fake iris with ies cannot be as shown in Fi s of colored, tra ured lens, soft len of real and fake i tion ent the geo cy. Contourlet mitation of wa at possess dir cture compris filter bank (DF ing them in t pyramidal an d that only tho direction of th the iris image composed in t level. Statistic and all are app d method the ir M of original an normalized iri of existing 2861 Spoof Iris Dete e structure of t textured len distinctly appea soft lens imag al edge show print attack (w localized) has igure 5. ansparent and no s and no lens. images. Row 2: A ometric struc t transform is f avelets in han rectional mom sing of Lapl FB) is used to to linear struc nd directional ose contourlet he image pro is unwrapped to 16 direction cal parameter pended to form ris feature vec nd gradient im is template. Ta methods in s ection a soft ns iris ars as e and ws the where s less o lens ATVS- ctures found ndling ments. lacian catch ctures filter s that oduce d in to ns at I rs are m the ctor is mages able I spoof Engineering, Technology & Applied Science Research Vol. 8, No. 2, 2018, 2859-2863 2862 www.etasr.com Suvarchala and Srinivas Kumar: Feature Set Fusion for Spoof Iris Detection TABLE I. COMPARISON OF EXISTING METHODS IN SPOOF DETECTION Method Attack Nature Feature Training/Testing Protocol Matching/Classifier [11] Print spoof Statistical features of local and global eye regions ATVS-Fir: 400 training and 1200 testing Neural networks [7] Contact lens Modified LBP IIIT-D: Weka classifier [14] Contact lens LBP features 300 fake+6000 live irises for training and 300 fake + 4000 live irises for testing AdaBoost classifier [13] Contact lens Iris textons and Haralick features CASIA and Bath: Equal train and test samples SVM classifier [19] Contact lens Scale invariant descriptors IIIT-D Vista & IIIT-D Cogent: equal number of training and testing samples SVM classifier [16] Print spoof Entropy with LBP and WP ATVS-Fir and CASIA-Iris syn databases -- [10] Print spoof Quality related measures and SFFS algorithm 200 real images for training and 1400 testing Quadratic classifier Proposed Contact lens and print spoof SGLDM, GLRLM features and contourlet features ATVS-Fir: 500 real for training and 1100 for testing SVM classifier III. DATASET PARTITION The datasets’ description is given in the Table II. Since there are unequal numbers of samples for left and right irises in the IIIT-D databases, we considered 5 samples/each class for the experiment. In ATVS-Fir database 500 samples are used for training and the rest 1100 (300 real + 800 fake) samples are used for testing. The eye image and gradient of the eye image are initially subdivided into blocks sized 80×80 which then form a total of 48 subimages. Feature vectors made up of statistical features defined in section II (B & C) of SGLDM and GLRLM of subimages are extracted. The size of the textural features will be 1056 (11 features/subimage × 48 subimages × 2 images). Contourlet decomposition of the iris template generates 20 subimages. The number of frequency domain features is 220 (11 features × 20 wedges). The final feature vector size is 1276 (1056+220). Those fake print attacked images in ATVS-Fir database which fail the segmentation step, will not have contourlet features but contain only the global textural features. At the end of each feature vector the class name is appended facilitating the supervised learning. TABLE II. DATABASE DETAILS Database Parameter| IIIT_D Vista IIIT-D Cogent ATVS-FIr Subjects 101 101 50 Categories Textured lens Soft lens Normal textured lens Soft lens Normal Fake Real Samples/Subject Left iris=5 Right iris=5 Left iris =5 Right iris=5 Left iris=8 Right iris=8 Classes 606 606 200 Total samples 3030 3030 1600 Eye image size 640×480 640×480 640×480 Train samples 3/class 3/class 5/class (real) Test samples 2/class 2/class 3/class (real) + all fake The classification of iris feature vectors is carried out using osusvm package available at [20]. The multiclass SVMs operated in one-against-all approach are tuned with linear support vector classifier kernel. SVMs are trained and tested at train sample:test sample ratio of 3:2 for IIITD databases. The classification of normal, soft lens and textured lens iris images in IIIT-D Vista and Cogent databases is conducted separately as well as mixed. The left and right irises are considered as separate classes in all databases. The train:test sample ratio is 5:3 for ATVS-Fir real database. In this database 500 real iris samples are taken for training and the remaining 1100 (real + fake) samples are considered for testing. IV. RESULTS AND DISCUSSION The experiments are conducted on a PC with Pentium i5, 2.67GHz processor and 4GB RAM in MATLAB 11.0 environment. The evaluation of the proposed method is done on ATVS-FIr, IIIT-D Vista and IIIT-D Cogent databases. The performance of the proposed method is appraised in terms of correct classification rate (CCR) and compared with existing methods in Table III. The method proposed in [7], obtained 76.83% and 65.03%. The method described in [19] obtained 89.78% and 83.88% and the proposed method achieved better CCR of 95.63% and 88.83% on IIITD-Vista and IIITD-Cogent databases respectively. On ATVS-Fir database the proposed method achieved 100% CCR which is same with [10] and Waleed and Hanna in [16] but better than Chun in [11] which obtained 99.75%. The CCR is less when only either textural or contourlet features are considered but significantly improved when both features are combined. For mixed IIITD database, SVMs are trained with normal and soft lens iris and tested with all three types including textured lens irises. The obtained results are given in Table IV. TABLE III. %CRR COMPARISON OF PROPOSED AND EXISTING METHODS Database Method CRR (%) ATVS-Fir [8] [11] 99.75 [16] 100 [10] 100 Proposed 100 IIITD-Vista [9] Lens* N-N T-T S-S Avg** [19] 87.77 98.59 82.99 89.78 [7] 62.10 92.95 75.44 76.83 Proposed 97.64 98.85 90.40 95.63 IIITD-Cogent [9] Lens* N-N T-T S-S Avg** [19] 79.80 95.54 76.29 83.88 [7] 65.99 80.81 48.31 65.03 Proposed 85.50 96.75 84.24 88.83 Note*: N-Normal, S-Soft lens and T-Textured lens. Avg**: Average Engineering, Technology & Applied Science Research Vol. 8, No. 2, 2018, 2859-2863 2863 www.etasr.com Suvarchala and Srinivas Kumar: Feature Set Fusion for Spoof Iris Detection TABLE IV. PERFORMANCE COMPARISON OF DIFFERENT FEATURES IN % CCR Database Lens SGLDM &GLRLM Contourlet Features Feature Fusion ATVS-Fir [8] -- 98.40 99.25 100 IIITD-Vista [9] N-N 94.68 96.56 97.64 T-T 93.65 95.35 98.85 S-S 86.24 87.80 90.40 Mixed* 84.52 86.23 89.46 IIITD-Cogent [9] N-N 80.40 82.78 85.50 T-T 90.24 93.45 96.75 S-S 79.98 81.25 84.24 Mixed* 78.54 80.82 82.66 Note: Mixed* is trained with normal and soft lens and tested with all 3 types TABLE V. COMPUTATION TIME FOR EACH BLOCK Parameter Computation time (s) of all databases* Image size 640×480 Gradient 0.12 SGLDM 0.105 GLRLM 0.11 Preprocessing 0.68 Normalization 0.15 Contourlet features 0.26 Authentication 0.012 Total time 1.437 All databases*: ATVS-Fir, IIITD-Vista and IIITD-Cogent V. CONCLUSIONS The method proposed in this paper identifies iris images with soft or textured lens and detects print attacks. The set of features consists of optimized textural and frequency domain descriptors. The print spoof images which fail the segmentation step have their contourlet features null. Hence the detection of print attacks is easy during classification. Hence the CCR is 100% for ATVS-Fir database. There is not much difference between normal and soft lens irises. So the classification rate is high with IIITD databases also. The proposed method is well suited for print spoof as well as lens images. ACKNOWLEDGMENT The ATVS-FIr data set is released by the Biometric Recognition Group, UAM, Madrid, Spain and the IIIT-D Vista and Cogent dataset by IIIT-Delhi, India. REFERENCES [1] K. W. Bowyer, K. Hollingsworth, P. J. Flynn, “Image understanding for iris biometrics: a survey”, Computer Vision and Image Understanding, Vol. 110, No. 2, pp. 281-307, 2008 [2] Unique Identification Authority of India, Role of biometric technology in aadhar enrollment, available at https://pdfs.semanticscholar.org /2db1/e82a61e1b30df3853c429f5a669cdb37c41e.pdf, 2012 [3] J. Lee, “Spoofing iris recognition technology with pictures”, available at: https://www.biometricupdate.com/201503/spoofing-iris-recognition- technology-with-pictures, 2015 [4] J. S. Doyle, P. J. Flynn, K. W. Bowyer, “Automated classification of contact lens type in iris images”, 2013 International Conference on Biometrics (ICB), Madrid, Spain, June 7-13, 2013 [5] N. Kohli, D. Yadav, M. Vatsa, R. Singh, “Revisiting iris recognition with color cosmetic contact lenses”, 2013 International Conference on Biometrics (ICB), Madrid, Spain, June 7-13, 2013 [6] J. Daugman, “Demodulation by complex-valued wavelets for stochastic pattern recognition”, International Journal of Wavelets, Multiresolution and Information Processing, Vol. 1, No. 1, pp. 1–17, 2003 [7] D. Yadav, N. Kohli, J. S. Doyle, R. Singh, M. Vatsa, K. W. Bowyer, “Unravelling the effect of textured contact lenses on iris recognition”, IEEE Transactions on Information Forensics and Security, Vol. 9, No. 5, pp. 851-862, 2014 [8] Biometric Recognition Group–ATVS, ATVS-Fir, Fake Iris Images Database, available at: https://atvs.ii.uam.es/atvs/fir_db.html [9] Image Analysis and Biometrics Lab at in IIIT Delhi, IIITD Contact Lens Database [10] J. Gallaby, J. Ortiz-Lopez, J. Fierrez, J. Ortega-Garcia, “Iris liveness detection based on quality related features”, 5th IAPR International Conference on Biometrics (ICB), New Delhi, India, pp. 271-276, March 29 - April 1, 2012 [11] C.-W. Tan, A. Kumar, “Integrating occular and iris descriptors for fake iris recognition”, 2014 International Workshop on Biometrics and Forensics (IWBF), Valletta, Malta, March 27-28, 2014 [12] X. He, S. An, P. Shi, “Statistical texture analysis-based approach for fake iris detection using support vector machines,” Lecture Notes in Computer Science, Vol. 4642, pp. 540-546, 2007 [13] Z. Wei, X. Qiu, Z. Sun, T. Tan, “Counterfeit iris detection based on texture analysis,” 19th International Conference on Pattern Recognition, Tampa, USA, December 8-11, 2008 [14] Z. He, Z. Sun, T. Tan, Z. Wei, “Efficient iris spoof detection via boosted local binary patterns”, Lecture Notes in Computer Science, Vol. 5558, pp. 1080-1090, 2009 [15] H. Zhang, Z. Sun, T. Tan, “Contact lens detection based on weighted LBP”, 20th International Conference on Pattern Recognition, Istanbul, Turkey, pp. 4279-4282, August 23-26, 2010. [16] W. S. -A. Fathy, H. S. Ali, “Entropy with local binary patterns for efficient iris liveness detection”, in: Wireless Personal Communications, Springer Science+Business Media, LLC, 2017 [17] J. Daugman, “How iris recognition works”, IEEE Transactions on Circuits and Systems for VideoTechnology, Vol. 14, No. 1, pp. 21-30, 2004 [18] R. M. Haralick, K. Shanmugam, I. Dinstein, “Textural features for image classification”, IEEE Transactions on Systems, Man and Cybernetics, Vol. SMC-3, No.6, pp. 610-62, 1973 [19] D. Gragnaniello, G. Poggi, C. Sansone, L. Verdoliva, “Contact lens detection and classification in iris images through scale invariant descriptor”, 10th International Conference on Signal-Image Technology and Internet-Based Systems, Marrakech, Morocco, November 23-27, 2014 [20] Support Vector Machines Toolbox, available at https://sourceforge.net/ projects/svm/ [21] M. N. Do, M. Vetterli, “The Contourlet transform: an efficient directional multiresolution image representation”, IEEE Transactions on Image Processing, Vol. 14, No. 12, pp. 2091–2106, 2005