Indonesian J our nal of Electrical Engineering and Computer Science V ol. 23, No. 3, September 2021, pp. 1458 1469 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v23.i3.pp1458-1469 r 1458 Multilinear principal component analysis f or iris biometric system Chetana Kamlaskar 1 , Aditya Abh yankar 2 1 School of Science and T echnology , Y . C. M. Open Uni v ersity , Maharashtra, India 2 Department of T echnology , Sa vitribai Phule Pune Uni v ersity , Pune, India Article Inf o Article history: Recei v ed Apr 1, 2021 Re vised Jul 12, 2021 Accepted Jul 28, 2021 K eyw ords: Feature fusion Iris biometric Multilinear principal component analysis Multilinear subspace learning W a v elet pack et decomposition ABSTRA CT Iris biometric modality possesses inherent characterist ics which mak e the iris recogni- tion system highly reliable and nonin v asi v e. No w adays, research in this area is chal- lenging compact template size and f ast v erification algorithms. Special ef forts ha v e been emplo yed to minimi ze the size of the e xtracted features without de grading the performance of the iris recognition system. In response, we propose an impro v ed feature fusion approach based on multilinear subspace learning to analyze Iris recog- nition. This approach consists of four stages. In the first stage, the e ye image is se g- mented t o e xtract the iris re gion. In the second step, w a v elet pack et decomposition is conducted to e xtract features of the iris image , since good tim e and frequenc y resolu- tions can be pro vided simultaneously by the w a v elet pack et decomposition. In the ne xt step, all decomposed nodes or pack ets are arranged as a 3 r d order tensor rather than a long v ector , in which feature fusion is directly implemented with multilinear prin- cipal component analysis (MPCA). This approach pro vides a more compact or useful lo w-dimensional representation directly from the original tensorial representation. Fi- nally , a discriminati v e tensor feature selection mechanism and classification strate gy are applied to iris r ecognition problem. The obtained results indicate the usefulness of MPCA to select discriminati v e features and fuse them ef fecti v ely . The e xperimental results re v eal that the proposed tensor -based MPCA approach achie v ed a competiti v e matching performance on the SDUMLA-HMT Iris database with an adequate accept- able rate. This is an open access article under the CC BY -SA license . Corresponding A uthor: Chetana Kamlaskar School of Science and T echnology Y . C. M. Open Uni v ersity , Maharashtra, India Email: chetana.kamlaskar@gmail.com 1. INTR ODUCTION Iris recognition is one of the most trusted biometric technologies in terms of human ide n t ification and v erification with a wide range of applications, including airport automatic check-in, access systems or humanitarian aid missions, and man y more. Compared with f ace and fingerprint biometric, iris pattern has rich te xture information [1] details s uch as rings, corona, crypts, contraction furro ws, ciliary processes, freckles, and colouration. Iris patterns are unique and highly distincti v e, and non-in v asi v e as well as highly stable with time. F or accurate iris recognition of indi viduals, the most discriminating information contained in the iris pattern needs to be e xtracted. Hence, it is crucial to choose a suitable method for feature e xtraction [2]. More discriminating features can be e xtracted in a w a v elet transform (WT) domain than in a time domain. This w ork uses significant features e xtraction based on w a v elet pack et decomposition (WPD) using Haar w a v elet. WPD J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1459 gi v es reasonably be tter performance because the dominant frequencies of iris te xture are located in the lo wer and middle frequenc y channels. In the first approach of the e xperiment, ener gy (E) based criterion is used to select the a p pr o pr iate w a v elet pack et after 3-l e v el decomposition. Then using an adapti v e threshold, appropriate pack et coef ficients are quantized into 1 , 0 or 1 . The feature v ector is generated by using the concatenation of quantized coef- ficients of appropriate ener gy pack ets. F or the classification of iris recognition system, triangle square ratio similarity measure is used. This approach is implemented just for comparison with other approaches. This w ork is mainly focused to propose a ne w ef ficient and rob ust algorithm for compact feature rep- resentation and classification of Iris images. The proposed algorithm dif fers fr om the e xisting Iris recognition system at the feature representation and classification stage. Here, in the second approach of the e xperiment, after 3-le v el WPD of a normalized iris image, all pack ets WPD e xcept the first pack et which represents DC component, are used to represent the features of the iris image. All of these pack ets are arranged into a 3 r d order tensor rather than a long v ector . This 3 r d order tensor is further processed by multilinear principal component analysis (MPCA), as MPCA represents multidimensional data as tensors rather than v ectors, with three k e y benefits, preserv e the multidimensional structure, lo wer computational demand, and requires fe wer parameters to estimate [3]. So, by using MPCA, the ef fecti v e components for each feature can be selected and e xtracted si- multaneously , and combined together . These fused components, as a ne w feature of the iris, is fed to a modified angle distance (MAD) classifier for automated classification. MPCA is pro v ed to be a more ef fecti v e method for multiple feature fusion and representation. The contrib ution of our w ork as follo ws: - Utilized MPCA for the discriminati v e feature selection from WPD tensors of iris images and MAD similarity measure for classification - The proposed approach create a compact lo w-dim ensional discriminati v e feature v ector and run with minimum computational time. - The proposed approach is e v aluat ed with recei v er operating characteristic (R OC) and equal error rate (EER) on the SDUMLA HMT Iris dataset. The paper is or g anized as follo ws: Section 2 describes related w ork, the proposed MPCA is outlined in Section 3, the e xperimental design is presented and the results are discussed in Section 4, and finally , Section 5 concludes the w ork. 2. RELA TED W ORK The most successful pro v en methods for iris recognition include w ork proposed by Daughman [4]. In order to e xtract iris features, he mak es use of quadrature 2D Gabor w a v elets and encodes the iris image to a binary code of 256 bytes (2048 bits) in length, referred as an iris code. Hamming distance is used to indicate the similarity of tw o iris codes. Iris recognition system proposed by [5] uses Laplacian of Gaussian filters for the decomposi tion of the iris re gion. Then, constructed Laplacian p yramid to generate a compact iris template. The similarity between tw o iris templates is determined using correlation comparison. Boles and Boashash [6] ha v e proposed a system based on dyadic 1D w a v elet transform with the zero crossing detectors for iris feature e xtraction and mak es us e of tw o dissimilarity functions for comparison of iris representation. The y claim that noise influences can be eliminated with the zero crossing detectors [6]. Zhu et al. [7] emplo y multi-channel Gabor filtering and the w a v elet transform to e xtract iris feature v ector and mak es use of weighted Euclidean distance classifier to identify the iris. An iris image is decomposed using 2D Haar w a v elet transform by [8]. In this w ork, 87-bit code feature is generat ed by quantizing the fourth le v el high frequenc y information, and a modified competiti v e learning neura l netw ork (L VQ) is used for classification. W a v elet pack et transform (WPT) using Haar w a v elet for e xtraction of iris te xture approach is used in [9]. In this study , only suitable sub images are selected by applying WPT decomposition. Then, WPT coef ficients of selected sub images are encoded as iris feature v ector and compared Manhattan distance between the tw o corresponding iris v ectors for matching. The approach proposed by Hariprasath and V enkatasubramanian [10] is based on 2D WPT . First, iris re gion is encoded into a sequence of 2D w a v elet pack et coef ficients with a size of the feature v ector of 1280 bits. Then, e xclusi v ely OR comparisons are made bet ween tw o dif ferent iris codes. The approach presented in [11] proposes Iris feature e xtraction using Haar w a v elet on IITD database and Hamming distance matcher to achie v e higher v erification performance. Recently , biometric authentication proposed by [12] uses continuous curv elet transform combined with PCA for Iris feature e xtraction. The perfor mance is e v aluated with three Multilinear principal component analysis for iris biometric system (Chetana Kamlaskar) Evaluation Warning : The document was created with Spire.PDF for Python.
1460 r ISSN: 2502-4752 classifiers - k-nearest neighbors (KNN), support v ector machine (SVM), neural netw ork (NN) and achie v ed a v erage recognition rates of 91.0%, 93.0%, and 97.0% respecti v ely . According to these pre vious studies a w a v elet transform is one of the rele v ant tools to e xtract the most distincti v e features contained in an iris image. Hence, for tensor representation, WPD w as chosen which has linear computational comple xity . MPCA based tensor feature e xtraction has found widespread use in v arious applications of computer vision and pattern recognition, recent applications include f ace recognition [13], signal processing, handwrit- ing, digital number recognition, content analysis, anomaly detection in data [14], g ait recognition [15]. A ne w frame w ork of MPCA for dimensionality reduction and feature e xtraction of the tensor object is proposed by [3] with an application to g ait recognition. Moti v ated by the success of MPCA in feature e xtracti on, in this w ork, we propose feature fusion using tensor based MPCA for Iris recognition. 3. THE PR OPOSED METHOD In general, the iris recognition system consists of four processing modules - Se gment ation, Norm alisa- tion, Feature e xtraction and encoding, and Matching. Figure 1 sho ws the block diagram of the proposed feature fusion method. Here, we aim to ef fecti v ely perfor m multiple feature fusion using tensor -based multi-linear subspace learning method. Figure 1. Block diagram of the proposed iris recognition system 3.1. Pr epr ocessing The first step is se gmentation, where iris re gion is isolated from an e ye image. This step plays a k e y role in the recognition performance. As improper se gmentation can lead to incorrect feature e xtraction, illumination normalization is performed pri or to iris se gmentation [16], [17]. In the ne xt step, normalization is done to transform or map the e xtracted iris re gion into a fix ed rectangular block as the size of the iris may dif fer from one e ye to another . F or this, Daugman’ s Rubber sheet model [4] is used. In this, each pix el of the isolated iris is remapped to a pair of polar coordinates to mak e iris representation in v ariant to the size of iris and pupil Indonesian J Elec Eng & Comp Sci, V ol. 23, No. 3, September 2021 : 1458 – 1469 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1461 dilation. Finally , the normalized iris is subjected to feature e xtraction. Before that, histogram equalization is performed to enhance the quality of normalized iris. 3.2. Iris featur e extraction using WPD Only the significant features of the iris pat tern must be e xtracted and encoded for accurate recognit ion of indi viduals. In this w ork, WT is used to e xtract features from the enhanced iris images. WT analyzes the signal or image at dif ferent frequenc y bands with dif ferent resolutions by decomposing it into approximation and detail coef ficients. The decomposition of the signal into dif ferent frequenc y bands is obtained simply by successi v e high pass and lo w pass filtering of the time domain signal. WT decomposes an image into four sub- images or sub-bands such as approximation coef ficients (LL), horizontal coef ficients (LH), v ertical coef ficients (HL) and diagonal coef ficients (HH). One or more of these sub-bands can be split into smaller sub-bands, which can be split ag ain, and so on. Hence, more discriminating features can be e xtracted in a WT domain than a time domain. Ho we v er , WT only displays suf ficient frequenc y resolution at lo w frequencies b ut poor frequenc y resolution at high frequencies. As an e xtension of WT , WPD is de v eloped to achie v e fine frequenc y resoluti on at both lo w and high frequencies. In WPD, each of approximation and detailed sub-bands are furthe r processed as opposed to WT where only approximat ion sub-bands ar e further proce ssed. This resul ts in s plitting the whole frequenc y plane into equally sized bands. Hence, WPD enables us to zoom into desired frequenc y channels for further decomposition and yield a better representation of signals [9]. F or this reason, WPD is more suitable to e xtract local patterns of each iris at dif ferent resolution le v els, which contains the main di v ersities of dif ferent irises. In this w ork, Haar WT w as carried out up to 3-le v el on the enhanced iris images after the no r malization step. Haar is the simplest orthogonal w a v elet system, compact support in time, has 1 v anishing moment. It pro vides a simple and computationally ef ficient approach for analysing the local aspects of a signal, defined by (1) and (2),   ( x ) = 8 > > > > < > > > > : 1 ; if 0 x < 1 2 ; 1 ; if 1 2 x < 1 ; 0 ; other w ise (1) and Haar scaling function computes a v erage or approximation. ( x ) = ( 1 ; if 0 x < 1 ; 0 ; other w ise (2) Figure 2 illustrates the WPD structure after 3-le v el WPD. W ith the le v els computed from top to bot- tom, time resolution decreases, whereas frequenc y resolution increases. A quadtree with 64 output sub-images is generated. The sub-images are referred as pack ets or nodes t hat ha v e coef ficients of approximation (A), horizontal detail (H), v ertical detail (V) and diagonal detail (D). Figure 2. WPD structure for 3-le v el decomposition Multilinear principal component analysis for iris biometric system (Chetana Kamlaskar) Evaluation Warning : The document was created with Spire.PDF for Python.
1462 r ISSN: 2502-4752 3.3. F eatur e subset selection and v ector cr eation In this paper , tw o approaches are proposed to select the optimal set of features. Approach 1: Ener gy Measure based P ack et Selection. Ener gy Measure (E): The ener gy-based crite- rion is used to choose useful sub-images for feature encoding as w a v elet maxima ener gy points are capable of detecting sharp v ariation points, and of formulating a signal the presentation that is well adapted for charac- terizing patterns. Ener gy distrib ution for an iris image f(x,y) with 1 <x<M , and 1 <y <N can be calculated using w a v elet pack ets [18] and ener gy measure using (3), where M is number of ro ws and N is number of columns of the enhanced normalized iris image. E = 1 M N M X x =1 N X y =1 j f ( x; y ) j 2 (3) Figure 3 sho ws the a v erage ener gy distrib ution of 244 dif ferent iris images with Haar WPT . It consis ts of total 64 pack ets ranging from (3,0)(corresponding to node 21) to (3,63)(corresponding to node 84) at 3-le v el decomposition. It is observ ed that if the image has distinct features with some frequenc y and direc tion, the corresponding sub-images or pack ets ha v e lar ger ener gies in w a v elet transform. Number of Packets or subimages at 3-level of WPT 0 10 20 30 40 50 60 70 Mean Energy 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 Mean Energy Distributionof subimages at 3-level WPT subimages 22,23,29,31 Figure 3. A v erage ener gy distrib ution of each pack ets From Figure 3, appropriate dominant ener gies are chosen to compute iris code. The node 21 corre- sponds to pack et number (3,0) has of fset (DC) information hence, not considered. The subimages at node 22, 23, 29, and 31 retain much higher ener gy than other sub-images so the y are chosen as candidates or samples for encoding. During the e xperiment, a combination of appropriate pack ets is used for the selection of sub-images, and their coef ficients are used to represent the feature v ector . Encoding WPT coef ficients: By applying soft threshold (T), an iris feature v ector referred as Iriscode is achie v ed by quantizing the coef ficients into one data element as (4), F ij = 8 > < > : 1 ; if C ij > T ; 1 ; if C ij < T ; 0 ; other w ise (4) where C ij the coef ficient of subimage, T is a soft threshold and F ij is encoded coef ficients of that subimage. In this e xperiment, T = 3 , which is more practical in engineering applications [9]. Here, is the standard de viation of the highest frequenc y sub-image coef ficients, pack et number(3, 63), that is, node 84. F or enhanced normalized im age size of 50x270, after 3-le v el w a v elet pack ets decomposit ion, the size of subimage at le v el 3 is 7x34 pix els. So e v ery single subimage generates a code of length 238. If the combination of N subimages is used then it w ould generate a code of length Nx238. Indonesian J Elec Eng & Comp Sci, V ol. 23, No. 3, September 2021 : 1458 – 1469 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1463 Approach 2: MPCA. After e xtracting w a v elet coef ficients at 3-le v el decomposition of the enhanced normalized iris image, we aim to fuse all of the features ef fecti v ely to classify iris images. In this w ork, MPCA, a tensor -based multi-linear subspace learning method is proposed to perform the multiple-feature fusion. It can ef fecti v ely combine and select all of the features e xtracted from the original image and consider the interre- lationship among dif ferent w a v elet pack et coef ficients, without reshaping tens ors into v ectors. A concept of ‘tensor’ is introduced to arrange all of the features of one normalized iris image. This w ork is moti v ated by mul- tilinear feature e xtraction methods presented in [3], the MPCA. Here, we propose a no v el tensor based feature fusion approach using MPCA to select and combine e xtracted iris features after w a v elet pack et decomposition. 3.3.1. T ensor notations and r epr esentation A tensor is a N-w ay array or a multidimensional array [14], and the order of a tensor is the number of dimensions, also kno wn as w ays or modes. In this paper , we denote scalars by lo wer -case le tters ( x; y ; : : : ) , v ectors (one-w ay array) by boldf ace letters ( x ; y ; : : : ) , matrices (tw o-w ay array) by boldf ace capital letters ( X ; Y ; : : : ) , and tensors of order three or higher(three-w ay or higher array) by calligraphic capital letters ( X ; Y ; : : : ) . An N th order tensor is denoted as, X 2 R I 1 x I 2 ::: x I N . A tensor of N th order contains N indices i n , where n = 1 ; : : : ; N , and each of which corresponds to the n-mode of X . F or e xample, in this study , size of the normalized iris is 50x270 pix els after 3-le v el w a v elet pack ets decomposition, the size of subimage or pack et at le v el 3 is 7x34 pix els. T otal 63 pack ets e xcluding the first one (DC component) are used together for feature fusion. So, the normalized te xture can be represented as 7x34x63 three-dimensional tensor objects (3 r d order tensor) with column, ro w , and number of w a v elet pa ck ets respecti v ely . The n-mode product of a tensor X 2 R I 1 x I 2 ::: x I N with a matrix U 2 R J n x I n denoted as X x n U , is a tensor with entries [14]. ( X x n U )( i 1 ; : : : ; i n 1 ; j n ; i n +1 ; : : : ; i N ) = P i n X ( i 1 ; : : : ; i n ) : U ( j n ; i n ) (5) 3.3.2. MPCA In pattern recognition, a tensor object is usually defined in high dimensional tensor space. Recog- nition methods serving directly on this space ha v e the curse of dimensionality problem and m an y classifiers beha v e inadequately gi v en a small number of training samples. Further , handling high dimensional samples are computationally costly . T o deal with this and learn features, directly from tensor without reshaping tensors into v ectors, researchers ha v e attempted multilinear subspace learning [19]. The PCA is mainly used to reduce data dimension and retain information that characterizes the v ariation of data as much as possible. Ho we v er , con v entional PCA w as originally proposed to process 1-D v ectors, which requires all input data to be con v erted into 1-D v ectors before analysis. Thi s unfolding process breaks the natural structure of the input data and loses compact or v aluable representations in the original form [14]. PCA also suf fers from a small sample prob- lem when the dimension of the unfol d e d data is much lar ger than the number of samples. T o o v ercome these dif ficulties, MPCA is proposed by e xtending the con v entional linear PCA based on multilinear algebra. The proposed MPCA is a multilinear algorithm performing dimensionality reduction in all tensor modes seeking those bases in each m o de that allo w projected tensors to capture most of the v ariation present in the original tensors. The core of the MPCA algorithm [3] is the eigen-decomposition in each mode so the distrib ution of the eigen v alues is e xpected to impact significantly on the performance of the algorithm. MPCA has been introduced in detail in [3]. Let, a set of M tensor objects fX 1 ; X 2 ; : : : ; X M g is a v ailable for training with zero mean. Each tensor sample X m 2 R I 1 x I 2 ::: x I N assumes v alues in a tensor space R I 1 R I 2 R I N , which is the tensor product of N v ector spaces R I 1 ; R I 2 ; : : : ; R I N . The MPCA objecti v e is the determination of N projection matrices f U ( n ) 2 R I n x P n ; n = 1 ; : : : ; N g to map original tensor s et fX m 2 R I 1 x I 2 ::: x I N g M m =1 into a tensor subspace fY m 2 R P 1 x P 2 ::: x P N g M m =1 with P n I n ; n = 1 ; : : : ; N the dimensionality of the pro- jected space is much lo wer than the original tensor space. Mathematically [3, 14], Y m = X m x 1 U (1) T x 2 U (2) T : : : x N U ( N ) T 2 R P 1 x P 2 ::: x P N ; m = 1 ; : : : ; M (6) The feature tensor after projection is obtained as fY m g which captures most of the v ariation in original tensor set, and v ariations are measured by the total tensor scatter , and U ( n ) is the mode-n projection matrix. Ho we v er , it is hard to obtain al l N projecti on matrices simult aneously . F or this purpose, the alternating least square (ALS) algorithm can be used to solv e the optimization of projection matrices approximately [3]. Thus, MPCA formulation is the estimation of the N projection matrices that maximize the total tensor scatter T y as (7), Multilinear principal component analysis for iris biometric system (Chetana Kamlaskar) Evaluation Warning : The document was created with Spire.PDF for Python.
1464 r ISSN: 2502-4752 f U ( n ) 2 R I n x P n g N n =1 = arg max f U ( n ) g T y (7) where T y = M P m =1 kY m Y k 2 F , Y denotes the mean of projected tensor feature calculated as Y = 1 M M P m =1 Y m . During testing, a test tensor sample X is first centered by subtracting the mean X obtained from the training data and then projected using (8) to the MPCA features. F or the implementation of MPCA algorithm, we referred to the code a v ailable at MathW orks. Y = ( X X ) x 1 U (1) T x 2 U (2) T : : : x N U ( N ) T (8) 3.4. Classification or matcher stage After the discriminant features are e xtracted, the final step of iris recognition is to design a rob ust matcher . The function of this step is to measure ho w similar or dif ferent templates are and to decide whether the y belong to the same person or not. In the training phase, iris code is generated using e xtracted features for each iris image and stored as a template in the g allery . During the testing phase, iris code of the query sample is compared using dif ferent distance measures. The result of this computation is then used as the score of match, with smaller v alues indicating better matches. F or Approach 1: The most popular distance measures are Euclidean, Manhattan and Cosine distance. Euclidean and Manhattan distance ignore the correlation which is important for measuring the similarity of tw o v ectors while Cosine measures the correlation b ut ignores the distance between tw o v ectors. T o tak e into account, both, the distance and the correlations between tw o v ectors, triangle square ratio similarity measure [20] is used, defined as (9), T S R ( a; b ) = k a b k 2 k a k 2 + k b k 2 = 1 2 k a kk b k k a k 2 + k b k 2 cos (9) if a and b are unit v ectors, then the triangle square ratio is equi v alent to Euclidean distance, while k a k = k b k , it is equi v alent to the cosine criterion [20]. F or Approach 2: The performance of tensor based feature fusion is measured usi ng a Modified angle distance (MAD) similarity measure [3]. It is weighted v ersions of the cosine angle, defined as (10), M AD ( a; b ) = N P i =1 a i :b i =w i N P i =1 a 2 i N P i =1 b 2 i (10) where N is the length of feature v ector and w is weight v ector . 4. RESUL TS AND DISCUSSION 4.1. Experiment on SDUMLA-HMT multimodal database The e xperiment is performed on the SDUMLA-HMT Multimodal Database from the Group of Ma- chine Learning and Applications, Shandong Uni v ersity [21]. This multimodal data set is a comprehensi v e collection of fi v e biometric modalities such as f ace, finger v ein, g ait, iris and fingerprint of the same subject and it consists of a total 106 subjects. W e ha v e chosen iris modality for testing the proposed algorithms. The details of iris modality are mentioned in T able 1 and representati v e images are sho wn in Figure 4. Figure 4. Representati v e iris images from SDUMLA database Indonesian J Elec Eng & Comp Sci, V ol. 23, No. 3, September 2021 : 1458 – 1469 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1465 T able 1. Iris SDUMLA-HMT database Image Resolution (Iris x impressions per Iris) T otal Images F ormat 768x576 Left Iris 106x5 = 530 1060 256 gray le v el BMP Right Iris 106x5 = 530 4.2. Pr oposed implementation In the pre-processing stage, the iris needs to be isolated from an e ye image and then normalized to the same size. Prior to this, video-based automat ic system for iris recognition (V ASIR) [22] is used to analyze the iris image quality . From the quality analysis, it is observ ed that iris images ha v e v ery lo w contrast between sclera and iris, and hence f ail to se gment iris correctly . So, iris image contrast enhancement is performed using ’imadjust’ and log transformation. Iris i mages are se gmented and then normalized using the method proposed in [23]. That means, the iris re gion is mapped into fix ed dimensions of 50(r) x 270 ( ) of polar image coordinates. The normalized image has lo w contrast , so adapti v e histogram equalization is performed on the image to adjust the contrast. In the feature e xtraction stage, 3-le v el w a v elet-pack et decomposition is used to e xtract WP coef ficients as iris features. In e xperiment 1, after 3-le v el Haar WPD of normalized iris, the size of subimage at le v el 3 is 7x34 pix els. So, e v ery single subimage generates an iris code of length 238. Based on the Ener gy measure criterion, N numbers of subimages are selected. The resultant feature v ector is the concatenation of quantized coef ficients of selected subimages. If a combination of N subimages is used then it generates iris code of length Nx238. The performance of dif ferent combinations of subimages is sho wn in T able 2. In the matching stage, NN classifier with a triangle square ratio is used. In e xperiment 2, a tensor model is b uilt in the feature e xtraction process. T o formulate a tensor repre- sentation of each enhanced normalized image of size 50x270, after 3-le v el Haar WPD, all 63 subimages (each of size 7x34) e xcluding the first one (DC component) are arranged as a three w ay tensor as X 2 R 7x34x63 . Here, MPCA is introduced, which is a tensor based dimensionality reduction algorithm. Then a no v el MPCA- MAD matcher is constructed which tak es adv antage of tensor feature e xtraction and a t the same time solv es the problem of formation of more discriminati v e single feature v ector by fusing 3-le v el WPD subimages coef- ficients. 4.3. P erf ormance e v aluation The performance of the system is reported using R OC curv e and EER. The R OC curv e is the per - centage of genuine attempts accepted (i.e. 1-f alse rejection rate (FRR)) on the y-axis, ag ainst the percentage of impostor attempts accepted (i.e. f alse acceptance rate (F AR)) on the x-axis. The R OC curv e measures the accu- rac y of the matching process for dif ferent threshold v al ues and sho ws the o v erall performance of the designed system. F AR is referred as the probability that impost er sample is tak en as genuine while FRR is referred as the probability that a genuine sample is tak en as an imposter one. On the R OC curv e, the point where F AR is equal to FRR is referred as EER. The obtained EER v alue is used to e v aluate the recognition accurac y in our e xperiments. In biometrics, the lo wer v alue of EER sho ws the better recognition performance of the system. In this w ork, out of 106 subjects, only 61 subjects are selected based on correct se gmentation of iris. W e use first 4 images per subject in the training set (61 Classes x 4 impressions per Class) and the remaining for testing. In our e xperiments, match scores are calculated by comparing e v ery single image with the remaining images. The match scores are di vided into inter -class and intra-class matching to justify the ef fecti v eness of the proposed method. The total number of intra-class comparisons is 366 and that of inter -class comparisons is 29280 . T able 2 sho ws that by selecti ng the appropriate combination of subimages to encode iris features, the performance of the recognition system can be impro v ed. The combination results in increasing the length of bit code and pro vides more te xture information from dif ferent subimages. If a subimage ha ving high-frequenc y noisy ener gy gets combined, it results in de grading the performance. The best EER performance of 2.2985% is obtained for iris code length of 952 using subimages 22 ; 23 ; 29 ; and 31 of Haar WPT . The R OC curv es and EER for v arious bit lengths of iriscode using Haar WPT are sho wn in Figure 5. Multilinear principal component analysis for iris biometric system (Chetana Kamlaskar) Evaluation Warning : The document was created with Spire.PDF for Python.
1466 r ISSN: 2502-4752 T able 2. EER performance for dif ferent combinations of subimages of HAAR WPT Subimages Feature Genuine (G) Imposter (I) Matcher EER% Encoded (Code length) T rials T rials (Similarity Measure) 22 ; 23 476 3 66 29280 T riangle 2.6981 22 ; 31 476 square 5.6660 23 ; 31 476 ratio 2.4556 22 ; 23 ; 31 714 2.4624 22 ; 23 ; 29 714 3.1523 23 ; 29 ; 31 714 2.7322 22 ; 23 ; 29 ; 31 952 2.2985 [1]T raining Images: No. of Class (N) = 61 and Images per Class (t) = 4 [2] Genuine T rials (G)= N t ( t 1) = 2 = 366 , [3] Imposter T rials (I)= N ( N 1) t t= 2 = 29280 False Accept Rate(%) 10 -3 10 -2 10 -1 10 0 10 1 10 2 Genuine Accept Rate(%) 10 1 10 2 ROC using combinations of WPT subimages Subimages 22,23,31, EER=2.4624% Subimages 22,31,EER=5.6660% Subimages 23,31,EER=2.4556% Subimages 23,29,31,EER= 2.7322% Subimages 22,23,29,31,EER=2.2985% Figure 5. R OC curv e for Haar WPT The result of the proposed tensor based MPCA feature fusion is sho wn in Figure 6 and T able 3. Here, the performance in terms of EER is tested by v arying the number of eigen v ectors which represents the dimension of iris features. Using MPCA, the best performance or t he lo west EER of 2.2814% is obtained for feature v ector length of 500 and for the remaining feature dimension, EER is slightly high. F or comparison purposes, the performance of the tensor based MPCA approach is also in v estig ated for the same bit length of iris code obtained using approach 1. It is seen that the best performance can be obtained by considering the tensor representation of all 3-le v el subimages of WPT and then performing multilinear subspace learning feature fusion using MPCA. Thus, approach 2 has the potential to fuse w a v elet pack et features of iris into a single feature v ector whi ch is a highly discriminati v e fused feature with fe wer dimensions. Figure 6 sho ws the R OC curv es of MPCA based feature fusion approach. Thus the proposed tensor -based MPCA approach not only brings the ef fect of dimension reduction b ut also significantly outperforms. False Accept Rate(%) 10 -3 10 -2 10 -1 10 0 10 1 10 2 Genuine Accept Rate(%) 10 1 10 2 ROC for tensor based MPCA EER=3.0396% EER=2.4010% EER=3.1216% EER=2.2814% Figure 6. R OC curv e for tensor based MPCA Indonesian J Elec Eng & Comp Sci, V ol. 23, No. 3, September 2021 : 1458 – 1469 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 r 1467 T able 3. Performance analysis for dif ferent feature dimension of MPCA based approach Feature(Code length) Genuine T rials(G) Imposter T rials (I) Matcher(Similarity Measure) EER% 256 366 29280 Modified 3.0191 476 Angle 3.3982 485 Distance 3.0396 500 2.2814 501 2.3497 714 2.4010 952 3.1216 [1] T raining Images: No. of Class (N) = 61 and Images per Class (t) = 4 [2] Genuine T rials (G)= N t ( t 1) = 2 = 366 , [3] Imposter T rials (I)= N ( N 1) t t= 2 = 29280 4.4. Comparison with existing methods The performance comparison of the proposed approaches with the other e xisting approaches for four Iris databases are sho wn in T able 4 along with the feature e xtraction, classification techniques, and their e v alu- ation protocols. No earlier w ork found based on MPCA for Iris recognition for direct performance comparison. From the pre vious w ork, it can be noticed that the deep learning approaches are generally outperformed as the y are able to analyse complicated data quite well. Ho we v er , deep learning-based approaches, while ef fecti v e, are v ery computationally e xpensi v e and time-consuming [24]. Further , the training dataset size pl ays a lar ge role in t he creation of good feature e xtractors [25]. Comparing with earlier w ork based on WT and WPD [26]-[29], our proposed approach sho ws the encouraging performance among typical algorithms. MP CA utilizes fe wer features while si gnificantly impro ving recognition accurac y compared to W a v elet pack et sel ection based on the Ener gy Measure me thod for feature v ector formation. Thus, our method is more po werful in representing the te xture features and reducing the probability of f alse match rate. Here, it is w orth noting that v ery fe w studies used the SDUMLA-HMT Iris database for e v aluating the biometric system. The feature e xtraction algorithm based on MPCA+LD A technique [30] can be e xam ined in the further study of the Iris recognition problem to tak e into consideration t he class relations [25] of the feature sets. Our prototype model ran on PC with 3.10GHz processor and 8GB RAM. The training and testing processing time sho wn in T able 4 sho ws a significant speedup of e x ecution. T able 4. Comparsion of iris recognition approaches with their e v aluation protocols Author Iris Database Method Feature V ector Classification Performance Measure Zhiping et al. [31] CASIA 2D-weighted PCA Normalized Adapti v e ANN CRR(%) = 97.7 size:240x20 Ng et al. [26] CASIA-Iris-V3 Haar WT 348 bits Hamming CRR(%) = 98.45 Rai and Y ada v [27] CASIA-IrisV1 Haar W a v elet 512 SVM Accurac y(%)= 91.33 decomposition Dhange et al. [28] IITD D WT+DCT+BPSO A vg.feature Euclidean CRR(%)= 97.81 Selected 56 Ohmaid et al. [29] UBIRIS D WT 24x216 KNN Accurac y(%)= 95 V ishi et al. [32] SDUMLA-HMT V eriEye 6.5 SDK - - EER= 3.30 % Kamlaskar et al. [33] SDUMLA-HMT 1D Log-Gabor 9600 Hamming EER= 2.59 % Alay et al. [24] SDUMLA-HMT V GG-16 CNN Feature map Fully connected Accurac y(%)= 98.58 model size: 7x7x512 layers: 4096 nodes Proposed SDUMLA-HMT WPD 952 T riangle square ratio EER= 2.298 % Proposed SDUMLA-HMT WPD+MPCA 500 MAD EER= 2.28% (T raining T ime 1.911s and T est time 0.0076s) 5. CONCLUSIONS This paper focuses on the e xtraction and fusion of features of iris modality . Iris features are e xtract ed using the WPD technique. W a v elet pack et coef ficients which are e xtracted at the 3-le v el, composed of approx- imations and details, represent iris features. Here, MPCA is proposed to consider the interrelationship among dif ferent w a v elet pack et coef ficients and generate more discriminating features with compact representation. So, all w a v elet pack ets are arranged in tensor format and performed a feature fusion based on a multilearning subspace learning algorithm. It aims to find transformations that preserv e the multidimensional data structure, search for lo w-dimensional multilinear proj ections, and perform dimensionality reduction ef ficiently . These characteristics mak e MPCA an ef ficient feature fusion tool for pattern recognition. The e xperimental results sho w the ef ficac y of our proposed approach in the fusion of w a v elet pack et feature sets e xtracted from an iris Multilinear principal component analysis for iris biometric system (Chetana Kamlaskar) Evaluation Warning : The document was created with Spire.PDF for Python.