Palmprint Recognition Based on Local Texture Features

Transcription

1 Palmprint Recognition Based on Local Teture Features Slobodan Ribaric, Markan Lopar Universit of Zagreb, Facult of EE and Computing Unska 3, 1 Zagreb, Croatia slobodan.ribaric@fer.hr markan.lopar@fer.hr Abstract. In this paper, we propose and evaluate palmprint recognition method based on local Haralick features. The Haralick features are etracted from overlapping square subimages of a palmprint region of interest (ROI). A biometric template is composed of N m-component feature vectors, where N is the total number of overlapping subimages, and m is the number of local Haralick features per subimage in the ROI. A live biometric template and templates from database are matched in N matching modules. Based on fusion at the matchingscore level, the total similarit measures between a live biometric template and templates from the database are calculated. B using the maimum of total similarit measure and the 1-NN classification rule, the final decision (person identit) is made. The proposed palmprint recognition sstem was tested on the PolU database. The results of open set identification are given. Kewords: Palmprint recognition, local Haralick features, Open set identification 1 Introduction The palmar surface of a human hand is a rich source of phsiological features. Highresolution palmar images (i.e. resolution higher than 5 dpi) offer finger and palmprint recognition based on features such as sweat pores, ridges, singular points and minutia points [l], []. From low-resolution palmar images the features such as principle lines (head, life and heart lines), wrinkles and teture-based features can be etracted from palmprint region. These features are relativel stabile and unique, and can be used for biometric verification and/or identification. In general, there are three main approaches to palmprint feature etraction [1]: principal line-based, appearance-based, and statistical-based approach. The principal line-based approach uses different edge detectors to etract palm lines [3]. The most popular group of feature etraction methods for low-resolution palmar images is the appearance-based methods such as principal component analsis (PCA) [4], [5], linear discriminant analsis (LDA) [6], the Most Discriminant Features (MDF) [7], and independent component analsis (ICA) [8]. Statistical approaches are either global or local. The global statistical approaches compute statistical features from the whole original or transformed palmprint region

2 of interest (ROI). The local statistical approaches divide the original or transformed palmprint ROI into small regions (subimages) and calculate statistical features for each small region. Recentl, more attention has been given to local (statistical) features, such as those etracted from original palmprint images using a Gabor bank of filters [9], and local binar patterns (LBPs) [1]. In this paper, we propose an approach to palmprint recognition based on local Haralick features that are obtained from the gra-level co-occurrence matrices (GLCMs) created on the small overlapping square subimages of the palmprint ROI. Related work In biometrics applications Haralick features [11], [1] are used for fingerprint classification [13], face [14] and iris recognition [15]. As far as we know, beside our paper [16], there are onl four other descriptions of applications of Haralick features for palmprint recognition. In [16] we have described the basic idea of using local Haralick features and results of the series of preliminar palmprint recognition eperiments for different parameters of the gra-level co-occurrence matri (GLCM) for 8 8 subimages. The eperiments were performed on two relativel small databases (55 hand images of 11 people and 134 hand images of 133 people). It was shown that the achieved recognition rate (98.91%) was better then the recognition rate for the same databases for the eigenpalm approach. In [17] the author used the Haralick features obtained from relativel large subimages ( ) along the palmprint principal lines. Verification eperiments were performed on the small part of the PolU database (onl 5 persons) and have shown a poor performance (EER above 14%). Based on such performance, it ma be concluded that the above described approach is not promising. In [18] Zhang et al. have used the image brightness and GLCM based entrop for palmprint aliveness detection. In the contactless palmprint and palm vein recognition sstem [19], the Haralick features (contrast, variance and correlation) are used for image qualit assessment. X. Zhu et al. described palmprint classification based on GLCMs and global Haralick features [], but the palmprints are classified onl into two categories according to traditional Chinese medicine epert knowledge. 3 Local Haralick features A brief overview of the gra-level co-occurrence matri (GLCM) and Haralick features substantiall follows [11], [1]. The GLCM P(g, g', δ, Θ), size G G, contains at the position (g, g') a number of occurrences of a pair of piels that are at a distance δ in the direction Θ, where one piel has a gra-level value g, and another piel has a gra-level value g'. The values of the piels are quantized into G levels. The GLCM has to be normalized in such a wa that each element of the matri is divided b R, where R is the sum of all the elements in the matri. An element of the normalized GLCM is the probabilit of the occurrence of pairs of piels of imposed values and a fied spatial laout. 1

3 Based on the normalized GLCM, Haralick has proposed 14 statistical features (f 1 - f 14 ), that can be calculated for each δ and Θ [11]: energ (f 1 ), contrast (f ), correlation (f 3 ), variance (f 4 ), inverse difference moment (f 5 ), sum average (f 6 ), sum variance (f 7 ), sum entrop (f 8 ), entrop (f 9 ), difference variance (f 1 ), difference entrop (f 11 ), information measures of correlation (f 1 and f 13 ), and maimal correlation coefficient (f 14 ). In our approach, the local Haralick features are calculated based on the GLCMs created on the d d piels subimages; d < D; defined b sliding-windows positioned on the palmprint ROI (size D D). Sliding-windows define overlapping subimages for the translation step t < d. The parameters δ; δ = 1,,...; Θ; Θ =, 45, 9 and 135 ; and G have a strong influence on the values of the Haralick features. For eample, in the literature [1] there is a recommendation for the values δ (for the global Haralick features) to be from 1 to 8 or 1. Of course, the values of δ depend on the size of the image S and the size and area of the basic teture element. Values of δ, for GLCMs for local Haralick features, are additionall limited b dimension d of the subimage. In order to obtain features that are robust to rotation and to reduce dimensionalit of the feature vectors, we have used the following approach: first, the four GLCMs are obtained (for Θ =, 45, 9 and 135 ), and then a new GLCM is calculated as the average of these four matrices. Finall, the Haralick features are obtained based on this new GLCM. A number of quantization gra-levels G of an image S is also an important parameter, because the computational compleit for the GLCM and Haralick features (ecept for f 14 ) is O(G ). At a first glance, the computation of the GLCMs and Haralick features seems a time-consuming process, but the computation can be speeded-up because the matrices are smmetrical, and, in practice, ver often sparse. 4 Description of the sstem The eperimental palmprint recognition sstem consists of the following modules: preprocessing module, module for localization and normalization of ROI, local Haralick feature etraction module, N matching modules, template database, N distance to similarit transform modules, fusion module, 1-NN and decision module. The short descriptions of functions of the above modules subsequentl follow. 4.1 Preprocessing In the preprocessing module some standard image preprocessing procedures are applied on the input hand-images from PolU database: global threshold, contour and relevant points etraction. Based on the contour of the hand and the relevant points on it, the palmprint ROI is automaticall localized. After that, the ROI is cropped from the gra-scale image, rotated to the same position, sized to the fied dimensions (96 96 piels) and then it is light normalized using histogram fitting. The procedures of preprocessing are similar to those described in [1]. Light normalization is described in 11

4 detail in [5]. We would like to stress that preprocessing is not in the focus of the paper and it is onl briefl described. 4. Feature etraction In the feature etraction module, the local Haralick features are etracted from a palmprint ROI as follows. The palmprint ROI represented as a D D piels gralevel image (D = 96, G = 56 gra levels) is divided into a set of N overlapping subimages. In our implementation, we use the square subimages obtained b using a sliding-window approach. A sliding-window of the size d d piels is positioned in the upper-left corner of the ROI. The first subimage is composed of all the piels of a palmprint ROI that fall inside the window. After that, the window is translated b t = d / piels to the right, and so on. When the sliding-window falls outside of the palmprint ROI, it is moved t piels down and all the wa to the left of the ROI. The process is concluded when the bottom-right corner of the sliding-window reaches the bottom-right corner of the palmprint ROI. Each sliding-window position defines one of N regions (subimages). The total number of the subimages is: D d N = + 1 (1) t For a piels palmprint ROI and for a 1 1 piels sliding-window, and sliding-window translation step t = 6 piels, there are N = 5 subimages. Based on ehausting testing, b using our own databases FER I (55 hand images of 11 people, 5 templates per person) and FER II (134 hand images of 133 people, approimatel 1 templates per person) [16]; FER I FER II = ; we have found that: i) three local Haralick features (energ, contrast, correlation) are sufficientl discriminator for palmprint recognition purposes, ii) the size of the sliding-window d = 1, the sliding-window translation step t = 6 and the distances δ from 1 to 6, give satisfactor recognition accurac, iii) the number of gra-levels of the palmprint ROI which is less than G = 18 degrades the recognition accurac. We have selected G = 56 levels. iv) the recognition accurac becomes lower if the resolution of the palmprint ROI is decreasing. The resolution was selected. The three selected local Haralick features are defined as follows: Energ: G 1 G 1 1 =, g= g = f ( p( g g )) () 1

5 Contrast: f G 1 G 1 =, g= g = ( g g ) p( g g ) (3) Correlation: G 1 G =, σ σ g= g = ( gg ) p( g g ) f µ µ (4) G 1 where µ gp ( g), p ( g) = p( g, g ), µ gp ( g), ( g ) = p( g, g ) = = g G 1 g = G 1 G 1 σ = p ( g)( g µ ), and σ = p ( g)( g µ ) g= g= G 1 = = g G 1 g = p, Based on the above selected parameters, a palmprint ROI is represented b N m- component feature vectors, where N = 5 and m = 6 3, where 6 is a number of distances and 3 a number of local Haralick features. Each biometric template is represented b 45, i.e. 5 18, features. 4.3 Matching, fusion at the matching-score level and decision The matching process between the m-component feature vector (m = 18), which is a component of the live template, and the corresponding m-component feature vector - a component of the template from a database, is performed in a matching module using the Euclidean distance. There are N matching modules. Note that the values of the feature vector components are within different dnamic ranges (from 1 - to 1 1 ) and the normalization of the features has to be applied before an calculation of the Euclidean distance. A straightforward technique of linear normalization via the respective estimates of the mean and the variance is used (z-score normalization). The outputs of the matching modules are transformed into the similarit measures s ij, i = 1,,..., N as follows: e d ij e e 1, if d ij d i s e ma ij = d i ma i = 1,,..., N e e, if d ij > d i ma (5) where d e ij is the Euclidean distance between the m-component feature vector obtained from the subimage i from the live template, and the corresponding m-content feature vector from the j-th database template. The distance d e i ma is the maimum distance of the corresponding m-component feature vectors, for subimage i, which is obtained from the training set. 13

6 In the proposed eperimental sstem we used fusion at the matching-score level, so the similarit measures obtained from N matchers are combined. The total similarit measure TSM j between the live template and the j-th database template is: TSM j = N i= 1 w s i ij (6) where w i, i = 1,,..., N, is the weight assigned to the subimage i. The weights w i, i = 1,,..., N, are set proportionall to the results of the recognition rate for each subimage i during the evaluation of the sstem: N i= 1 w = 1 (7) i w i = N recognition _ rate (8) j= 1 recognition _ rate where recognition_rate i is a recognition rate for the subimage i. Note that the weights wi, i = 1,,..., N are calculated from the training set (PolU 1 database; see Section 5.). B using the 1-NN classification rule based on the maimum TSMj, the final decision (person identit) based on a decision threshold is made. i j 5 Eperiments and results For the development, evaluation and testing of the proposed eperimental palmprint recognition sstem we used the database PolU [], which consists of 7751 palmprint images of 386 persons (about images per person). The snta of the palmprint image notation is "PolU L_NN.bmp", where denotes ID of palmprint (1-386), L denotes session (F - first, S - second) and NN denotes the number of the palmprint image (1 - ). The database PolU was divided in three sub-databases: PolU k ; k = 1,, 3. In order to determine the values of the weights w i, i = 1,,..., N, associated with each subimage, we used a training database PolU 1, (41 palmprint images of 1 people; ID = 1 to 1). The following operations are performed: 1) For all the samples in the database PolU 1, all the features are calculated according to parameters selected in advance (d, t, δ, G, a number of local Haralick); ) The separated template (the active sample) is matched with all the remaining templates in the PolU 1. The features obtained from the subimages of the active samples are compared to the features obtained from the corresponding subimages of the samples in the database. For eample, the local Haralick features, etracted from the 14

7 subimage defined b the sliding-window positioned in the upper-left corner of the palmprint ROI for the active sample, are matched with features obtained from the upper-left corner of the palmprint ROI from all the other samples in the PolU 1. 3) The 1-NN classification of the active sample is made for each subimage and the result of the correct classification is recorded. The above procedure is repeated for all samples in the database PolU 1. The values of the weights w i, i = 1,,..., N, are calculated according to the equation (8). 5.1 Open set identification eperiment For the open set identification eperiment we used the PolU database (931 palmprint images of 156 people; about 6 images per person - 3 from F, and 3 from S session; ID = 11 to 76) and the PolU 3 database (4419 palmprint images - 1 palmprint images of 156 users (about 14 images per user; ID 11-76; plus 18 palmprint images of 11 people which have role as imposters; ID ) This set up makes for 1 client eperiments (about ) and 18 impostor eperiments (about 11 ). The open set identification eperiment was repeated times for 6 different client templates (3 from the first session, 3 from the second session) in the template database. The results of the open set identification are represented in terms of FAR (False Accepted Rate), FRR (False Rejection Rate), and FIR (False Identification Rate): FAR [%] = (# of accepted imposters as clients) 1 / (total # of imposters), FRR [%] = (# of rejected clients) 1 / (total # of clients), FIR [%] = (# of wrongl classified clients) 1 / (total # of clients). Table 1. shows the mean value and standard deviation for FAR, FRR and FIR as a function of the threshold T. Table 1. FAR, FRR and FIR as a function of threshold T (for open set identification) Threshold T FAR (mean [%]; std. deviation [%]) FRR (mean [%]; std. deviation [%]) FIR (mean [%]; std. deviation [%]) ;.5 1.3;.335.9; ; ;.354.9; ; ;.34.6; ; ;.334.6; ; ;.335.4; ; ;.34.4; ; ;.383.4; ; ;.393.4; ;.36.5;.44.4; ;.38.19;.41.; ;.6.37;.45.; ;.1.51;.57.;.4 15

8 .896.;..69;.533.;.4 Fig. 1 presents the open set identification test results and shows the dependence of the FAR and the FRR on the threshold value T. From Fig. 1, and Table 1 it is clear that our identification sstem achieves EER (Equal Error Rate; FAR FRR) equal to (1.3 % ±.143% and 1.64 % ±.335 %) for threshold T =.888. For the same threshold value, the false identification rate FIR is.4% ±.6 %. Fig. 1. FAR (T) and FRR(T) for the open set identification test. A minimal total error TER = FAR + FRR is.18 %, where FAR is.13% ±.36 % and FRR is.5 % ±.44 %, is achieved with the threshold value.89. At the same time false identification rate FIR is.4 % ±.3%. Fig. shows the ROC curve. Fig.. The ROC curve 16

9 5. Identification time The total identification time per person consists of a fied time, which is required for the hand-image preprocessing, ROI etraction, normalization, feature etraction, and a variable time for the one-to-man matching process: total_identification_time = fied_time + M (one_to_one_matching_time), where M is the number of all the templates stored in the database during the enrolment process. The fied time is 17 ms and the time for one-to-one matching is.7 ms for a personal computer (one-core processor, working frequenc.4 GHz, 166 MHz FSB). 6 Conclusion We have developed an eperimental palmprint recognition sstem based on local Haralick features. For open set identification, based on PolU database, the following results are achieved: EER (FAR-FRR) is (1.3 % ±.143% and 1.64 % ±.335 %) for threshold T =.888, a minimal total error TER is.18 %, where FAR is.13% ±.36 % and FRR is.5 % ±.44 %, at threshold value T =.89. At the same time false identification rate FIR is.4% ±.3%. Compared with the approach based on the features obtained from a palmprint b means of an LBP operator applied on a Gabor response [3], for the same database, our eperimental sstem achieved better results in terms of the recognition accurac and the computational speed. References 1. A. Kong, D. Zhang, M. Kamel. A surve of palmprint recognition, Pattern Recognition, 4, pp , 9.. D. Maltoni, D. Maio, A. K. Jain, S. Prabhakar. Handbook of fingerprint recognition, Springer-Verlang, New York, D. Zhang. Palmprint Authentication, Kluwer Academic, Norwell, G. Lu, D. Zhang, K. Wang. Palmprint recognition using eigenpalms features, Pattern Recognition Letters, 4, (9), pp , Sept S. Ribaric, I. Fratric. A biometric identification sstem based on eigenpalm and eigenfinger features, IEEE Transactions on Pattern Analsis and Machine Intelligence, vol. 7 no. 11, pp , Nov X. Wu, D. Zhang, K. Wang. Fisherpalms based palmprint recognition, Pattern Recognition Letters, vol. 4, no. 15, pp , Nov D. L. Swets, J. J. Weng. Using discriminant eigenfeatures for image retrieval, IEEE Transactions on Pattern Analsis and Machine Intelligence, vol. 18, no. 8, pp , Aug L. Shang, D.S. Huang, J.X. Du, Z.K. Huang, Palmprint recognition using ICA based on winner-take-all-network and radial basic probabilistic neural network, in: ser. Lecture Notes in Computer Science, 397, Springer, pp. 16-1, 6,. 9. D. Zhang, W.K. Kong, J. You. Online Palmprint identification, IEEE Transactions on Pattern Analsis and Machine Intelligence, vol. 8, no. 9, pp , Sep. 3, 17

10 1. X. Wang, H. Gong, H. Zhang, B. Li, Z. Zhuang. Palmprint identification using boosting local binar pattern, in: Proc. of International Conference on Pattern Recognition, pp , R. M. Haralic. Statistical and structural approaches to teture, Proceedings of the IEEE, vol. 76. no. 5, pp , Ma R. M. Haralic, K. Shanmugam, I. Dinstein. Tetural features for image classification, IEEE Transactions on Sstems, Man and Cbernetics, vol. SMC-3, no. 6, pp , Nov M. Yazdi, K. Ghesari. A New Approach for the Fingerprint Classification Based on Gra-Level Co-Occurrence Matri, International Journal of Computer and Information Science and Engineering ;3, pp , M. F. Augusteijn, T. L. Skufca. Identification of Human Faces through Teture-Based Feature Recognition and Neural Network Technolog, in: Proc. IEEE Int. Conference on Neural Networks, pp , Apr R. Szewczk, P. Jablonski, Z. Kulesza, A. Napieralski, J. Cabestan, M. Moreno. Automatic people identification on the basis of iris pattern etraction features and classification, in: Proc. of the MIEL, pp ,. 16. S. Ribaric, M. Lopar. Palmprint recognition based on local Haralick features, in: Proc. IEEE Melecon 1, , March D. S. Martins, Biometric recognition based on the teture along palmprint principal lines, Faculdade de Engenharia da Universidade do Porto, accessed March 8, D. Zhang, Z. Guo, G. Lu, L. Zhang, Y. Liu, W. Zuo. Online joint palmprint and palmvein verification, Epert Sstem with Applications, doi:1.116/j.eswa , G. K. Ong, T. Connie, A.B.J. Teoh. A Contactless Bimetric Sstem Using Palm Print and Palm Vein Features, Advanced Biometric Technologies, pp , 11.. X. Zhu, D. Liu, Q. Zhang. Research of Thenar Palmprint Classification Based on GLCM and SVM, Journal of Computers, vol. 6, No. 7, pp , B. Jähne, H. Haußecker, P. Geißler. Handbook of Computer Vision and Application, Vol. : Signal Processing and Pattern Recognition, Academic Press, San Diego, PolU Palmprint Database, edu.hk/~biometrics/ 3. L. Shen, Z. Ji, L. Zhang, Z. Guo. Appling LBP Operator to Gabor Response for Palmprint Identification, in: Proc. International Conference on Information Engineering and Computer Science, pp. 1-3, Dec