Inter national J our nal of Electrical and Computer Engineering (IJECE) V ol. 6, No. 6, December 2016, pp. 3255 3261 ISSN: 2088-8708 3255 Combination a Sk eleton Filter and Reduction Dimension of K er nelPCA-Based on P almprint Recognition Muhammad K usban 1 , Adhi Susanto 2 , and Oyas W ah yunggor o 3 1,2,3 Department of Electrical Engineering and Information T echnology , Uni v ersitas Gadjah Mada, Indonesia 1 Department of Electrical Engineering, Uni v eristas Muhammadiyah Surakarta, Indonesia Article Inf o Article history: Recei v ed Jun 29, 2016 Re vised Oct 18, 2016 Accepted No v 4, 2016 K eyw ord: Sk eleton K ernel PCA P almprint recognition Feature information EER ABSTRA CT P almprint identification is part of biometric recognition, which attracted man y re- searchers, especially when fusion with f ace identification that will be applied in the airport to hasten kno wing indi vidual identity . T o accelerate the process of v erifica- tion feature palms, dimens ion reduction method is the dominant technique to e xtract the feature information of palms. The mechanism will boost if the R OI images are processed prior to get normalize image enhancement. In this paper with three sample input database, a k ernel PCA method used as a dimension reduction compared with three others and a sk eleton filter used as a image enhancement method compared with six others. The final results sho w that the proposed method success fully achie v e the tar get in terms of the processing time of 0 : 7415 second, the EER performa nce rate of 0.19 % and the success of v erification process about 99,82 %. Copyright c 2016 Institute of Advanced Engineering and Science . All rights r eserved. Corresponding A uthor: Muhammad K usban Departemen T eknik Elektro dan T eknologi Informasi (DTETI) - UGM & T eknik Elektro Uni v ersitas Muhammadiyah Surakarta muhammadkusban.s3te13@mail.ugm.ac.id Muhammad.K usban@ums.ac.id 1. INTR ODUCTION P almprint recognition is part of the biometric system which attracted man y researchers. One w ay to optimize the v erification and identification is to impro v e the appearance image of palm R OI ( r e gion of inter est ). The trick is to use a proper filter method to g ain a uniform brightness le v el of the image so that the ne xt process to obtain feature information becomes easier . Another trick is by selecting a dimension reduction method in accordance with the pattern of information from the palm so that the ongoing process can distinguish between the original pattern and the f ak e is more accurate and ef ficient in the term s of EER ( equal err or r ate ) and time. There are some researchers who ha v e discussed the use of the R OI filter and dimension reduct ion methods to impro v e the palmprint recognition. The filter function, among others, is to get the sharpness of the image [1] and also to access feature information more ef fecti v ely [2]. Some researchers ha v e used the filter method in the field are the Laplacian, the Gaussian, and the unsharp masking [3]. The method has a dra wback that is by increasing the noise when an acquisition process tak es place. By W ang and Leedham to eliminate the noise is by using a median filter that continued to use a 2D Gaussian la w pass [4]. Zhao et al. using a series of filters, namely: match filter , the W iener , and smoothing filter that the o v erall aims to ele v ate signal-to-noise ratio, remo v e noise, and to impro v e the appearance of the image [5]. Some researchers ha v e used a dimension reduction algorithm to obtain an optimal palmprint recogni- tion. In practice, the feature e xtraction of image tak en directly from a 2D image matrix based instead of the v ectors-based on scatter dif ference criterion. As a result, let to the small sample size (SSS) which will af fect the appearance of the singularity problem. W an has conducted research using two-dimensional gr aph embedding local discrimant analysis (2DLGED A) to o v ercome the singularity LD A [6]. Ho we v er , Xinchun found that the use of principal component analysis (PCA) is the best selection of dimension reduction compared with other J ournal Homepage: http://iaesjournal.com/online/inde x.php/IJECE DOI:  10.11591/ijece.v6i6.11677 Evaluation Warning : The document was created with Spire.PDF for Python.
3256 ISSN: 2088-8708 algorithms [7]. The idea w as reinforced by Imtiaz that for a more rob ust palmprint recognition, the use of PCA done after the process of 2D D WT in adv ance with the aim to get more ef ficient the local v ariables in each se gment of palm [8]. From all of the dimension reduction methods which ha v e been used in the study , none of them has an optimal performance on all side [9][10]. Therefore, in this paper of fered a proposal t o impro v e t he detection system through palm print with K ernel PCA (KPCA) that processed after the process sk eleton filter in all R OI of palms. The KPCA is able to produce an ef ficient discriminant rate [11]. Although the preliminary research, K usban [12] stated that Gabor parameters of 8 5 and the dimension reduction of PCA can g ai n a great achie v ement in v erifying of palms. From the research that ha v e been conducted, the sk eleton filter method for enhancement image and the KPCA for reduction dimension, produces a promising outcome. As a comparison, the sk eleton filter compared with six other filters and the KPCA compared with three other dimension reduction. All simulations run using three kinds input of the database. 2. PR OPOSED METHOD 2.1. Image Enhancement Under normal condition, the entire acquisition image of palms can not be directly benefited to e xtract feature of palm R OI because it contains big unw anted information [2]. T o solv e the problem, image enhance- ment frequently is used to g ain local pattern more clearly , so it helps in strengthening the output rate of feature information [13]. Some e xamples of the use of v arious filters to the R OI image of palms visible in Figure 1. Filter sk eleton has been pre viously used by Lin in his research to obtain all lines minutiae in palm [14]. The result is an ability of sk eleton filter to impro v e the appearance of images, thus impro ving sys tem per - formance and upgrade the point pattern matching approach. Ho we v er , there is the weakness of this method that is increasing the number of feature information meaningless that resulted adds to the comple xity of the com- puting process. T o o v ercome this problem, the authors propose sk eleton filter that only analyze the principal lines of palms by the threshold method which generate outcome as sho wn in Figure 1 (e). (a) (b) (c) (d) (e) (f) (g) Figure 1: A series visualization R OI of palm from se v en dif ferent filter process: (a) Original (b) Anisotropic (c) Multiple (d) Shock (e) W a v elet (f) Sk eleton (d) Histogram Equalization. The notation of the sk eleton S ( G ) from the set G is a point p of S ( G ) and d p is the biggest disk with center in p , then the G in this disc is a ’maximum disk’. The disk d p touches the boundary G in tw o or more IJECE V ol. 6, No. 6, December 2016: 3255 3261 Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 3257 dif ferent places. The sk eleton of G is defined for morphology operation by the erosion and opening function. S ( G ) = K [ k =0 S k ( G ) ; (1) with S k ( G ) = ( G k H ) ( G k H ) H and H is the matrix dilation and erotion or structuring element with ( G k H ) for k sequential erosions of G : ( G k H ) = ( : : : (( G H ) H ) : : : ) H ; k times, and K is the last iterati v e process before G erode to an empty v alue, or K = max f k j ( G k H ) 6 = ;g . In its application, the filter sk eleton is approached by using the principles distance between each point and their boundary . If kno wn the initial x 0 ( m 1 ; n 1 ) = x ( m 2 ; n 2 ) are tw o points equidistant, then the distance transform defined x k ( m 2 ; n 2 ) = x 0 ( m 1 ; n 1 ) + min f x k 1 ( i; j ) : d ( m 1 ; n 1 ; i; j ) 1 g (2) The sk eleton is the set of v alue whose distance from the nearest boundary is has maximum in locally . f ( m 1 ; n 1 ) : x k ( m 2 ; n 2 ) x k ( i; j ) ; d ( m 1 ; n 1 ; i; j ) 1 g (3) 2.2. K er nel PCA The main concern in biometric recognition is the amount of data continued increase significantly . It is wh y the dimension reduction method is absolutely necessary . In an ef fort to di vide the data i n t o smaller , the PCA method is widely used in palmprint recognition [15]. Ho we v er , the technique has dra wbacks dif ficult to ackno wledge the feature information in a single image that ha v e a v ariation orientation. T o o v ercome this problem, it is necessary to use non-linear method and k ernel PCA (KPCA) successfully used in biometrics [11]. If it is kno wn in the PCA apply an association la w C v w = w , with C v is the matrix co v ariance and x is the center data, then: C v = 1 n q X i =1 x i x 0 i (4) under the condition w 2 f x 1 ; : : : ; x n g if 6 = 0 , < x i , w x i , and C v w > i = 1 ; : : : ; q : When the is a high-dimension space, then the KPCA applies the same situations. C v = 1 q P q i =1 ( x i )( x i ) T . F or k ernel matrix K with size q q will ha v e a v alue of association k ( x i ; x j ) = h ( x i ) ; ( x j ) i , with the centering data is as follo w ^ ( x ) = ( x ) s ( x ) = ( x ) 1 q q X i =1 ( x i ) ; (5) thus, the transformation k ernel space is ^ k ( x; z ) = h ^ ; ^ ( z ) i = * ( x ) 1 q q X i =1 ; ^ ( z ) 1 q q X i =1 ( x i ) + = k ( x; z ) 1 q q X i =1 k ( x; x 1 ) 1 q q X i =1 k ( z ; x i ) + 1 q 2 q X i;j =1 k ( x i ; x j ) (6) 3. RESEARCH AND DISCUSSION Research conducted using three database input of palm image, namely: Casia ( C ), IITD India ( I ), and PolyU ( P ). The number of data samples, respecti v ely , are 550, 450, and 650 with each item ha v e a v ariety of image appearance as much as 5, 6, and 10. So that the t otal amount of imagery used is 11.950 palm images that dif ferent from each other . Softw are for simulation is Matlab R2014b under W indo ws 7 Pro. While, the hardw are is PC Intel i7 4500K with 8 GHz of main memory . A Sk eleton F ilter and K ernelPCA On P almprint Reco gnition (Muhammad K usban) Evaluation Warning : The document was created with Spire.PDF for Python.
3258 ISSN: 2088-8708 From the research that has been done for palmprint recognition, a rate of research sho wn in T able 1. F our types reduction dimension, namel y KPCA, KF A, LD A, and PCA are used to get feature information of palm image from three dif ferent databases and se v en distinct filter . The results process from se v enth filter is sho wn in Fig 1. Finally , The result of research are the rate process and the performance of EER v alue from each filters. T able 1: The results of using multiple filters in f o ur dimension reduction based (RD) to obtain a rate of time process and rate of performance (EER) from three types databases: Casia ( C ), IITD India ( D ), and PolyU ( P ) D R Method Casia IITD-India PolyU T ime EER V er . T ime EER V er . T ime EER V er . KF A Original 1,8587 0,4261 0,5746 0,5868 0,3363 0,6656 1,1003 0,3011 0,6991 Anisotropic 1,7318 0,4807 0,5215 0,5742 0,3364 0,6644 1,0355 0,3771 0,6227 Multiple 1,8023 0,4175 0,5823 0,5626 0,3376 0,6622 1,1562 0,3307 0,6700 Shock 1,7188 0,4696 0,5315 0,5709 0,3998 0,6000 1,1946 0,3984 0,6009 W a v elet 1,7338 0,4709 0,5292 0,5935 0,4278 0,5722 1,1024 0,4145 0,5855 Sk eleton 1,7303 0,4446 0,5554 0,5666 0,3174 0,6822 1,1060 0,3502 0,6500 Histogram 1,7513 0,4472 0,5531 0,6010 0,3067 0,6922 1,1536 0,2787 0,7209 KPCA Original 1,3050 0,0249 0,9754 0,4066 0,0189 0,9811 0,7514 0,0036 0,9964 Anisotropic 1,3502 0,2267 0,7739 0,4225 0,5000 0,4611 0,7368 0,5000 0,5755 Multiple 1,3572 0,5000 0,5085 0,4191 0,0200 0,9800 0,7676 0,0027 0,9973 Shock 1,2729 0,0109 0,9892 0,3978 0,0233 0,9767 0,7873 0,0073 0,9927 W a v elet 1,2697 0,0663 0,9339 0,4163 0,0434 0,9567 0,7717 0,0084 0,9918 Sk eleton 1,3039 0,0146 0,9854 0,4073 0,5000 0,4933 0,7415 0,0019 0,9982 Histogram 1,3331 0,0154 0,9846 0,4053 0,0178 0,9822 0,9600 0,0025 0,9973 LD A Original 4,1741 0,0123 0,9877 2,0252 0,0148 0,9856 3,1772 0,0037 0,9964 Anisotropic 4,2425 0,2405 0,7592 2,1251 0,0199 0,9800 3,1990 0,0082 0,9918 Multiple 4,3174 0,0169 0,9831 2,0290 0,0166 0,9833 3,3382 0,0046 0,9955 Shock 4,2774 0,0129 0,9869 2,0957 0,0266 0,9733 3,2568 0,0073 0,9927 W a v elet 4,2314 0,0646 0,9354 2,1914 0,0557 0,9444 3,4504 0,0136 0,9864 Sk eleton 4,2408 0,0093 0,9908 2,0870 0,0167 0,9833 3,3474 0,0036 0,9964 Histogram 4,3522 0,0094 0,9908 2,0972 0,0167 0,9833 3,3670 0,0025 0,9973 PCA Original 3,1740 0,0347 0,9654 1,6684 0,0224 0,9778 2,0115 0,0079 0,9918 Anisotropic 3,2567 0,2732 0,7269 1,4563 0,0256 0,9744 2,1097 0,0112 0,9891 Multiple 3,3113 0,0308 0,9692 1,4778 0,0244 0,9756 2,1430 0,0064 0,9936 Shock 3,3131 0,0324 0,9677 1,4440 0,0256 0,9744 2,2786 0,0091 0,9909 W a v elet 3,1949 0,1033 0,8969 1,4141 0,0656 0,9344 2,0819 0,0199 0,9800 Sk eleton 3,2092 0,0215 0,9785 1,4256 0,0201 0,9800 2,2206 0,0046 0,9955 Histogram 3,2495 0,0223 0,9777 1,4204 0,0210 0,9789 2,1732 0,0046 0,9955 The v alues listed in the T able 1 directly sho w that the sk eleton filter is superior at the time consumed compared others on three dif ferent databases. It is seen that the use of sk eleton contains the same kind balanced line thickness with the color of white in the background. While the other type has lines dif ferent thickness with dark, distinct gradation of the background. In Figure 2 sho w some of the performance curv e R OC (the recei v er operating characteristic), CMC (the cumulati v e match curv e), DET (the detection error tradeof f), and EPC (the e xpected performance curv e). Seen from the display of four curv es, the sk eleton filter with a yello w color representation dominate other types. Although the R OC curv e and the DET in Figure 2 (a) and (c) the sk eleton inferior to shock filter , b ut e v entually the method has best v alue and the statement is reinforced with a vie w EPC curv e sho wn in Figure 2 (d) that the yello w color is the bottom line or has an error rate of the smallest. The ne xt process is the selection method of dimension reduction after all R OI image of palms filtered by the sk eleton method. A KPCA (the k ernel principal component analysis) method has the outstanding per - formance compared with other reduction dimension such as: LD A (linear discriminant analysis ), KF A (k ernel fisher analysis) and PCA (princi pal component analysis) as seen in the Figure 3. The KPCA method with black color lines appear to ha v e the most e xcellent display of curv e performance R OC, CMC, DET , and EPC. Espe- cially for R OC in Figure 3 (a), where the black color lines is in the top position curv e so that the f alse rejection IJECE V ol. 6, No. 6, December 2016: 3255 3261 Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 3259 T able 2: Performance of proposed system and other systems No. T ype EER (%) 1. V arious noise density le v els of palms on phase-dif ference information, [16] 5.8475 2. W a v elet combination on ne w w a v elet based method, [17] 4.0702 3. Fused on SIFT -based Image Alignment, [18] 0.4846 & 0.5078 (left-right) 4. PV -Full-1.0 by FVC-onGoing on Latent P almprint Matching, [19] 5.6 5. EER on Multifeature-Based High-Resolution, [20] 4.8 6. Proposed approach 0.19 rate has the highest v alue or close to 1. While the EPC curv e in Fi gure 3 (d)looks the black color lines to be in the lo west position error rate or highest percentage of v erification. The statement reinforced by e vidence from T able 1 that the use of KPCA method e v en rob ust since data input increasing in number ranging from small to lar ger that is IITD India, Casia, and then PolyU. F rom the output research, it yields better performance in terms of error equal rate (EER) when compared with other similar studies as sho wn in T able 2. False Accept Rate 1 0 6 1 0 4 1 0 2 1 0 0 False Rejection Rate 0 0 . 2 0 . 4 0 . 6 0 . 8 1 Original Anisotropic Multiple Shock Skeleton Wavelet Histogram (a) Rank 0 . 6 5 0 . 7 0 . 7 5 0 . 8 0 . 8 5 0 . 9 0 . 9 5 1 Recognition Rate 0 10 0 20 0 30 0 40 0 50 0 Original Anisotropic Multiple Shock Skeleton Wavelet Histogram (b) False Alarm proba bility (in %) 1 0 3 1 0 2 1 0 1 1 0 0 Miss probability (in %) 0 0 . 2 0 . 4 0 . 6 0 . 8 1 Original Anisotropic Multiple Shock Skeleton Wavelet Histogram (c) Alpha 1 0 1 1 0 0 Error rate 0 . 0 2 0 . 0 4 0 . 0 6 0 . 0 8 0 . 1 0 . 1 2 0 . 1 4 Original Anisotropic Multiple Shock Skeleton Wavelet Histogram (d) Figure 2: Curv es are used to sho w the achie v ement of v arious the filter method in R OI image including the type of performance to: (a) R OC (b) CMC (c) DET (d) EPC. 4. CONCLUSION Globally , the sk eleton method has the best performance compared to other filters: original, anisot ropic, multiple, shock, w a v elet, and histogram with the highest v alue is 99,82 % in successfully v erification process with error rate about 0,19 % in PolyU database and KPCA-based. Ov erall the KPCA is the most suitable method of dimension reduction to obtain the feature from v erification and identification of palms. A Sk eleton F ilter and K ernelPCA On P almprint Reco gnition (Muhammad K usban) Evaluation Warning : The document was created with Spire.PDF for Python.
3260 ISSN: 2088-8708 False Accept Rate 1 0 6 1 0 4 1 0 2 1 0 0 False Rejection Rate 0 0 . 2 0 . 4 0 . 6 0 . 8 1 KFA KPCA LDA PCA (a) Rank 1 0 3 1 0 2 1 0 1 1 0 0 Recognition Rate 0 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0 KFA KPCA LDA PCA (b) False Alarm probability (in %) 1 0 3 1 0 2 1 0 1 1 0 0 Miss probability (in %) 0 0 . 2 0 . 4 0 . 6 0 . 8 1 KFA KPCA LDA PCA (c) Alpha 1 0 1 1 0 0 Error rate 0 0 . 2 0 . 4 0 . 6 0 . 8 1 KFA KPCA LDA PCA (d) Figure 3: Curv es are used to sho w the achie v ement of v arious the dimension reduction method in palmprint recognition including the type of performance to: (a) R OC (b) CMC (c) DET (d) EPC. REFERENCES [1] Goh Kah Ong Michael, T ee Connie, and Andre w Beng J in T eoh. A contactless biometric syste m using multiple hand features. J . V is. Comun. Ima g e Repr esent , 23:1068–1084, 2012. [2] C. Lakshmi Deepika et al. P almprint authentication using modified le gendre moments. Pr ocedia Computer Science , pages 164–172, 2010. [3] O. Nibouche and J. Jiang. P almprint matching using feature points and svd f actorisation. Digital Signal Pr ocessing: A Re vie w J ournal , 23:1154–1162, 2013. [4] L. W ang et al. Infrared imaging of hand v ein patterns for biometric purposes. IET Computer V ision , 1:113–122, 2007. [5] S. Zhao et al. E xtracting hand v ein patterns from lo w-quality images: A ne w biometric technique using lo w-cost de vices. Ima g e and Gr aphics, 2007. ICIG 2007. F ourth International Confer ence on , 2007. [6] W an et al. Fuzzy tw o-dimensional local graph embedding discriminant analysis (f 2dlgeda) with its application to f ace and palm biometrics. Neur al Computing and Applications , 23:201–207, 2013. [7] Xinchun et al. P almprint recognition based on curv elet transform decision fusion. Pr ocedia Engineering , 23:303–309, 2011. [8] Imtiaz et al. A w a v elet-based dominant feature e xtraction algorithm for palm-print recognition. Digital Signal Pr o- cessing: A Re vie w J ournal , 23:244–258, 2013. [9] Y ong Xu et al. A feature e xtraction method for use with bimodal biometrics. P attern Reco gnition , 43:1106–1115, 2010. [10] Milad Jaf ari Barani. et al., ”impl ementation of g abor filters combined with binary features for gender recognition”. International J ournal of Electr onical and Computer Engineering (IJECE) , 4:108–115, 2013. [11] M. Ekinci and M. A ykut. Gabor -based k ernel pca for palmprint recognition. Electr onics Letter s , 43:1077–1079, 2007. [12] M. K usban. V erifikasi dan identifikasi telapak tang an deng an k ernel g abor . J urnal Nasional T eknik Elektr o dan T eknolo gi Informasi (JNTETI) , 4:2, Mei 2015. [13] W ang et al. Discriminati v e and generati v e v ocab ulary tree: W ith application to v ein image authentication and recog- nition. Ima g e and V ision Computing , 34:51–62, 2015. IJECE V ol. 6, No. 6, December 2016: 3255 3261 Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE ISSN: 2088-8708 3261 [14] J. Li and G. Shi. A no v el palmprint feature processing method based on sk eleton image. Signal Ima g e T ec hnolo gy and Internet Based Systems, 2008. SITIS ’08. IEEE International Confer ence on , pages 221–228, 2008. [15] Sara Zokaee and Karim F aez. Human identification based on electrocardiogram and palmprint. International J ournal of Electr onical and Computer Engineering (IJECE) , 2:261–266, 2012. [16] G. S. Badrinath and Phalguni Gupta. P almprint based recognition system using phase-dif ference information. Futur e Gener ation Computer Systems , 28(1):287–305, 2012. [17] Naresh Bhuma and C H Himabin. N EW P ALM PRINT A UTHENTICA TION SYSTEM BY USING W A VELET B ASED METHOD. 2(1), 2011. [18] W ei Bu, Qiushi Zhao, and Xia ngqian W u. SIFT -based Image Alignment for Contactless P almprint V erification. pages 1–6, 2013. [19] J. Dai and J. Zhou. Multifeature-based high-resolution palmprint recognition. IEEE T r ansactions on P attern Analysis and Mac hine Intellig ence , 33(5):945–957, May 2011. [20] E. Liu, A. K. Jain, and J. T ian. A coarse to fine minutiae-based latent palmprint matching. IEEE T r ansactions on P attern Analysis and Mac hine Intellig ence , 35(10):2307–2322, Oct 2013. A Sk eleton F ilter and K ernelPCA On P almprint Reco gnition (Muhammad K usban) Evaluation Warning : The document was created with Spire.PDF for Python.