Signature Recognition by Pixel Variance Analysis Using Multiple Morphological Dilations

Transcription

1 Signature Recognition by Pixel Variance Analysis Using Multiple Morphological Dilations H B Kekre 1, Department of Computer Engineering, V A Bharadi 2, Department of Electronics and Telecommunication** A A Ambardekar 3, Department of Computer Science ** Thadomal Shahani Engineering College, Bandra(E), Mumbai-51 1 Thakur College of Engineering & Technology, Kandivali (E), Mumbai University of Nevada, Reno, USA 3 hbkekre@yahoo.com 1, vinu_bharadi@rediffmail.com 2, amol_ambardekar@rediffmail.com 3 Abstract: In this paper we propose a technique that can be used for signature recognition. This is a contour based morphological technique. In this technique the contour of signature is generated by dilating the signature by various levels which generates a different region corresponding to pixel deviation. This template is use to analyze variation in signature segments. We calculate the variation using set operations on the test signature & dilated template. The FAR & FRR of system is a evaluated and we discuss the feasibility of the proposed technique. Keywords Signature Recognition, Morphology, Template Matching. S I. INTRODUCTION IGNATURE verification is an important research area in the field of authentication of a person as well as documents [1][4][5] in e-commerce and banking. We can generally distinguish between two different categories of signature verification systems: online, for which the signature signal is captured during the writing process, thus making the dynamic information available, and offline for which the signature is captured once the writing process is over and thus, only a static image is available. In this paper we deal with an Off-line signature verification System. We design a system capable of verifying authenticity of a signature based on test performed with genuine signatures (Verification Mode) and person identification from signature (Recognition Mode). Various approaches are possible for signature recognition with a lot of scope of research. In this paper we deal with an Off-line signature recognition technique, where the signature is captured and presented to the user in the format of image only. We use various image processing techniques to extract the parameters of signatures and verify the signature based on these parameters. Signature recognition is a two class pattern classification problem, where authentic signatures belong to one class and the forged signatures belong to the other class. ** - A part of this work is submitted to Mumbai University as a B.E. Electronics Degree project work by these authors and further developed and detailed mathematical analysis is presented here for publication. We use various signature feature points, along with the conventional global features, In this paper we propose a technique based on contour generation of signature by multiple morphological dilation operation. II. CONTOUR GENERATION The scanned signature is first pre-processed to get a normalized binary template. We follow a series of operations like noise removal using filtering, Scaling, Smoothening, Intensity normalization, Thinning [1][9][10]. This gives us a binary signature template. The scanned signature and its normalized template is shown in Fig.1. Fig. 1. Normalized Signature template We use the normalized template for further processing. The signature has intra-class variations i.e. Signatures of a same user have variations, but these variations are limited. Forged signatures (Simple forgeries) and different user s signatures have vast variations (Inter Class variations). We try to detect these variations in signature segments. The signature pixels are having specific limit of variation for the genuine set of signature and the forgeries have variations greater than the limit of a specific person s signature. We try to classify the signature based on this variation. To quantify this variance we have used the proposed morphological [4] technique. We generate a contour of signature; this contour is actually the external boundary of the signature. Dilation algorithm is used for this and various levels of dilations are used. We use morphological dilation process [4][5] with three different structuring elements [5]; the structuring elements are circles with radius r1, r2, r3, r4 these radii correspond to the allowed pixel variance. This process gives four dilated sets D1, D2, D3, and D4 as shown below, Di S B i (1)

2 Where S=Signature template Bi = structuring element- Circle of Radius r i. i=1, 2, 3. This is achieved in programming environment by drawing circles of various radii on the templates and filling them with appropriate colour; the circles are drawn with radii r1, r2, r3, r4. Where r4 > r3 > r2 > r1. This operation gives a structure with bands of varying thickness. These bands of colours will represent the variation extent of each pixel and hence the signature segments. This structure is shown in Fig 2.The four bands in the testing program were generated with r1=3 Pixel, r2=6 Pixel, r3=10 Pixel and r4=16 Pixel radius and filled with Black, Red, Green, Blue Colours Respectively. The lowest variation band is C 00, We find the element count for each set for Cij if i j then it is cross intersection and corresponding to the variation pixel. The values of i, j correspond to different degree of variation. We count the number of pixel in each band. For genuine signatures with least variation the element count of Cij where i=j is very high and C 00 contain the perfect of least variation pixels. Higher is the element count higher is the matching. If the element content of cross intersection is high i.e. Cij, where i j then the pixels are having more variation and correspond to mismatch. Each set is given a weight Wij corresponding to degree of match, we evaluate the matching score S as follows, Wij (i=j) > Wij (i j). i 4 j 4 S NCij * Wij (4) i 1 j 1 Values of Wij are based on intra class variation and evaluated by correlation of same person s signature and evaluating the variation or we can set static values by trial & error process. The higher the value of S higher is the matching between the template. We can set a threshold for S to classify the signatures. Fig. 2. Check Pattern for standard Signature shown in Fig1. The colours are represented in (R, G, B) format where Red =(255,0,0), Green=(0,255,0), Blue=(0, 0, 255) and white=(255,255,255), Black=(0, 0, 0). We call this bands pattern as a Check pattern. In the next section we discuss detection process. III. PIXEL VARIANCE ANALYSIS For training purpose we take three standard signatures from a user and generate dilation sets Di and same band structure for them. For detection purpose we use two templates at a time, one is the test signature and one from the standard templates. Similar check pattern is generated for the test signature. The test template T will be used to generate the dilation sets we get four test dilation sets as discussed above with the help of structuring elements TDi T B i (2) TDi= test dilation pattern, Bi = structuring element- Circle of Radius r i. i=1, 2, 3, 4. Now consider each of this dilation set as an allowable pixel variation area, we have the lowest dilation band D1 as the 3 pixel dilated signature template as the allowed variation and if any pixel is going outside this region will be a misplaced pixel of high variation pixel. If the pixel from test signature are going in the higher dilation bands of original signature Di then they have variation more than the specified radius of the band, we can count the number of pixels from test signature which are having such variation by set intersection process. We take the intersection of dilation sets and try to find the cross occurrence [7][2]. This is given by Cij Di TD j. i,j=1, 2,3, 4. (3) IV. PRACTICAL APPROACH We have implemented the algorithm mentioned above in Visual Basic 6.0. To find the intersection of sets we evaluate the bitwise EX-OR of the colour bands generated by the dilation process. This is illustrated as follows, First template sign1(x, y)..x=0,1..n, y=0,1..m (5) For a M Rows X N Columns template. Second template sign2(x, y)..x=0,1..n, y=0,1..m (6) for a M rows X N Columns template. EX-OR operation will generate third template test(x, y) where test(x i, y i ) = sign1(xi, yi).(r,g,b) Θ sign2(xi, yi).(r,g,b) (7) This EX-OR operation will be performed on the (R, G, B) colour triplet of each pixel because of this in the test(x, y) template various colours are generated by the EX-ORing of R G B bands R(255,0,0) Ex-Or G(0,255,0) will yield (255,255,0) If same colour is there it will generate White(0, 0, 0). Black (0, 0, 0) Ex-Ored with Any of the R, G, B bands will give the same colour only i.e. R, G, B only. The R G B bands in sign template are actually the allowed variations for the pixel and if the pixel is in the allowed variations it will generate Either of the pure R, G, B or otherwise any combination of R, G, B. Pure Black pixels are un-deviated pixels (Black (0, 0, 0) Ex-Or Black (0, 0, 0) Will give (0, 0, 0) i.e. Black only). By scanning the check pattern generated we can find out the pixel variance and hence the matching of signature template, The colour codes used are as follows in Table I.

3 Fig. 3. Genuine and test signatures and Generated test pattern. TABLE I COLOUR CODES USED IN PROGRAM Colour R G B 0. Black Red Green Blue Background colour White Colour Colour Colour Test result background We consider a set of signatures for a person as follows, where signatures from database are shown as well as the signature to be checked is also shown in Fig. 3. We perform the dilation based contour analysis and the check pattern generated is shown in the same figure. Next step of the operation is to scan this check pattern and count the number of pixels of each colour, Red (NR), Green (NG), Blue (NB), Black (NBK) and White (NW) respectively. We give these parameters as inputs to a neuro-fuzzy classifier. The fuzzy logic detector has four participation sets Perfect, Good, Acceptable and, for which the output of neural network is given [3]. This is discussed in the next section. V. RESULTS For testing the methodology we collect three genuine signatures as shown in Fig. 3. We also consider two test signatures from same user as shown in Fig. 4(1) & (2). We collect two forged signatures for the same person; they are shown in Fig. 4(3) & (4) as follows. We test all these signatures against the signatures shown in Fig. 3. & evaluate the matching. The test results are summarized in Table II Figure 4 Test Signatures 1,2 are genuine signatures 3,4 are forged signatures TABLE II TEST SIGNATURES MATCHING RESULTS Parameter Sign1 Sign2 Sign3 Sign4 Black Pixel Red Pixel Green Pixel Blue Pixel Missing or extra Pixels(original) Pixels(duplicate) Matching (%) Remarks Acceptable Good The results indicate that Sign1 & Sign2 are matching with genuine signature and they are accepted, Sign3 was rejected due to drastic scale change at initial comparison only, sign4 is rejected because of low matching score. The matching results indicate that the forged signatures are rejected. The morphological dilation analysis is shown in Fig. 6; the signature with higher matching has high content of black pixels in the check pattern which correspond to minimum variance band. For sign3 the dilation analysis is not performed because the morphological parameters have high variation and this is detected before dilation of template. The matching analysis is shown in Fig. 6. Genuine signatures have higher matching. We can select threshold for specifying the acceptance level of signatures. Next we present the study performed to evaluate the False Acceptance Ratio (FAR) and False Rejection Ratio (FRR) for the system developed.

4 % Acceptance Sign Sign2 Sign3 Sign Fig. 5. Pixel Matching Analysis for signature tests. comparison. The first comparison table is as follows. Following table shows performance of individual features and the final system, The final system uses all the features, thresholds calculated from a set of training signatures and a Euclidian Distance based classifier. Recognition Mode Inputs Test Signatures Cases That Should be Cases That Should be TABLE III SIGNATURE RECOGNITION RESULTS / Signatures Cases Actually Cases Falsely Cases Actually Cases Falsely Performance Metrics % 130 TAR FRR TRR FAR Matching(%) Matching(%) FAR-FRR Plot FAR FRR Fig. 6. Signature Matching Analysis. We have used a signature database consisting 500 signatures from 100 different persons. Per person 5 signatures are collected out of which 3 signatures are used for thresholds calculation and record creation. Remaining signatures are used as genuine test signatures. From some arbitrary persons we have collected forged signatures for testing purpose. Total number of signatures used for testing is 610. Out of 610 samples 300 signatures were used for user enrollment, 200 signatures were genuine test signatures, 40 skilled forgeries, 60 casual or unskilled forgeries, 10 signatures were unusable due to distortion. The final system uses user specific thresholds and Euclidian distance based classifier. We have implemented a weighted comparator based classifier in the final system. The metrics FAR & FRR (False Acceptance Ratio & False Rejection Ratio) [4][5][8][10] are evaluated for each feature. In table III we present the test analysis. The entries indicate that out of total 257 tests conducted 244 tests gave correct classification and 13 tests were failed hence the overall accuracy reported is 94.94%. The system developed has 3.7% FRR & FAR are 6.56%. This is shown in FAR-FRR plot in Fig. 7. The system has reported overall accuracy of 94.94%. These values are not fixed but can be modified by changing the decision threshold depending on the requirement. We present a typical set of values for the Fig. 7. FAR-FRR Plot for Signature Recognition System (EER = 6.1%) Next we present the classified results in Table III, we present system performance calculated separately for the Genuine and forged signatures. In the forged signature group we further have group of Casual and Skilled forgery signatures. The system is having 100% Accuracy for the rejection of casual forgery and for skilled forgery the False Acceptance ratio is 5.79%. For the genuine signatures True Acceptance Ratio is 92.77%. All sample of a subject Test Samples Table III Performance Metrics for Final System Forged Genuine Casual Skilled Threshold Ratio Results obtained on the given test bed (%) TAR FRR 7.23 FAR TRR FAR TRR 94.21

5 VI. CONCLUSION Here we have presented an off-line signature recognition system based on the pixel variance analysis by multiple morphological dilations. The system is based on a Euclidian distance based fuzzy classifier. This is a contour based signature recognition technique. Along with various parameters like number of pixels, Angle of rotation, width, height we are using check pattern generated by multiple morphological dilation. We are using check pattern to find out variation in signature pixels. The system developed has reported 94.94% accuracy. The FAR is 6.56% and FRR is 3.7%. The system performance can be improved further by using more training signature and nuro-computing approach. REFERENCES [1] B. Majhi, Y S Reddy, D Prasanna Babu, Novel Features for Off-line Signature verification, International Journal of Computers, Communications & Control Vol. I (2006) [2] Y. M. Y. Hassan,L. J. Karam Morphological Reversible Contour Representation, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 22, No. 3, March 2000 [3] H. Baltzakis, N. Papamarkos, A new signature verification technique based on a two-stage neural network classifier, Engineering Applications of Artificial Intelligence 14 (2001), [4] E. Gose, R. Johnson Baugh, S. Jost, Rafe Pattern Recognition & Image Analysis, PHI [5] A. K. Jain, A. Ross, S. Prabhakar, An Introduction to Biometric Recognition, IEEE Transactions on Circuits and Systems for Video Technology, Vol. 14, No. 1, January 2004 [6] L Bai-ling Zhang, Off-line Signature Recognition and Verification by Kernel Principal Component Self-regression, 5 th International Conference on Machine Learning and Applications (ICMLA'06), 2006 [7] S. Chen and S. Srihari, Use of Exterior Contours and Shape Features in Off-line Signature Verification, Eight International Conference on Document Analysis and Recognition (ICDAR 05) [8] T. Kaewkongka, K. Chamnongthai, B. Thipakom, Off-Line Signature Recognition using parameterized Hough Transform, Proceedings of Fifth International Symposium on Signal Processing and its Applications, ISSPA 99, Australia, August, 1999 [9] D. Kalenova, Personal Authentication Using Signature Recognition, Department of Information Technology, Laboratory of Information Processing, Lappeenranta University of Technology. [10] H. Kekre, S. Pinge, Signature Identification using Neural Networks, Proceedings of National Conference on Image Processing 2005 (NCIP2005), Organized by TSEC, Mumbai, pp 31-39