HANDWRITTEN SIGNATURE VERIFICATIONS USING ADAPTIVE RESONANCE THEORY TYPE-2 (ART-2) NET

Transcription

1 Volume 3, No. 8, August 2012 Journal of Global Research in Comuter Science RESEARCH PAPER Aailable Online at.jgrcs.info HANDWRITTEN SIGNATURE VERIFICATIONS USING ADAPTIVE RESONANCE THEORY TYPE-2 (ART-2) NET Tirtharaj Dash *1, Subhagata Chattoadhyay 2 and Tanistha Nayak 3 *1 Deartment of Comuter Science & Engineering Veer Surendra Sai Uniersity of Technology Burla , India tirtharajnist446@gmail.com 2 Deartment of Comuter Science & Engineering Bankura Unnayari Institute of Engineering Bankura , India subhagatachatterjee@yahoo.com 3 School of Comuting National Institute of Science and Technology Berhamur , India tanisthanist213@gmail.com Abstract Authorizing hand-ritten signature has alays been a challenge to reent illegal transactions, esecially hen the forged and the original signatures are ery similar-looking in nature. In this aer, e aim to automate forged signature erification rocess, offline, using Adatie Resonance Theory tye-2 (ART-2), hich has been imlemented in C language using both sequential and arallel rogramming. The said netork has been trained ith the original signature and tested ith tele ery similar-looking but forged signatures. The mismatch threshold is set as 5% hoeer, it is set flexible as er the requirement from case-to-case. In order to obtain the desired result, the igilance arameter (ρ) and the cluster size (m) has been tuned by carefully conducted arametric studies. The accuracy of the ART-2 net has been comuted as almost 100% ith ρ = 0.97 and m = 20. Keyords- handritten signature automatic erification ART-2 forged signatures INTRODUCTION Neural netorks hae been idely used in attern recognition, esecially here the atterns are comlex due to close resemblance of original and generated atterns [1-4]. An imortant roerty of a neural netork classifier is that, it learns the exemlary atterns (as inuts) by udating its nodal connectors eights. The draback of such tye of learning is that, hen ne atterns are fed, the eights are udated and as a result, it loses the memory of older atterns and stores the imression of ne atterns [5]. To handle this issue, Grossberg and Carenter (1987) roosed the concet of Adatie Resonance theory (ART) netorks, here the netorks retain the earlier learning, hich is certainly adantageous oer the conentional neural classifier [6]. ART is of to tyes i.e. tye-1 and tye-2. ART-1 takes binary inut ector, hereas, ART-2 takes analog/continuous inut ector and therefore more meritorious [7]. In our earlier ork, ART-1 netork has been considered for automatic erification of hand-ritten signature offline, ith high leel of accuracy (99.97%) [8]. In that aer, hoeer, only to forged signatures ere considered. In this aer, ART-2 has been considered for the offline erification of tele ery similar looking but forged handritten signatures. Handritten signature is the rincial biometric measure for ersonal identification. It is an imortant method for erforming legal transactions. Hoeer, there are chances hen signatures could be delicately coied, such that these aarently resemble originals and are difficult to be identified by the naked eyes. Hence, automating such a detection rocess could be of real adantage to us. Hoeer, it requires ast research rior its ractical use. In this ie, this aer is an attemt here ART-2 net has been used and imlemented using both sequential and arallel rogramming techniques to note its detection accuracy and seed. Automatic erification of handritten signature is an age-old research toic. Aailable literatures sho that seeral traditional and soft comuting techniques hae been used for accomlishing the said task. Due to sace constraints, detail discussion of all the techniques are beyond the scoe of this aer. Hence, some releant studies hae been shon, belo. Handritten signature recognition using traditional techniques: Author and year Technique(s) Ag. accuracy Inglis and Witten, 1994[9] Temlate matching 75% Mizukami et al., (2002) [10] Kholmato and Yanikoglu, (2005) [11] Shanker and Rajagoalan, (2007) [12] Dislacement extraction method 75% 3-D ector and DWT 98.6%. modified and basic DWT 98%, 71% Tian et al., 2007 [13] DWT 86.04%. Radhika et al., (2011) [14] Zernike moments 98% JGRCS 2010, All Rights Resered 21

2 Tirtharaj Dash et al, Journal of Global Research in Comuter Science, 3 (8), August 2012, Handritten signature recognition using soft comuting techniques: Author and year Technique Ag. accuracy Hossain et al., 2003 [15] HMM and BPNN hybrid 88.9% L et al., 2005 [16] SVM 95% Aksoy and Mathkour, (2011) [17] Rule-based inductie learning system 97% Dash et al., (2012) [8] ART % Dash et al., (2012) [18] AMN 92.3% the total number of ixels. The aerage SI comuted for the forged signatures is around 51%. Ste-2: Feature extraction: In this ste, e hae extracted the ixel alues in RGB (Red- Green-Blue) format, hich are then conerted into gray scale alues by using the folloing standard relation. Gray alue= 0.33 R G B (2) Ste-3 and 4: Deeloment of ART-2 algorithm and its training: As already mentioned, the ART-2 net has been deeloed on C language. Figure 2 shos the schematic diagram of ART- 2 structure. The imlementation algorithm is as follos: From these studies, it may be noted that ART has not been tested idely in this field, hich leaes an oortunity to inestigate ART-2, hich is the motiation behind this ork. In the folloing section, e hae described the methodology of ART-2 imlementation using C language using both the sequential and arallel rogramming. METHODOLOGY To accomlish the task, folloing stes hae been taken: Ste-1: Acquisition of hand ritten signatures one original and tele forged (refer to Fig.1a and 1b) and comuting the Similarity Index (SI) ixel-ise using equation 1 Ste-2: Feature extraction to obtain all the gray-scale alues ith equation 2 Ste-3: Deeloment of ART-2 net on C language ith (i) sequential and (ii) arallel rogramming Ste-4: Training the net ith the original signature, and finally Ste-5: Testing the net ith the forged signatures and comute the error by estimating the ercentage of mismatch (see equation 3). The stes are discussed in detail in the folloing section. Ste-1: Acquisition of Handritten signatures: Figures 1a and 1b sho the samles of hand-ritten original and one forged signature. Figure 1a: Original signature Figure 1b: One forged signature The size of the signatures is di ith bit deth of 4 bit. Similarity Index (SI) is then comuted beteen all forged signatures and the original signature using equation 1. 1 D SI 100 (1) T The arrangements of ixels ( ON and OFF ) in the forged signatures are comared ith that of the original signature, ro and column-ise. The disarities are then comuted. In equation 1, D is the number of dissimilar ixels and T is Figure-2: Structure of ART-2. Ste 0: Initialize the folloing arameters: a, b, c, d, e, α, ρ, θ. Also secify the number of eochs of training (ne) and number of learning iterations (nit). Ste 1: Perform Stes 2-12 (ne) times. Ste 2: Perform Stes 3-11 for each inut ectors. Ste 3: Udate F 1 unit actiations: U i = 0 i = s i i =0 q i = 0 i =f(x i ) S e s Udate F1 unit actiation again: e i = s i +au i P i = U i q i = i e i = f(x i ) + bf(q i ). Ste 4: Calculate signals to F 2 units: Yi = n i 1 b ij i Ste 5: Perform Stes 6 and 7 hen reset is true. Ste 6: Find F2 unit Y J ith largest signal (J is defined such that Y J Y j, j=1 to m). Ste 7: check for reset: JGRCS 2010, All Rights Resered 22

3 Tirtharaj Dash et al, Journal of Global Research in Comuter Science, 3 (8), August 2012, e P i =U i +dt ji u r i = i ci e u c If r < (ρ-e), then y J = -1 (inhibit J). Reset is true Perform Ste 5. If r (ρ-e), then i =s+au i q i = i e i =f(x i ) + bf(q i ). Reset is false. Proceed to Ste 8. Ste 8: Perform Stes 9-11 for secified number of learning iterations. Ste 9: Udate the eights for inning unit J: t ji = αdu i + {[1+ αd(d-1)]}t ji b ij = αdu i + {[1+ αd(d-1)]}b ij Ste 10: Udate F1 actiations: e i = s i +a U i P i = U i +dt ji q i = i e i =f(x i ) + bf(q i ). Ste 11: check for the stoing condition of eight udation. Ste 12: Check for the stoing condition for the number of eochs The symbols used in the aboe algorithm are: n = number of comonents in inut training ector m = maximum number of cluster units that can be formed ρ = igilance arameter (set beteen 0 and 1) s = binary inut ector x = actiation ector for F 1 (b) layer α = learning trials b ij = bottom-u eights t ji = to-don eights (Weights from Y j of F 2 layer to X i unit of F 1 (b) layer) x = norm of ector x and is defined the square root of the sum of the squares of the resectie comonents of x i (i=1 to n). It is imortant to mention that both the sequential and arallel imlementation is executed in Linux enironment (Ubuntu 10.04). It may be noted that, both sequential and arallel rocessing has been executed in a Intel dual core PC haing 1 GB RAM and 2 GHz rocessor seed. Ste-5: Testing the erformance of ART-2 Each of the forged signatures is then assed through the trained ART-2 netork and the number of matched b ij is counted. Equation 3 exresses the mismatch ercentage as follos, * b ij mismatch (3) count In this equation, count denotes the total number of bottomu eights (b ij ) and b* ij are the eights hich are matched ith the training cases. In this ork, e hae chosen 5% mismatch ercentage to decide on hether a signature could be acceted or not, i.e. mismatches u to 5% are alloed else rejected. It is imortant to note that mismatch acts as a threshold and threshold setting must be situation secific and the choice of user /administrator [19]. RESULTS The aerage similarity index (SI) beteen the original and forged signatures near 51%, hich may hae higher chance of matching, instead of rejecting the forged signatures. It is desired that een ith slightest difference, the netork must be able to differentiate those from the original signature based on its learning and assigned igilance. The aer suggests that igilance arameter (ρ) needs to be otimally set, hich is the first challenge. In this ork, otimum ρ has been set through a detail arametric study (see table-1 and 2). The second challenge is to assure that the netork learns the exemlary atterns through seeral obserations (number of clusters). Table 3 shos ho the cluster size (m) influences the accuracy and the comutational times. In table 1, it may be seen that ith forged signatures 11 and 12 the mismatch is <5% and therefore these are acceted as original. In case, the mismatch threshold is set <1%, the algorithm ould be able to detect all forged signatures. Hence, e hae made the algorithm ery flexible to allo such modification, hich deends on the situations. Table 2 shos that ith ρ=0.97, the detection accuracy is almost 100% ith minimum time in both sequential and arallel rogramming. Table-1 Result obtained from ART-2 net ith different alues of igilance arameter (ρ) Test Cases Vigilance Parameter (ρ) (Original s.) Original Forged Forged Forged Forged Forged Forged Forged Forged Forged Forged Forged Forged JGRCS 2010, All Rights Resered 23

4 Accuracy (%) Aerage Accuracy (%) Tirtharaj Dash et al, Journal of Global Research in Comuter Science, 3 (8), August 2012, Table-2 Accuracy rate at different igilance arameter ith corresonding comutation time ρ Accuracy (%) Comutation Time (seconds) Sequential Parallel Fig.3 lots the ρ s. accuracy arametric study. As seen in table 2, for ρ = 0.97, the accuracies are in both the sequential and arallel rogramming, e hae shon the lot for arallel rocessing Vigilance Parameter Figure-3: Parametric study ith igilance arameter and the detection accuracy. Table-3 Imroement of accuracy ith number of clustering units (m) in the F 2 layer #m Accuracy (%) Comutation Time (sec.) Sequential Parallel Table-3 shos that ith ρ=0.97 and m = 20, the detection accuracy is the highest. Fig. 4 lots the number of clusters (m) s. the resectie accuracy leels achieed Number of Cluster units Figure-4: Parametric study ith number of cluster units and the detection accuracy. Therefore, e conclude that through arametric studies, our method gies more accurate result hen comared ith other techniques, described in section I. CONCLUSIONS AND FUTURE WORK An ART-2 tye net has been deeloed in this ork for automating the erification of ery similar looking (SI ~51%) forged signatures, offline. It has been imlemented ith both sequential and arallel rocessing to achiee faster and accurate detection ith a mismatch threshold of 5%. Through arametric studies the best ρ and m are obtained. The accuracy is found to be 99.98%. It is imortant to mention that, in this study e hae tested only tele forged signatures, hich is a small sized samle. This is certainly a limitation of this ork. The net needs to be tested ith many different tyes of original as ell as forged signatures for obtaining a more authentic roof of its erformance. We are currently orking on it. REFERENCES [1] S. Chattoadhyay, P. Kaur, F.A. Rabhi, U.R. Acharya Neural Netork Aroaches to Grade Adult Deression, Journal of Medical Systems (2011), ublished online on 21/07/2011 DOI: /s [2] S. Chattoadhyay, P. Kaur, F.A. Rabhi, U.R. Acharya An Automated System to Diagnose the Seerity of Adult Deression. In the roceedings of 2 nd International Conference on Emerging Alications of Information Technology (CSI EAIT-2011), th February Kolkata India, (Editors - Jana D., and Pal P.), Publishers: IEEE Comuter Society and Conference Publishing Serices IEEE Xlore, Google Scholar, DOI: /EAIT JGRCS 2010, All Rights Resered 24

5 Tirtharaj Dash et al, Journal of Global Research in Comuter Science, 3 (8), August 2012, [3] Chattoadhyay S., Pratihar D. K, De Sarkar S. C. "Statistical Modelling of Psychoses Data. Comuter Methods and Programs in Biomedicine (2010) 100(3): [4] S. Chattoadhyay., D. K. Pratihar, S.C. De Sarkar - "Some Studies on Fuzzy clustering of sychosis data". International Journal of Business Intelligence and Data Mining (2007) 2(2): [5] S.K. Ahmed, A.K. Ramasamy, A. Khairuddin, J. Omar. Automatic online signature erification: A rototye using neural netorks. TENCON [6] G.A. Carenter, S. Grossberg, A self organizing neural netork for suerised learning, recognition, and rediction. Can neural netorks learn to recognize ne objects ithout forgetting familiar ones? IEEE Communication Magazine, 1992, [7] S.N. Sianandan, S.N. Deea Princile of Soft Comuting (2nd Edition) WILEY-INDIA, ISBN , [8] Dash T., Nayak T., Chattoadhyay S., Offline Verification of Hand Written Signature Using Adatie Resonance Theory Net (Tye-1), ICECT-2012, Vol-2, , [9] S. Inglis, I.H. Witten. Comression-based Temlate Matching. Proc. IEEE Data Comression Conference. Los Alamitos, CA, , [10] Y. Mizukami, M. Yoshimura, H. Miike, I. Yoshimura. An off-line signature erification system using an extracted dislacement function. Pattern Recognition Letters. 23, , [11] A. Kholmato, B. Yanikoglu. Identity authentication using imroed online signature erification method. Pattern Recognition Letters. 26, , [12] A.P. Shanker, A.N. Rajagoalan. Off-line signature erification using DTW. Pattern Recognition Letters. 28, , [13] W. Tian, Y. Qiao, Z. Ma. A Ne Scheme for Off-line Signature Verification Using DWT and Fuzzy Net. ACIS Int. Conf. Softare Engineering, Artificial Intelligence, Netorking, and Parallel/Distributed Comuting,. Vol.3,.30-35, [14] K.R Radhika, M.K. Venkatesha, G.N. Sekhar. An aroach for online signature authentication using Zernike moments. Pattern Recognition Letters. 32, , [15] A.K.M. A., Hossain, Md. G. Rahman, R.C. Debnath. Comaratie study of signature erification system using Suerised (BPNN) and Unsuerised (HMM) Models. ICCIT-2003, ,, [16] H. L, W. Wang, C. Wang, Q. Zhuo. Off-line Chinese signature erification based on suort ector machines. Pattern Recognition letters. 26, , [17] M.S.Aksoy, H. Mathkour. Signature erification using rules 3-ext inductie learning system. Int. J. Physical Sciences. 6(18): , [18] Dash T., Nayak T., Chattoadhyay S., Offline Hand Written Signature Verification Using Associatie Memory Net. Int. J. of Adanced Research in Comuter Engineering & Technology. 1(4): ,2012. [19] S. Chattoadhyay, S. Banerjee., F.A, Rabhi, U.R.Acharya A Case-based Reasoning System for Comlex Medical Diagnoses. Exert System :The Journal of Knoledge Engineering (2012) DOI: /j x (in ress). JGRCS 2010, All Rights Resered 25