Internatonal Journal of Sgnal Processng Systems Vol, No June 203 Offlne Verfcaton of Hand Wrtten Sgnature usng Adaptve Resonance Theory Net (Type-) Trtharaj Dash Veer Surendra Sa Unversty of Technology, Dept of Computer Scence and Engneerng, Burla-76808, Inda Emal: trtharajnst446@gmalcom Tanstha Nayak Natonal Insttute of Scence and Technology, Department of Informaton Technology, Berhampur-76008, Inda Emal: tansthanst23@gmalcom Subhagata Chattopadhyay Camella Insttute of Engneerng, Dept of Computer Scence and Engneerng, Kolkata-70029, West Bengal, Inda Emal: subhagatachatterjee@yahoocom Abstract Recognzng hand-wrtten sgnature s mportant due to several legal ssues Manual verfcaton s often dffcult when very much smlar-lookng but forged sgnature s produced In ths work an effort has been made to automate such knd of sgnature verfcaton process offlne, usng Adaptve Resonance Theory type- (ART-) It s mplemented usng both seral and parallel processng, the performance of whch are then compared The sad network has been traned wth the orgnal sgnature and tested wth two forged sgnatures The grade of smlarty has been computed by ntroducng the term Smlarty Index (SI) Performance analyss reveals that after a careful tunng of vglance parameter (ρ), both seral and parallel processng are able to learn the exemplary patterns wth 00% accuracy Whle testng, t s noted that parallel processng performs better than the seral processng n terms of speed as well as dentfyng the forged sgnatures by computng the msmatch network has been consdered for automatc verfcaton (offlne) of hand-wrtten sgnature The reason for consderng ART- n ths study s that ART- learns faster, consumes less memory space, and can adapt newer patterns wthout losng the ntally stored patterns [3] Sgnature of a person s the most wdely used technque for dentty verfcaton However, the ssue s that, every new sgnature of a same person may vary due to re-ntalzaton of scrbblng n each tme, whch mght be msused or forged by others ART s able to store all the patterns ncludng the varatons, whch mght not be possble wth other NN-based technques, such as ADALINE/MADALINE or Perceptron Hence, t nvtes a vast scope of research Due to ncrease n computaton power, a lot of fast algorthms have been developed for hand wrtten sgnature verfcaton [4] Here, we have mplemented handwrtten sgnature verfcaton (offlne) wth ART usng both (a) seral and (b) parallel programmng technques In fact, for parallel programmng we have used an API package commonly known as openmp (Open Multprocessng) [4] We compared the two algorthms wth respect to tme of executon and results Avalable lterature shows that usng NNs and based on the parameters, such as () heght, () slant, () pressure of pen tp and (v) velocty of sgnng, sgnatures could be verfed wth 97% accuracy for the best case [5]-[7] Hdden Markov Model (HMM) was used for sgnatureverfcaton It was performed by smple analyss of alphabet sequences wthn the sgnature Accordng to ths model, a sgnature was consdered as a sequence of vectors related to each pont of sgnature n ts trajectory In HMM matchng between model and sgnature was performed by usng the probablty of how the orgnal sgnature was calculated If the probablty s hgh, then only the sgnature was accepted The use of Back propagaton NN was also proposed wth success for offlne sgnature verfcaton [5] Index Terms -hand-wrtten sgnature, automatc verfcaton, ART-, smlarty ndex, forged sgnatures I INTRODUCTION Learnng exemplary patterns s an mportant property of a Neural Network (NN) Based on the learnng mechansms, there are several types of NN, such as Adaptve Lnear Net (ADALINE), multple ADALINE (MADALINE), Perceptrons, Hopfeld net and so forth (uses supervsed learnng methods), and Kohonen s Self Organzng Map, Adaptve Resonance Theory (ART) net, whch work by unsupervsed methods [] Ths paper focuses on ART type- Adaptve Resonance theory (ART) networks were frst developed by Steven Grossberg and Gal Carpenter n 987[2] ART s of two types e type- and type-2 ART- takes bnary nput vector, whereas, ART-2 takes analog/contnuous nput vector [3] In ths research, ART Manuscrpt receved Feb 2, 203; revsed June, 203 do: 02720/jsps7-22 7
Internatonal Journal of Sgnal Processng Systems Vol, No June 203 Contour technque was proposed by [8], where the contour of the sgnature was generated by dlatng the sgnature by varous levels generatng a band lke structure Ths dlated sgnature template was then compared wth the test sgnature template Here they used color bands for varous dlaton levels and an EX-OR were used for color bands to fnd varaton n sgnature segments Ths technque was fast and smpler to mplement and gave an accuracy of 75% Statstcal Approach usng correlaton was proposed by [9] But t faled n case of sklled forgeres As ths method usually descrbes the characterstcs of the sgnature related to the shape of sgnature Structural or Syntactc approach s an pattern recognton technque whch represents patterns n the form of symbolc data structures vz Trees, graph, strngs etc To verfy a forge sgnature, ts symbolc data was compared wth number of prototypes whch are stored n the database These structural features used modfed drecton and transton dstance feature (MDF) [0] whch extracts the transton locatons MDF are based on relatonal organzaton of low-level features nto hgherlevel structures Wavelet-based approach was used by [] where both statc and pseudo dynamc features could be extracted from the orgnal sgnature and processed by wavelet transform, whch were then converted to sub-bands Ths ncreased the dfference between orgnal and forged sgnatures Ths method gave an average error rate of 257% for Englsh sgnature and 396% for Chnese Sgnature Handwrtten Sgnature verfcaton (offlne) s also done by usng Adaptve Resonance Theory Method [2] In ths work, they have used ART-2 net for recognzng very smlar lookng but forged sgnature In ths study, the accuracy rate s almost 00% Assocatve Memory Net (AMN) approach was used for recognzng Handwrtten Sgnature Verfcaton [3], [4] Here, the accuracy rate 923% Adaptve Resonance Theory (ART) method s also used for sgnature verfcaton purpose Based on -bt quantzed pressure pattern n tme doman, the work was beng proposed The tmng nformaton s used for screenng for frst stage screenng of ncomng sgnatures usng ART- networks wth varous values of vglance parameter [5] Edge Detecton can be another method for sgnature verfcaton, where edge of the sgnature can be consdered for tranng and testng [6] II METHODOLOGY The study has been performed n the followng steps: Step-: Collectng hand wrtten sgnatures Step-2: Extractng the features of all the sgnatures by to get all the gray-scale values Step-3: development of ART- net on C language wth () seral and () parallel processng Step-4: Tranng ART net wth the orgnal sgnature, and Step-5: Testng ART net wth two forged sgnatures and compute the error by notng the % of msmatch Step-: Hand sgnature collecton: Fg, 2a, and 2b show the hand-wrtten sgnatures both orgnal and forged, collected for ths work The sze of the sgnatures s 324x20 dp In the next stage, Smlarty Index (SI) has been computed between each of the forged sgnatures and the orgnal sgnature usng equaton The arrangements of pxels ( on and off ) n the forged sgnatures are compared wth that of the orgnal sgnature row and column-wse The dspartes/dssmlartes are then noted In equaton, D s the number of dssmlar p pxels and Tp s the total number of pxels The SIs computed for forged sgnature and 2 are ~5% Dp SI 00 () T p Step-2: Feature extracton: We extracted the pxel values n RGB (Red-Green- Blue) format and then converted t to gray scale format by usng the followng standard relaton Gray value= 033 R + 056 G + 0 B (2) It s mportant to note that we have used a text fle to extract the pxel value n the bnary form usng the followng condtonal statements If gray value > 0 Then append a to the text fle Else Append a 0 A sample of pxel values are gven n Appendx-A Fgure The orgnal sgnature Fgure 2(a): Forged sgnature Fgure-2(b): Forged sgnature 2 8
Internatonal Journal of Sgnal Processng Systems Vol, No June 203 Step-3 and 4: Development of ART- algorthm and ts tranng: As already mentoned, t has been developed on C language Fg 3 shows the schematc dagram of ART- structure The mplementaton algorthm s as follows: Step-0: Intalze the parameters: The symbols used n ths algorthm: n = number of components n nput tranng vector; m = maxmum number of cluster unts that can be formed; ρ = vglance parameter (set between 0 and ); α = learnng trals; b j = bottom-up weghts; α > and 0 < ρ t j = top-down weghts (Weghts from Y j of F 2 layer to Intalze the weghts: X unt of F (b) layer); 0 < b j (0) < α/(α-+n) and t j (0) = s = bnary nput vector; x = actvaton vector for F (b) layer; x = norm of vector x and s defned as the sum of components of x (= to n) Step-: Perform Steps 2-3 when stoppng condton s false Step-2: Perform Steps 3-2 for each tranng nput Step-3: Set actvaton of all F2 unts to zero Set the actvaton of F(a) unts to nput vectors Step-4: Calculate the norm of s: s = s It s mportant to menton that both the seral and parallel processng usng OpenMP were executed n Lnux envronment System specfcaton: It may be noted that, both seral and parallel processng has been executed n a PC wth Intel dual core processor, GB RAM, and 2 GHz Step-5: Testng the performance of ART- The forged sgnature s passed to the traned ART- network and the number of updated b j s counted Now (3) Step-5: Send nput sgnal form F (a) layer to F (b) layer: x = s (4) the b j updated durng the tranng and the new b j after Step-6: For each F node that s not nhbted, the followng rule should hold: If y j -, then y j = b j x testng s compared The rato of equal b j s to the total b j of tranng 00 (5) gves the percentage of matchng of the two sgnatures from ths result the tested sgnature can be accepted or rejected Step-7: Perform Steps 8- when reset s true Step-8: Fnd J for y J y j for all nodes j If y J = -, then all the nodes are nhbted and note that ths pattern cannot be clustered Step-9: Recalculate actvaton of X of F (b): x = s t (6) Step-0: Calculate the norm of vector x: x = x (7) Step-: Test for the reset condton If x / s < ρ, then nhbt node J, y J = - Go back to Fgure 3: Structure of ART- Step 7 agan Else f x / s ρ, then proceed to the next step (Step 2) Step-2: Perform weght updaton for node J (fast learnng): b J (new) = α x / (α -+ x ) (8) t J (new) = x (9) Step-3: Test for the stoppng condton The stoppng condtons may be: a No change n weghts b No reset of unts c Maxmum number of epoch reached Fgure 4: Tranng and testng of ART- 9
Internatonal Journal of Sgnal Processng Systems Vol, No June 203 III RESULTS Fg 5 s a plot drawn to compare the results of seral and parallel processng It shows that, at ρ=03 the parallel mplementaton gves better result by consumng only 56 sec durng executon Fg 6a and 6b shows the parametrc study on dfferent vglance parameters (ρ) vs msmatch (e the squared errors) n case of frst forged sgnature for the seral and parallel processng, whle fg 7a and 7b plots the same for the second forged sgnature These are dscussed below The average smlarty ndex (SI) between the orgnal and forged sgnatures near 60%, whch may have hgh chance of matchng, nstead of rejectng the forged sgnatures It s desred that even wth slghtest dfference, the network must be able to dfferentate those from the orgnal sgnature based on ts learnng and assgned vglance It s certanly a real-world challenge to curb ths ssue The paper suggests that vglance parameter (ρ) needs to be optmally set, whch s the frst challenge In ths work, optmum ρ has been set based on the percentage of msmatch (whch s the squared error) through a detal parametrc study Table I and Table II show the parametrc study of settng ρ The second challenge s to assure that the network learns the exemplary patterns through several observatons The thrd challenge s that the learnng and ts executon must be accomplshed at mnmum tme When orgnal sgnature s checked wth tself, t s noted that the net learns 00% exemplary patterns at average tme of 370 sec n case of seral processng (see Table I) In case of parallel processng, 00% exemplary pattern has been learnt wth much less tme, e, 620 sec (see Table II) Whle detectng the forged sgnatures, t s noted that up to % msmatch could be detected by seral processng wth average tme of 885 sec wth ρ=099 In case of parallel processng such amount of msmatch could be detected wth average of 70 sec, whch s much less, than the tme consumed durng seral processng Also, durng the parallel processng the values of ρ wdely vares yeldng more flexblty to the network In ths study, msmatches are dentfed wth ρ values 0, 0, 068 and 099 respectvely 8 Parallel Seral Avg tme(sec) 6 4 2 0 8 6 4 0 02 03 04 05 06 07 08 09 Vglance Parameter Fgure 5 Plot showng the average executon tme vs Vglance parameter (ρ) n Seral and Parallel processng 2 8 6 4 2 ρ RESULTS OF SERIAL PROCESSING [Α = 00] vs vs Forged 488 0 00 00 320 489 440 478 390 020 00047 400 48975 320 7 390 030 000 380 2 390 70 380 0 00 370 2 390 7 360 068 0004 37 2 390 07 340 080 000 400 4897 40 4876 70 099 00 390 007 870 07 900 vs vs Forged 822 480 607 9 437 020 002 6059 4807 595 7 5 030 000 620 0 47 72 593 0 002 69 9 820 7 78 068 002 698 09 825 70 780 080 080 00 709 4897 899 00 099 00 640 007 689 07 809 00 04 05 06 07 08 09 5 5 5 vs Forged2 00 03 Fgure 6a Plot showng the Squared error produced wth several Vglance parameter (ρ) n Seral processng for frst forged sgnature TABLE II RESULTS OF PARALLEL PROCESSING [Α = 00] ρ 02 Vglance Parameter vs Forged2 TABLE I 485 48 0 02 03 04 05 06 07 08 09 Vglance Parameter Fgure 6b Plot showng the Squared error produced wth several Vglance parameter (ρ) n Parallel processng for frst forged sgnature From Fg 6a and Fg 6b, t may be noted that for the frst forged sgnature, n case of seral and parallel processng the maxmum msmatches are found wth ρ = 20
Internatonal Journal of Sgnal Processng Systems Vol, No June 203 099 and 05, respectvely On the other hand, from Fg 7a and Fg 7b, t may be noted that for the second forged sgnature, n case of seral and parallel processng the maxmum msmatches are found wth ρ = 099 and 045 & 07, respectvely Therefore, ρ values between 045 099 mght be able to catch the forged sgnatures mostly IV An ART- type NN has been developed n ths work for automatng the forged sgnature verfcaton task offlne It has been mplemented wth both seral and parallel processng Whle tranng, the accuracy of the sad net n both the processng technques s closed to 00% The accuracy s tested on two closely resembled but forged sgnatures wth smlarty ndex (SI) >60% The study observes that wth parallel processng the msmatches of both the forged sgnatures are detected completely wth ρ equals to 0, 068, and 099 Whle comparng wth other avalable technques, the developed net has outperformed other technques n terms of accuracy n predctng forged sgnatures Here msmatch acts as threshold and threshold settng must be stuaton specfc 5 5 485 48 475 0 02 03 04 05 06 07 08 09 Vglance Parameter APPENDIX A SIGNATURE PIXELS Fgure 7a: Plot showng the Squared error produced wth several Vglance parameter (ρ) n Parallel processng for second forged sgnature Sample pxel grds of Sgnature: 0 0 5 5 CONCLUSIONS AND FUTURE WORK 5 Sample Pxel grd of Forged Sgnature : 0 02 03 04 05 06 07 08 09 0 Vglance Parameter Fgure 7b: Plot showng the Squared error produced wth several Vglance parameter (ρ) n Parallel processng for second forged sgnature Fnally, the performance (based on detecton accuracy) has been compared wth other technques used n handwrtten sgnature verfcatons (refer to Table-3) It shows that our method gves more accurate result when compared wth other technques However, n ths study we have tested only two forged sgnatures Ths s a lmtaton The net needs to be tested wth many dfferent types of forged sgnatures We are at present workng on ths ssue Sample Pxel grd of Forged Sgnature 2: 0 00 TABLE III COMPARISON TABLE FOR DIFFERNT METHODS Model used Accuracy HMM[5] BPNN [5] Modular Neural Net(MNN) [6] ANN [6,7] Template Matchng Technque [8] Wavelet-based approach [] ART-2 net [2] AMN[3] ART-(n ths work) 889 839 966 9427 75 8743; 8604 9998 923 9999 0 REFERENCES [] 2 K Krshna, A Goyal, and S Chattopadhyay, Non-correlated character recognton usng Hopfeld network: a study, n Proc Internatonal Conference on Computer and Computatonal Intellgence, Bangkok, Thaland, 20, pp 385-389
Internatonal Journal of Sgnal Processng Systems Vol, No June 203 [2] G A Carpenter and S Grossberg, A self organzng neural network for supervsed learnng, recognton, and predcton, IEEE Communcaton Magazne, vol 30, no 9, pp 38-, 992 [3] S N Svanandan and S N Deepa, Prncple of Soft Computng, 2 nd ed, WILEY-INDIA,, 20 [4] R Chandra, L Dagum, D Kohr, Maydan, J McDonald, and R Menon, Parallel Programmng n OpenMP, Morgan Kaufmann Publsher, 200 [5] A K M A Hossan, Md G Rahman, and R C Debnath, Comparatve study of sgnature verfcaton sustem usng supervsed (BPNN) and unsupervsed (HMM) Models, n Proc 6 th Internatonal Conference on Computer and Informaton Technology, Dhaka, Bangladesh, 2003 [6] O Mrzae, H Iran, and H R Pourreza, Offlne sgnature recognton usng modular neural networks wth fuzzy response ntegraton n Proc Internatonal Conference on Network and Electroncs Engneerng, vol, Sngapore, 20, pp 53-59 [7] A McCabe, J Trevathan, and W Read, Neural network-based handwrtten sgnature verfcaton, Journal of Computers, vol 3, no 8, pp 9-22, 2008 [8] S Ingls and I H Wtten, Compresson-based template matchng, n Proc IEEE Data Compresson Conference, Los Alamtos, CA, 994, pp 06-5 [9] I W McKeague A statstcal model of sgnature verfcaton [onlne] pp -25 2004 Avalable: http://wwwcolumbaedu/~m23/ps/sg-revpdf [0] M Blumensten, X Y Lu, and B Verma, A modfed drecton feature for cursve character recognton, n Proc Internatonal Jont Conference on Neural Networks, 2004, pp 2983-2987 [] W Tan, Y Qao, and Z Ma, A new scheme for off-lne sgnature verfcaton usng DWT and fuzzy net, n Proc 8 th ACIS Internatonal Conference on Software Engneerng, Artfcal Intellgence, Networkng, and Parallel/Dstrbuted Computng, 2007, vol 3, pp 30-35 [2] T Dash, S Chattopadhyay, and T Nayak, Handwrtten sgnature verfcaton usng adaptve resonance theory type-2 (ART-2) Net Journal of Global Research n Computer Scence, vol 3 no 8, pp 2-25, 202 [3] T Dash, T Nayak, and S Chattopadhyay, Offlne handwrtten sgnature verfcaton usng assocatve memory net, Internatonal Journal Advanced Research n Computer Engneerng & Technology, vol no 4, pp 370-374, 202 [4] T Dash, T Nayak, and S Chattopadhyay, Handwrtten sgnature verfcaton (offlne) usng neural network approaches: A comparatve study, Internatonal Journal of Computer Applcatons, vol 57 no 7, pp 33-4, 202 [5] L Y Tseng and T H Huang, An onlne Chnese sgnature verfcaton scheme based on the ART neural network, n Proc Int Conf Neural Networks, 992, vol 3, pp 624-630 [6] T Dash and T Nayak, A java based tool for basc mage processng and edge detecton, Journal of Global Research n Computer Scence, vol 3 no 5, pp57-60, 202 Trtharaj Dash s pursung hs Master of Technology n Department of Computer Sc & Engg at Veer Surendra Sa Unversty of Technology, Burla He has publshed more than 20 research artcles n journals of nternatonal repute and conferences Hs research area ncludes Parallel Computng, Algorthms, Soft Computng and AI Tanstha Nayak has done her Bachelor of Technology n Informaton Technology at Natonal Insttute of Scence and Technology, Berhampur She has contrbuted 8 papers Soft Computng, Quantum Computng and Image processng researches Dr Subhagata Chattopadhyay receved hs PhD n Informaton Technology from Indan Insttute of Technology (IIT), Kharagpur and postdoctoral fellowshp n e-health nformaton system at Australan School of Busness, the Unversty of New South Wales (UNSW) Sydney Australa He has publshed 98 research papers n the feld of Soft Computng, Pattern Recognton, AI, Image Processng, Knowledge Engneerng and Management Recently he has publshed one edted book named "ADVANCES IN THERAPEUTIC ENGINEERING", by CRC Press USA Dr Chattopadhyay s presently workng as a full professor n the Dept of Computer Scence and Engneerng at Camella Insttute of Engneerng, Inda He s also the Drector of the sad nsttute 22