Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Offline Handwritten Signature Verification Using Associative Memory Net

Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

ISSN: 2278 1323 International Journal of Advanced Research in Computer Engineering & Technology Volume 1, Issue 4, June 2012

Offline Handwritten Signature Verification using Associative Memory Net


Tirtharaj Dash, Tanistha Nayak, Subhagata Chattopadhyay
Abstract Handwritten signature verification must be accurate due to several legal issues. The verification mandates faster detection as well. Human eye often fails to differentiate accurately a very similar-looking forged signature from its genuine version. The clarification is often a time-taking process. Therefore, such verification, from the computer science perspective, is a bi-objective optimization problem, where cost functions are errors and computational time. This paper studies application of Associative Memory Net (AMN) in correctly detecting forged signatures, fast. Here, the cost functions are handled with detail parametric studies and parallel processing using OpenMP. The algorithms are trained with the original or genuine signature and tested with a sample of twelve very similar-looking forged signatures. The study concludes that AMN detects forgery with accuracy 92.3%, which is comparable to other methods cited in this paper. Index Terms Signature verification, Offline, Bi-objective optimization, Associative Memory Net, OpenMP;

Adaptive Resonance Theory Net (ART) and Kohonens Self Organizing Map (SOM) rely on unsupervised learning and can store previously learned patterns. Associative Memory Net (AMN) updates its knowledge base through the concept of supervised learning. Depending on target pattern, AMN can be Auto-AMN or Hetero-AMN. In the former type, target is similar to the training pattern, while it is different in case of its counterpart. In this study, we have developed an algorithm based on the principle of Auto-AMN. As signature verification is a bi-objective optimization problem where lowest error should be expected with least time, parallel methods have been proposed. The parallel implementation distributes the computation work to multiple processors and achieves the same result in a minimal CPU time. The verification is performed offline. II.
RELATED WORKS

I. INTRODUCTION In the developing countries, hand-written signature is the most commonly used biometric verification techniques [1]. Skillful copy of a signature is not so hard and if cannot be detected, several legal issues may surface up. Detecting such a false but very similar looking signature is quite challenging in machine intelligence research. To address this issue, various signature detection schemes are being proposed since couple of decades, which include both traditional approaches and soft computing techniques. Neural Network (NN) learns patterns by examples or observations [1]. Learning can be supervised or unsupervised. Adaptive Linear Net (ADALIN), Multiple ADALIN (MADALIN), Perceptron Network, Radial Basis Function Net (RBFN) etc. are some important examples of supervised learning methods. These NNs learn faster and accurately, but the problem with these is that, new training is required each time it learns new input patterns and as a result, the previously learned patterns are lost. On the other hand, networks, such as Counter Propagation Network (CPN),

Manuscript received May, 2012. Tirtharaj Dash, Dept. of Information Technology, National Institute of Science and Technology, Berhampur, India, Mob: 09853190787 (e-mail: tirtharajnist446@gmail.com). Tanistha Nayak, Dept. of Information Technology, National Institute of Science and Technology, Berhampur, India, Mob: 09692882373 (e-mail: tanisthanist213@gmail.com). Subhagata Chattopadhyay, Dept. of Computer Science and Engineering, National Institute of Science and Technology, Berhampur, India (e-mail: subhagatachatterjee@yahoo.com).

We have conducted a literature survey on various traditional and soft computing-based methods, used for signature verification. The major sources of the literatures are Google Scholar, Scopus, Science Direct, and IEEE Xplore. Relevant studies are briefly discussed below. A method of signature verification using ART-1 was developed by [1] which resulted in an accuracy of 99.9% in case of skilled forgeries. This method was based on Training and testing of the network. A Bayesian network representation has been proposed by Xiao and Leedham, (2002). They proposed a decision tree like network, where each node of the tree computes the conditional probability as the chance of matching [2]. Displacement extraction method has been proposed for offline signature verification. In this work, the authors have extracted a displacement function between a pair of signatures, original and fake. The study concludes that the average accuracy rate of detection is around 75% [3]. In the year 2005, Kholmatov and Yanikoglu presented an online signature recognition scheme using Dynamic Time Warping (DTW) and three dimensional feature vectors. The overall accuracy rate of the method was 98.6%. The authors have used and [4]. Optimal function of features was used for online verification [5]. In this work, the authors first choose a candidate function and optimized it to produce an optimal function. Basically the optimization was done using well known Genetic Algorithm (GA). The error rate in this work was only 0.1%. Cooperating NN was used for offline signature verification. The features used in this work were geometrical 370

All Rights Reserved 2012 IJARCET

ISSN: 2278 1323 International Journal of Advanced Research in Computer Engineering & Technology Volume 1, Issue 4, June 2012 parameters, outline and image of the signature. The accuracy rate in this work was 96% [6]. Wavelet thinning features were used for offline signature verification using Matching Algorithm. Similarity measurement was evaluated using Euclidean distance of all found corresponding feature points. The accuracy in this case was 81.4% [7]. Hidden Markov Model (HMM) was successfully used for signature verification. It was performed by analysis of alphabets within the signature. According to this model, a signature is collection of vectors related each point in its outline. The average accuracy rate was 88.9% in this case [8]. The method using distance statistics yielded an accuracy of 78% for verification and 93% for signature identification [9]. A combination of shape contexts and local features were used for online signature verification. However, DTW technique was also used for elastic matching between signatures. The proposed method suggested that by combining local features with shape contexts the performance of the algorithm could be increased. The average error rate in this work was 6.77% [10]. III. METHODOLOGY A step-wise method has been followed in this work. The steps include: Step-1: Collection of the signature samples (Original as well as Forgeries) Step-2: Feature Extraction Step-3: Implementation of AMN in C language Step-4: Training of the networks with genuine signatures Step-5: Testing of the developed AMN Step-1: Collection of Signature samples Original signature is produced at first and then forged signature samples are collected from twelve different persons at different times (refer to Appendix-I). Each person has been given ample time (10 days each) to enable copying the original signature with almost no visually detectable mistakes. All signatures are then saved into BMP files of size: 20063dpi with Bit depth as 4. The signature templates are shown in Appendix-I. Step-2: Feature Extraction A user-defined image function in C-Program is used for extraction of pixels from the BMP files [11]. The method of feature extraction is given in Figure-1. 3.1. Serial Implementation of the Network Our developed algorithm for the approach and the training of the AMN network is given below. It will be wise to mention that training of the AMN had been done with the genuine sample only. INITIALIZE weight (W) to 0 INPUT the original sign. to the first layer of AMN FOR i=1 to n DO FOR j=1 to n DO Calculate the weight as Wij(new)=Wij (old) + INPUTiTARGETj END //end of for loop on j END //end of for loop on i FOR i=1 to n DO FOR j=1 to n DO Calculate the net output node as, Yinj=

input

to

each

x W
i i 1

ij

IF(Yinj>0) Yj=+1; ELSE Yj=-1; END //end of for loop on j END //end of FOR loop on i The structure of AMN and its training method is represented in Figure-2.

Figure-2 Structure of AMN and its training 3.2. Parallel Implementation of the Networks We developed parallel version of all the above algorithms using OpenMP (www.openmp.org). By doing this, three major things could be achieved; i. Reduction in computation time, ii. Utilization of all the processor of the system iii. Inherent parallelism property of NN could be used. The function used to calculate the time details in Sequential implementation is clock() and in parallel implementation is omp_get_wtime(). The header file used in the parallel programming is omp.h. [12] 3.3. System architecture specification

Figure-1 Feature Extraction using our C-program A sample pixel value of a signature is given in Appendix-II. Step-3: Implementation of AMN and its Trainings

371
All Rights Reserved 2012 IJARCET

ISSN: 2278 1323 International Journal of Advanced Research in Computer Engineering & Technology Volume 1, Issue 4, June 2012 The implementation (both Serial and parallel) are carried out in a system having Intel Quad Core Processor with 4GB RAM having a processor speed of 2.10GHz. The operating system used is Linux (Ubuntu Version 11.10). It is mentioned that the package used for parallel processing is the OpenMP 3.0. Step-4: Testing of the developed AMN The developed and trained AMN network was then trained with 12 samples of Forged signatures (Refer Fig.2). 4.1. Calculation of mismatch The numbers of YJ which are -1 are counted (count) (see Section 3.1). The mismatch percentage can be calculated using the following Equation-1. Mismatch = [count/Total Number of Pixels] 100 (1) C. Acceptance accuracy of each forged signature A plot has been given in Figure-3 to see which one of the forged signatures sample is getting verified with highest accuracy.
80 70 60
Mismatch --->

Forged8 Forged9 Forged10 Forged11 Forged12

29.08 31.02 29.09 21.57 22.09

Reject Reject Reject Accept Accept

12.08 12.69 6.304 7.91 14.52

4.33 4.54 2.00 2.89 4.83

Looking at the mismatch results of AMN from Table-1, it can be said that the network is performing better for all the forged signatures (Forged 1-10). But the network is rejecting the original signature itself (See Test case-1 of Table-1). The network is also accepting two of the forged signatures, Forged-11 and Forged-12. The happened due to setting of the mismatch threshold to 25%. However, the average time taken by the serial algorithm is 9.85 seconds where as the parallel algorithm takes only 2.98 seconds.

4.2. Setting the threshold of mismatch Threshold is the minimum mismatch percentage after which the tested signature could be termed as illegal. As the key task behind this work is to impose stricter security and safety applications, we set the threshold as low as 25% to avoid acceptance of skilled forgeries. It is important to note that, such threshold setting must be situation specific and the choice of the administrator/user [13]. IV. RESULTS AND DISCUSSIONS The contribution of our work can be outlined as; development of an algorithm for signature verification scheme using AMN neural network. It is wise to note that all the implementation is carried out using both serial and parallel processing techniques. In this paper, study on signature area and its affect on computation time is provided. A. Performance of AMN Technique It should be noted that the AMN network implemented here is auto-associative. The reason is that the forged signature will be checked with the genuine version of itself. So the network becomes more complex due to increase in pixels of the signature image. AMN is a two layer NN with the first layer being the pixels values of the forged signatures and second layer contains the pixel nodes from the original signature. B. Mismatch results of AMN Network The result in Table-1 gives the performance of our algorithm for signature verification using AMN. Table-1 Result of execution of Test cases Computation Time Test Case Mismatc (seconds) (Original h Decision vs.) (%) Serial Parallel Original 21.31 Accept 6.15 1.98 Forged1 29.29 Reject 7.9 2.44 Forged2 28.92 Reject 8.8 2.67 Forged3 29.86 Reject 12.06 3.38 Forged4 28.21 Reject 7.13 1.80 Forged5 28.44 Reject 10.24 2.87 Forged6 28.67 Reject 10.85 2.10 Forged7 29.65 Reject 11.46 3.02

50 40 30 20 10 0

10

11

12

Test cases --->

Figure-3 Accuracy achieve by the algorithm for each forged signatures Overall accuracy can be calculated as the number of correctly recognized sample test cases to the total number of test cases multiplied by 100. In our study, the accuracy is found to be 92.3%. D. Forged Signature versus Computation time plot A plot is given in Figure-4 showing CPU utilization time for AMN to recognize each of the forged signature samples.

372
All Rights Reserved 2012 IJARCET

ISSN: 2278 1323 International Journal of Advanced Research in Computer Engineering & Technology Volume 1, Issue 4, June 2012
15 14 12
Computation time --->

Serial Parallel

APPENDIX-I

10 8 6 4 2 0 0

5 6 7 8 Forged Signature samples --->

10

11

12

Figure-4 Time taken by algorithm to detect each forged signatures E. Comparison of our algorithm with other methods Table-2 shows the standing of our method in the team of other methods proposed for signature verification. Table-2 Comparative analysis of different methods of signature verification Yea Accurac Method r y ART-1[1] 2012 99.9% Displacement Extraction Method[3] 2002 75% DTW and 3D feature vector[4] 2005 98.6% GA and Optimal Function[5] 2004 99.9 Co-operative NN[6] 1994 96% Matching Algorithm[7] 2007 81.4% Hidden Markov Model[8] 1994 88.9% Distance statistics[9] 2004 78% DTW[10] 2006 93.23 Associative Memory Net (this work) 2012 92.3% V. CONCLUSION AND FUTURE WORK In this paper, we proposed a neural network based technique for offline hand-written signature verification. The method has been optimized with respect to accuracy and computation time. We proposed two versions of the algorithm, serial and parallel. The study revealed that our algorithm takes an average of 9.58 seconds in serial and 2.98 seconds in parallel to give a detection accuracy of 92.3%. The error threshold is set to 25% in this work to make decision. It is important to mention that in this work we have tested our algorithms on a sample of only twelve forged signatures. It will be impressive to test the algorithms with more training and test cases and then develop a GUI for the applications to make it user friendly tool.
[2] [3]

APPENDIX-II 15 15 15 15 15 15 15 15 15 15 15 15 8 7 7 0 0 0 0 0 0 0 0 0 0 0 7 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 15 8 0 0 0 0 0 7 15 15 15 15 15 15 15 15 15 15 REFERENCES
[1] Dash T., Nayak T., Chattopadhyay S., Offline Verification of Hand Written Signature Using Adaptive Resonance Theory Net (Type-1), in Proc: IEEE Int. Conf. Electronics Computer Technology (ICECT) (2012), Volume 2, pp. 205-210. Xiao X., Leedham G., Signature verification using a modified Bayesian network, Pattern Recognition (2002) Volume 35, 983-995. Mizukami Y., Yoshimura M., Miike H., Yoshimura I., An off-line signature verification system using an extracted displacement function, Pattern Recognition Letters (2002), Volume 23, 1569-1577. Kholmatov A., B. Yanikoglu, Identity authentication using improved online signature verification method, Pattern Recognition Letters (2005), Volume 26, 1400-2408. Romo J.C.M., Silva R.A., Optimal Prototype functions of features for online signature verification, International Journal of Pattern Recognition and Artificial Intelligence, Volume: 18, Issue: 7(2004) pp. 1189-1206. Cardot H., Revenu M., Victorri B., Revillet M.J., A static signature verification system based on cooperating neural networks architecture, International Journal of Pattern Recognition and Artificial Intelligence, Volume 8, Issue 3(1994) pp. 679-692. Fang B., You X., Chen W.S., Tang Y.Y., Matching algorithm using wavelet thinning features for offline signature verification, International Journal of Pattern Recognition and Artificial Intelligence, Volume 5, Issue: 1(2007) pp. 27-38. Inglis S., Witten I.H., Compression-based Template Matching, in proc: IEEE Data Compression Conference (1994), pp. 106-115.

[4]

[5]

[6]

[7]

[8]

373
All Rights Reserved 2012 IJARCET

ISSN: 2278 1323 International Journal of Advanced Research in Computer Engineering & Technology Volume 1, Issue 4, June 2012
[9] Kalera M.K., Srihari S., Xu A., Offline signature verification and identification using distance statistics, International Journal of Pattern Recognition and Artificial Intelligence, Volume 18, 7(2004) pp. 1339-1360. [10] Li B., Zhang D., Wang K., Online signature verification by combining shape contexts and local features, International Journal of Pattern Recognition and Artificial Intelligence, Volume 6, Issue 3 (2006) pp. 407-420. [11] Dash, T., Nayak T. A Java Based Tool for Basic Image Processing and Edge Detection, Journal of Global Research in Computer Science, Volume 3, Issue 5 (2012) pp. 57-60. [12] Chandra R., Dagum L., Kohr D., Maydan D., McDonald J., Menon R., Parallel programming in OpenMP, Morgan Kaufmann Publisher (2001). [13] Chattopadhyay S., Banerjee S., Rabhi F.A, Acharya R. U. A Case-based Reasoning System for Complex Medical Diagnoses. Expert Systems: the Journal of Knowledge Engineering (2012); DOI: 10.1111/j.1468-0394.2012.00618.x (in press). Tirtharaj Dash and Tanistha Nayak are the B.Tech. final year students, in the Dept. of Information Technology at NIST, India. Soft computing, Parallel computing, and Algorithm are their key research interests. They have published five research papers. Dr. Subhagata Chattopadhyay is a full professor in the Department of Computer Science and Engineering at National Institute of Science and Technology, Berhampur. He received his PhD in Information Technology from IIT, Kharagpur and Postdoctoral fellowship in e-Health Information Systems at Australian School of Business, the University of New South Wales (UNSW) Sydney Australia. He has published over 80 research papers in the field of Soft computing, Pattern recognition and AI.

374
All Rights Reserved 2012 IJARCET

You might also like