comparative study of offline signature verification techniques · comparative study of offline...

6
International Journal of Advancements in Research & Technology, Volume 2, Issue2, Feb ruary-2013 1 ISSN 2278-7763 Copyright © 2013 SciResPub. Comparative Study of Offline Signature Verification Techniques Devshri Satyarthi 1 , Y.P.S. Maravi 2 , Poonam Sharma 3 , R.K. Gupta 4 Department of CSE & IT, Madhav institute of technology and science Gwalior, 474002, India Email: [email protected], Tel +919944914443 ABSTRACT Today in the world of profit aiming mind, trust of people is at high risk where forgery is monotonous in which fake signature has come into limelight. The fraudulent signature thus needs to be verified using verification techniques. The signature forgery can be restricted by either online signature verification or offline signature verification techniques. Online signature on one hand verifies the signature by performing a match with the pre-processed signature dynamically by detecting the motion of stylus during signature while on other hand, offline verifies by performing a match using the two dimensional scanned image of the signature. This paper studies about the various techniques available in offline signature verification along with their shadows. Keywords : Signature verification, false rejection rate, Quality performance measures, Hidden markov model, Support Vector Machine, Neural network, Template Matching, Structural or syntactic, dynamic time warping. 1 INTRODUCTION In the real world instances authentication of an individual is desired frequently by distinguishing feature for person identi- fication. This authentication is generally offered by signature from past decades. Still in present, especially in financial, sig- nature is preferred as the demanding option for authentica- tion. Henceforth, with the increasing number of transactions, need of automatic signature verification will arise for authen- ticating the individual. Signature verification is an approach that is not only capable to find the identity addressing two individual efficiently but also strongly related tasks. On one side, the preliminary task is to identify the signature owner; on the other side, the secondary task is the decision, whether the signature is genuine or forged. Signature verification is spread into two categories: One is online signature verification and other is offline signature verification. Online signature verification dynamically scans the user sig- nature by tracing the motion on the stylus and verifies it against pre-processed signature information. On other hand, Offline signature verification gives static signature verification information, in which signature is scanned from document and verified against 2D scanned image of the signature. In this paper, the variations offered in offline signature verifi- cation schemes are discussed. The schemes vary from Hidden Markov model (HMM), Dynamic time warping (DTW), Sup- port vector machines (SVM), Neural network (NN), Wavelet transform to Structural or syntactic. 2. Quality Performance Measures In evaluating the performance of a signature verification system, there are two important factors: the false rejection rate (FRR) of genuine signatures and the false acceptance rate (FAR) of forgery signatures and these two are inversely related. The false rejection rate (FRR), the false acceptance rate (FAR), the equal error rate (EER) and the average rate (AER) are used as quality performance measures. The FRR is the ratio of genuine test signatures rejected to the total number of genuine test signatures submitted. The FRR called the type I error and is defined as, The FAR is the ratio of the number of forgeries accepted to the total number of forgeries submitted. The FAR is also called the type II error and is defined as, The average of the FRR and FAR is called the AER. The EER is find the equal error rate between forgeries accepted and forgeries submitted number. The EER is also called the type III error and is defined as, 3. Hidden Markov Model This model is based on statistical learning theory capable to ab- sorb both the variability and the similarity between the patterns. HMMs are based on the empirical risk minimization (ERM) prin- ciple. It is the simplest of induction principles. It is based on de- cision rule and decision rule is based on a finite number for ex- ample training set.

Upload: others

Post on 30-Jun-2020

15 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Comparative Study of Offline Signature Verification Techniques · Comparative Study of Offline Signature Verification Techniques Devshri Satyarthi1, Y.P.S. Maravi2, Poonam Sharma3,

International Journal of Advancements in Research & Technology, Volume 2, Issue2, Feb ruary-2013 1 ISSN 2278-7763

Copyright © 2013 SciResPub.

Comparative Study of Offline Signature Verification Techniques Devshri Satyarthi1, Y.P.S. Maravi2, Poonam Sharma3, R.K. Gupta4

Department of CSE & IT, Madhav institute of technology and science Gwalior, 474002, India Email: [email protected], Tel +919944914443

ABSTRACT

Today in the world of profit aiming mind, trust of people is at high risk where forgery is monotonous in which fake signature has come into limelight. The fraudulent signature thus needs to be verified using verification techniques. The signature forgery can be restricted by either online signature verification or offline signature verification techniques. Online signature on one hand verifies the signature by performing a match with the pre-processed signature dynamically by detecting the motion of stylus during signature while on other hand, offline verifies by performing a match using the two dimensional scanned image of the signature. This paper studies about the various techniques available in offline signature verification along with their shadows.

Keywords : Signature verification, false rejection rate, Quality performance measures, Hidden markov model, Support Vector Machine, Neural network, Template Matching, Structural or syntactic, dynamic time warping.

1 INTRODUCTION In the real world instances authentication of an individual is desired frequently by distinguishing feature for person identi-fication. This authentication is generally offered by signature from past decades. Still in present, especially in financial, sig-nature is preferred as the demanding option for authentica-tion. Henceforth, with the increasing number of transactions, need of automatic signature verification will arise for authen-ticating the individual. Signature verification is an approach that is not only capable to find the identity addressing two individual efficiently but also strongly related tasks. On one side, the preliminary task is to identify the signature owner; on the other side, the secondary task is the decision, whether the signature is genuine or forged. Signature verification is spread into two categories: One is online signature verification and other is offline signature verification. Online signature verification dynamically scans the user sig-nature by tracing the motion on the stylus and verifies it against pre-processed signature information. On other hand, Offline signature verification gives static signature verification information, in which signature is scanned from document and verified against 2D scanned image of the signature. In this paper, the variations offered in offline signature verifi-cation schemes are discussed. The schemes vary from Hidden Markov model (HMM), Dynamic time warping (DTW), Sup-port vector machines (SVM), Neural network (NN), Wavelet transform to Structural or syntactic. 2. Quality Performance Measures

In evaluating the performance of a signature verification system, there are two important factors: the false rejection rate (FRR) of genuine signatures and the false acceptance rate (FAR) of forgery signatures and these two are inversely related.

The false rejection rate (FRR), the false acceptance rate (FAR), the equal error rate (EER) and the average rate (AER) are used as quality performance measures.

The FRR is the ratio of genuine test signatures rejected to the total number of genuine test signatures submitted. The FRR called the type I error and is defined as,

The FAR is the ratio of the number of forgeries accepted to the total number of forgeries submitted. The FAR is also called the type II error and is defined as,

The average of the FRR and FAR is called the AER.

The EER is find the equal error rate between forgeries accepted and forgeries submitted number. The EER is also called the type III error and is defined as,

3. Hidden Markov Model

This model is based on statistical learning theory capable to ab-sorb both the variability and the similarity between the patterns. HMMs are based on the empirical risk minimization (ERM) prin-ciple. It is the simplest of induction principles. It is based on de-cision rule and decision rule is based on a finite number for ex-ample training set.

Page 2: Comparative Study of Offline Signature Verification Techniques · Comparative Study of Offline Signature Verification Techniques Devshri Satyarthi1, Y.P.S. Maravi2, Poonam Sharma3,

International Journal of Advancements in Research & Technology, Volume 2, Issue2, Feb ruary-2013 2 ISSN 2278-7763

Copyright © 2013 SciResPub.

HMM consists of N states, where each state has two principle elements: first principle is the nature of a group of observations that is associated with the state, and the second principal is a his-togram which describes the making of a transition.

HMM’s efficiency, it uses cross-validation process for achieving best representative signature model from the database [3], and the reason for use these method is that it define automatic generated threshold in order of false accept rate (FAR) and false reject rate (FRR) signature. HMM’s ideally suited for these applications since the observations are time dependent. Offline signature con-tain no temporal information, the extraction of observations (that is time dependent) is less importance. Hidden Markov model work on different topology, the right topology chosen depending on find correct signature in the pattern set.

Signature modelling using HMM’s contains two phases: The first phase is cross-validation procedure used during the learning pro-cess and the second phase is verification procedure. The learning phase is to generate an HMM λ= {A, B, π} model that adequately characterizes each author signature model from the different writers [11].Cross-validation procedure is used to find optimal and dynamic solution for define the optimal state number for each specific signature model. In Verification phase, before the verification of a signature, the signature is transformed in to a sequence of observations using our feature extraction scheme. The verification process is basically made up of the Forward al-gorithm [11].

In order to define an HMM completely, the following elements are needed.

1. A set of N state, (S1...........SN) where qi is the state at time (t).

2. A set of K observation symbol, (V1..........Vk) where Ot is the observation at time (t)

3. A state transition probability matrix (A = Aij) where the probability of transition from state Si at time (t) to state Sj at (t+1) is a:

aij = P (qt+1 = Sj / qt = Si)

4. A set of output probability distribution B, where for each state j:

bj (k) = P (Ot = Vk / qt = Si)

5. An initial state distribution: π = (πi)

where, πp = P (qi = Si) [10].

In discrete models, two factors are important [4].The first is the number of states to be used. The second is the number of transi-tions among these states. The number of states depends on the

signature length hand the best results in terms of learning proba-bility [1].

El-Yacoubi et al. Use HMMs and the cross-validation principle for random forgery detection. A grid is superimposed on each signature image, segmenting it into local square cells. From each cell, the pixel density is computed, so that each pixel density rep-resents a local feature. Each signature image is therefore repre-sented by a sequence of feature vectors, where each feature vec-tor represents the pixel densities associated with a column of cells. The cross-validation principle involves the use of a subset (vali-dation set) of each writer’s training set for validation purposes. Since this system only attempts to detect random forgeries, sub-sets of other writers’ training sets are used for impostor valida-tion. Two experiments are conducted on two independent data sets. These data sets contain the signatures of 40 and 60 writers respectively. Both experiments use 20 genuine signatures (from each writer) for training and 10 for validation. Both experiments use the forgeries of the first experiment for impostor validation. Each test signature is analyzed under several resolutions and the majority-vote rule is used to make a decision. AERs of 0.46% and 0.91% are reported for the respective data sets.

4. Support Vector Machine

SVM is proposed. SVM is a new learning method introduced by V. Vapnik et al. SVMs are very universal learners. With a set of examples from two classes, a SVM finds the hyper plane, which maximizes the distance from either class to the hyper plane and separates the largest possible number of points belonging to the same class on the same side [2]. SVM is based on the structural risk minimization principle (SRM). SVM becomes very popular because of its success in hand written digit recognition. SRM consists of two main principles: The first principal is to control the empirical risk on the training set and the second principle is to control the capacity of decision function used to obtain this risk value.

SVM consist of two classes: Linear separable and Classification problem. Linear separable find the hyper plane with maximum Euclidean distance from the training set. There will be just one optimal hyper plane with the maximal margin d, defined as the sum of distances from the hyper plane to the closest points of the classes. This linear classifier threshold is the optimal separating hyper plane [8].

SVM is to solve the following optimization problem

Subject to yi (wT ϕ (xi) + b) ≥ 1 – ξi, C > 0, and ξi ≥ 0

Where, the original input space is mapped into a higher dimen-sional feature space by the function Φ [8].

Page 3: Comparative Study of Offline Signature Verification Techniques · Comparative Study of Offline Signature Verification Techniques Devshri Satyarthi1, Y.P.S. Maravi2, Poonam Sharma3,

International Journal of Advancements in Research & Technology, Volume 2, Issue2, Feb ruary-2013 3 ISSN 2278-7763

Copyright © 2013 SciResPub.

Then SVM can find an optimal linear separating hyper plane with the maximal margin in this higher dimensional feature space. C is the penalty parameter of the error term. For a two-class problem, the nonlinear decision function derived from the SVM classifier can be formulated as

Where is called the kernel function [8].

A SVM with GDTW applied to a handwriting recognition task achieved superior recognition rates compared to a conventional hidden Markov model (HMM) based method. The authors point out that the kernel is not positive definite but offer some theoreti-cal and empirical explanations to why it sometimes works well in practice [4].

5. Dynamic Time Warping

DTW is an algorithm for measuring similarity between two se-quences in respect of time or speed. DTW has been applied to video, audio, and graphics, one well known application has been automatic speech recognition, to cope with different speaking speeds.

The sequences to be analyzed by DTW are "warped" non-linearly in the time dimension to determine a measure of their similarity independent of certain non-linear variations in the time dimension.

The basic principle is to allow a range of 'steps' in the space of (time frames in sample, time frames in template) and to find the path through that space that maximizes the local match between the aligned time frames, subject to the constraints implicit in the allowable steps. The total similarity cost found by this algorithm is a good indication of how well the sample and template match, which can be used to choose the best-matching template.

Consider two signatures and . An alignment between L and L’ is represented by warping path, where T is the length of warp-ing path, and two warping function:

, mean that the -the point in L corresponds to the -the element in L’.

The matching cost between L and L’ along is defined as:

In our approach, is calculated by using online context. The optional warping path is the one which minimize the above cost function

The optimization problem of DTW can be solved efficiently us-ing dynamic programming.

6. Neural Network

NN used in signature verification for the purpose of hand written signature and verify its authenticity. But yeast day neural network mainly used in pattern recognition are their power and ease of use. The signature verification process parallels this learning mechanism. Neural network based on two processes: training and learning [9].The first step is training with the help of extract, a feature set representing the signature with several samples from different signers. The second step is learning the relationship between a signature and its class. Once this process is done then the network can be classified as belonging to a particular signer. NNs therefore are highly suited to modelling global aspects of handwritten signatures.

Alan McCabe et al. proposed a method for verifying handwritten signatures by using NN architecture. Various static (e.g., height, slant, etc.) and dynamic (e.g., velocity, pen tip pressure, etc.) signature features are extracted and used to train the NN. Several Network topologies are tested and their accuracy is compared. The resulting system performs reasonably well with an overall error rate of 3.3% being reported for the best case. Rasha Abbas in his earlier research investigated the suitability of using back propagation neural networks for the purpose of offline signature verification however later on in the suitability of using multilayered feed forward neural network was investigated.

Fig.1 Neural network architecture [11].

7. Structural or Syntactic

SS pattern recognition is the representation of pattern by symbol-ic data (Signatures etc) structure such as string, trees, and graphs. Symbolic representation is compared with a number of proto-types stored in a database. Structural feature use modification direction and transition distance feature (MDF) which extracts the transition locations and are based on the relational organiza-tion of low-level features into higher- level structures. The Modi-fied Direction Feature (MDF) [11]. Utilizes the location of transi-tions from background to foreground pixels in the vertical and horizontal directions of the boundary representation of an object.

y1

y2

yn

Σ f

Weights

W1j

W2j

Wnj

Bias

j

Activation function

Weighted Sum

Inputs (output from previous layers

Page 4: Comparative Study of Offline Signature Verification Techniques · Comparative Study of Offline Signature Verification Techniques Devshri Satyarthi1, Y.P.S. Maravi2, Poonam Sharma3,

International Journal of Advancements in Research & Technology, Volume 2, Issue2, Feb ruary-2013 4 ISSN 2278-7763

Copyright © 2013 SciResPub.

Nguyen et al presents a new method in which structural features are extracted from the signature's contour using the (MDF) and its extended version: the Enhanced MDF (EMDF) and further two neural network-based techniques and Support Vector Ma-chines (SVMs) are investigated and compared for the process of signature verification. The classifiers were trained using genuine specimens and other randomly selected signatures taken from a publicly available database of 3840 genuine signatures from 160 volunteers and 4800 targeted forged signatures. A distinguishing error rate (DER) of 17.78% was obtained with the SVM whilst keeping the false acceptance rate for random forgeries (FARR) below 0.16%.

Huang and Yan use a signature database that contains the signa-tures of 53 writers. For each writer, 24 genuine signatures, of which eight are used for training and 16 for testing, as well as 144 skilled forgeries are submitted. The forgeries are either simu-lated or traced. Statistical models, which are based on the pixel distribution and structural layout of the signatures, are used for an initial classification. During the initial classification the system rejects 2.2% of the genuine signatures, accepts 3.6% of the for-geries and is undecided on 32.7% of the signatures. For these, “questionable” signatures, that is 32.7% of the data set, a struc-tural feature verification algorithm is evoked. This algorithm compares the detailed structural correlation between the input and reference signatures. The system rejects 31.2% of the questiona-ble genuine signatures and accepts 23.2% of the questionable forgeries. This implies that, for the combined classifier, an FRR of 6.3% and an FAR of 8.2% is achieved. 8. Template Matching A process of pattern comparison is called template matching. According to Deng developed a system that uses a closed contour tracing algorithm to represent the edges of each signature with several closed contour. The curvature data of the traced closed contours are decomposed into multi resolution signals using wavelet transforms. The zero-crossing corresponding to the cur-vature data are extracted as features for matching based on the accuracy of the feature extraction process is calculated. Matching is done through dynamic time warping [9].

Verification performance is affected by the variation of signature stroke widths and a registered signature selected from a set of reference samples in off-line signature verification using a pattern matching. Katsuhiko Ueda in [11] proposed the modified pattern matching method, which is independent of signature stroke width and the selection method of a registered signature for Japanese signature verification [8].

According to Fang et al. proposed two methods for the detection of skilled forgeries using template matching. The first method is used optimal matching of the one-dimensional projection profiles of the signature patterns and the second method is based on the elastic matching of the strokes in the 2-dimension signature.

(A) Authors (R) Representation (D) Description (V) Verification

Error rate (%)

FRR

FAR

EER /AER

(A)EI-Yacoabi et.al.(2000) (R) Superimposed gird (D )Pixel density (V) HMM

0.75 1.17

0.18 0.64

0.46 0.91

(A) (A) Justino et.al.(2001)

(R) Superimposed gird (B) (D) Pixel

distribu-tion, slant

(C) (V)HMM

2.8 1.4 2.1

(A) Guo et al.(2001) (R) Storke segments (D) Geometric, gray-levels (V) DTW

2 - -

(A) Baltzakls et.al.(2001)

(B) (R) Global ,grid, texture (C) (D)Geometric, pixel

density (D) (V) NN

3 - 9.8

(A) (A) Huang et al.(2002) (B) (R) Structure layout (C) (D) Structure correlation (D) (V) Statistical, structure

6.3 - -

(A) H.W>JI et al.(2005) (R)Wavelength trans-

form, Zero-crossing (D) Common benchmark (V) SVM

5.25 5 5.50

(A) Fang et al.(2003) (R) Profiles, ”elements” (D) Geometric (V) TM

- - 20.8 18.1 23.4

Page 5: Comparative Study of Offline Signature Verification Techniques · Comparative Study of Offline Signature Verification Techniques Devshri Satyarthi1, Y.P.S. Maravi2, Poonam Sharma3,

International Journal of Advancements in Research & Technology, Volume 2, Issue2, Feb ruary-2013 5 ISSN 2278-7763

Copyright © 2013 SciResPub.

Table1.Comparison between various tec. Bases on FRR, FAR and AER/ERR.

Fig. 2 Genuine and forged signatures from database.

PROS AND CONS OF HMM, SVM, DTW, NN, AND SS

1. HMM has been found as a powerful statistical technique which is applied to handwriting recognition and signa-ture verification. SVM is a new classification technique which has been applied with success in pattern recogni-tion application like face and speech recognition. DTW is a well established distance measure for time series.

2. The SVM is linear kernel performing better than the

HMM. The Kernel is a symmetric function. The Kernel function implicitly defines the feature space and the mapping.

3. Justino et.al. Performed a comparison between HMM and SVM classifiers in detection of random, simple and skilled forgeries by using a grid segmentation scheme.

4. A SVM with GDTW applied to a handwriting recogni-tion task achieved superior recognition rates compared to a conventional HMM based method.

5. The pair wise similarities and a dynamic programming algorithm calculated in DTW algorithms are used to find a path through the similarity matrix which makes the ac-cumulated similarity as large as possible; gives higher performance when applied to speaker-dependent speech recognition problem.

6. The standard DTW with linear predictive modelling and distortion measurements is equivalent to the probabilis-tic modelling except that it searches for the best transi-tion path to minimize the accumulative distortion, while the probabilistic technique sums the density along every possible path.

7. DTW and HMM are two well-studied non-linear se-

quence alignment (or, pattern matching) algorithm. Non-linear sequence alignment (or, pattern matching) has vast range of applications in DNA matching, string matching, speech recognition, etc.

8. DTW has been applied to mostly in speech recognition, since it’s obvious that the speech tends to have different temporal rate, and alignment is very important for a good performance.

9. DTW conditional probabilities are used instead of local distance in standard DTW, and transition probabilities instead of path costs. This actually is related to HMM.

10. In DWT to compute path Dist (W) is the distance (typi-

cally Euclidean distance) of warp path W, and Dist (wki, wkj) is the distance between the two data point indexes (one from X and one from Y) in the kth element of the warp path, while In HMM, Viterbi algorithm is used for searching the optimal state transition sequence, for a given observation sequence.

11. TM is suitable for rigid matching to detect genuine sig-nature however these methods are not very efficient in detecting skilled forgeries.

12. NN commonly used classifiers for pattern recognition problems. It is a promising approach as instead of train-ing the entire network only the three new small NNs (one for each set of features) needs to be trained when we add a new person or signature.

13. The structural techniques become tedious and hectic due to large training sets and computational efforts

Page 6: Comparative Study of Offline Signature Verification Techniques · Comparative Study of Offline Signature Verification Techniques Devshri Satyarthi1, Y.P.S. Maravi2, Poonam Sharma3,

International Journal of Advancements in Research & Technology, Volume 2, Issue2, Feb ruary-2013 6 ISSN 2278-7763

Copyright © 2013 SciResPub.

References

[1] J.R. Justino, F.V Bortolozzi, R. Sabourin, “A compari-son of SVM and HMM classifiers in the offline signa-ture verification”, Proceedings of Elsevier, Pattern Recognition Letters, vol. 26, no. 9, pp. 1377-1385, 2005.

[2] E.Özgündüz, T.Şentürk and M.E Karslıgil, “Off-Line Signature Verification and Recognition by Support Vec-tor Machine”, Proceedings of European Signal Pro-cessing, 2005.

[3] E.J.R. Justino , El Yacoubi, A. Bortolozzi, F. Sabourin, “An off-line signature verification system using hidden Markov model and cross-validation”, Pro-ceedings of IEEE Computer Graphics and image pro-cessing, pp. 105-112, 2000.

[4] E.J.R.Justino, F.Bortolozzi, R. Sabourin, “Off-line sig-nature verification using HMM for random, simple and skilled forgeries” Proceedings of Sixth International Conference on Document Analysis and Recognition, pp.1031–1034, 2000.

[5] B.Majhi, Y.S Reddy, D.P Babu, “Novel Features for Off-line Signature Verification”, Proceeding of Interna-tional Journal of Computers, Communications & Con-trol, vol. I, no. 1, pp.17-24, 2006.

[6] A.P.Shanker, A.N.Rajagopalan,“Offline signature verification using DTW”, Proceeding of Pattern Recognition Letters, vol. 28, pp. 1407–1414, 2007.

[7] J.W.Hong and H.Q.Hua, “Signature Verification Using Wavelet Transform and Support Vector Machine”, Pro-ceeding of Advances in Intelligent Computing, vol. 3644, pp. 671-678, 2005.

[8] M.S Arya and V. S Inamdar “A Preliminary Study on Various Off-line Hand Written Signature Verification Approaches”, Proceeding of International Journal of Computer Applications, vol. 1, no.9, 2010.

[9] S.A. Daram T. S. Ibiyemi ola, “Offline Signature

Recognition using Hidden Markov Model” Proceeding of International Journal of Computer Applications, vol. 10, no.2, 2010.

[10] S.N. Gunjal, M. Lipton, “Robust Offline Signature Ver-

ification Based on Polygon Matching Technique”, Pro-ceeding of International Journal of Emerging Technolo-gy and Advanced Engineering, vol. 1, 2011.

[11] Katsuhiko Ueda, “Investigation of Off-Line Japanese Signature Verification Using a Pattern Matching”, Proc. of the 7th ICDAR, 2003

.