Discriminative feature selection for on-line signature verification

Pattern Recognition 74 (2018) 422–433

Contents lists available at ScienceDirect

Pattern Recognition journal homepage: www.elsevier.com/locate/patcog

Discriminative feature selection for on-line signature verification Xinghua Xia a,∗, Xiaoyu Song a, Fangun Luan a, Jungang Zheng a, Zhili Chen a, Xiaofu Ma b a b

School of Information and Control Engineering, Shenyang JianZhu University, 110168, Shenyang, China Ruckus Wireless Inc., 350W Java Dr, Sunnyvale, CA 94089, USA

a r t i c l e

i n f o

Article history: Received 14 April 2017 Revised 13 September 2017 Accepted 19 September 2017 Available online 27 September 2017 Keywords: On-line signature verification Discriminative feature selection Factorial experiment design Orthogonal experiment design Signature alignment Signature curve constraint

a b s t r a c t On-line handwritten signatures are collected as real-time dynamical signals which are written on collective devices by users. Since individuals have different writing habits, consistent and discriminative features should be selected to distinguish genuine signatures from forged signatures. In this paper, two methods, which are based on full factorial experiment design and optimal orthogonal experiment design, are proposed for selecting discriminative features among candidates. To improve the robustness, consistency of feature is analyzed at first, and more consistent features are selected as candidates for discriminative feature selection. To reduce the influences of fluctuations caused by internal and external writing environments changes before verification, signatures are effectively aligned to their reference templates based on Gaussian mixture model. A modified dynamic time warping with signature curve constraint is presented for verification to improve the efficiency. Comprehensive experiments are implemented based on the data of the open access databases MCYT and SVC2004 Task2. Experimental results verify the effectiveness and robustness of our proposed methods. © 2017 Elsevier Ltd. All rights reserved.

1. Introduction As the requirements of information security and identity verification keep increasing, biometrics is gaining popularity as a more trustable alternative to password based security systems. On-line handwritten signature verification is one of the most acceptable technologies of biometrics due to the fact that handwritten signatures have long been established as the most widespread means of personal verification. On-line signatures are difficult to be imitated and forged because they are unique and consistent for a given period. Experimental results have indicated that the accuracy of online signature verification is not lower than other biometrics [1,2]. On-line signature verification could generally be divided into two groups, i.e., parametric approaches and functional approaches. In parametric approaches, signatures are represented by series of parameters or vectors. Several common parameters are used the most extensively, such as position, displacement, numbers of pen ups and pen downs, speed, acceleration, pen down time ratio, aspect ratio, etc. [3,4]. When functional approaches are concerned, signatures are usually characterized in terms of time functions, some of the most commonly used functions are position trajectory, velocity, acceleration, centripetal acceleration, pressure, direction of pen movement, azimuth angle and altitude angle, etc. [5,6]. Generally, the functional approaches would obtain higher ac∗

Corresponding author. E-mail address: [email protected] (X. Xia).

https://doi.org/10.1016/j.patcog.2017.09.033 0031-3203/© 2017 Elsevier Ltd. All rights reserved.

curacy and reliability because they contain more dynamic information [2,4,6]. But functional approaches often require heavy computation during the process of matching or dissimilarity evaluation and it is less efficient in most cases. In verification, the authenticity of test signature is evaluated by matching its features with those stored in knowledge base for a given user. There are some commonly used verification methods, such as template matching methods [7,8], statistics based methods [9,10] and structure based methods [11,12].

2. Related works On-line handwritten signatures are collected as real-time signals and presented as time series. By reasons of internal and external environment changes, there are fluctuations of size, location and rotation angle of signatures within the same user at different inputs. Moreover, signatures will not keep higher consistency for a long time since the writing habits and external environments change. In this study, it is necessary to reduce the influence of fluctuations caused by variances of size, location and rotation angle, which could worsen the performance of verification. Thus, it is very important to effectively align the test signatures to references before verification. Furthermore, consistent and discriminative features should be extracted and selected for reducing the influences of these fluctuations and distinguishing genuine signatures from skilled forgery signatures.

X. Xia et al. / Pattern Recognition 74 (2018) 422–433

Methods of signature alignment include size, location and rotation angle matching. In most of research works, signatures are aligned by size, location and rotation angle, respectively. In methods of size alignment, max-min normalization and z-score normalization [13,14] are the most used. Methods of location and rotation angle alignment are mainly mapping the location and rotation angle of test signatures onto the reference coordinate system. The location alignment commonly uses the initial point and signature centroid as reference point [15,16], while the rotation angle are aligned by minimum moment of inertia to reference system [17,18]. Recently, alignment methods based on Gaussian mixture model (GMM) are developed [19–22]. In these methods, one point set is treated as the GMM centroid with equal isotropic covariance, the alignment of two point sets is altered to the centroid of GMM matching by the maximum likelihood estimation, and the optimal solution can be obtained by the expectation maximization (EM) algorithm. Effective and discriminative feature extraction and selection are important for the performance of on-line signature verification. Consistent and discriminative features are analyzed and selected for on-line signature verification in [2,6,11,15,23,24]. Different methods and criterions are presented in relevant works. Different conclusions are concluded based on different criterions. Most of these methods are mainly based on consistencies which mainly analyze the inter-similarities for a given user. To achieve higher accuracy for on-line signature verification, more distinctive and discriminative features should be extracted and selected to distinguish the genuine signatures from forgeries. In recent studies, feature’s capabilities in distinguish between genuine and forged signatures are analyzed directly by using equal error rate (EER) values [25,26]. When template matching approaches are considered, dynamic time warping (DTW) is widely used [7,8,27]. DTW is a nonlinear optimization matching method. During the process of DTW matching, it allows the time axes being compressed or expanded of two signatures to obtain the minimum distance. However, heavy computation is one of the defects of DTW when sampled points included in signatures increases, which will decrease efficiency of on-line signature verification. Some researchers proposed modified methods to improve the efficiency of DTW [8,27–30]. Most of these works mainly emphasize on the data reduction, and some information of signature might be discarded during the verification. 2.1. Framework and motivations of the work This work emphasize on reducing the inconsistencies of signatures and improving the effectiveness. Discriminative features are selected for improving the accuracy of verification. Our proposed method consists of several components, i.e., preprocessing, feature extraction and selection, verification, as shown in Fig. 1. For a given user, the input signatures are collected by collective devices and are presented as dynamic time series. There might be noises, distortion and variation during signature acquisition caused by collective devices and writing habits. Moreover, by reasons of internal and external environments changes, there are fluctuations of size, location and rotation angle of signatures within the same user at different inputs. In preprocessing stage, signatures are preprocessed, including smoothing and alignment, to decrease the fluctuations which are caused by noises, distortion and variation of signatures. In the feature extraction and selection stage, original features are extracted empirically at first. For improving the accuracy and robustness, consistent and discriminative features should be selected to distinguish genuine signatures from skilled forgery signatures. Thus, there are two steps included in the selection stage. (1) Step 1: consistent features are selected from original features

423

to improve the robustness of the system; (2) Step 2: for improving the accuracy, discriminative features are selected among the consistent features. EER values are used as the performance indicator to evaluate the distinguishing quality for the features. Two methods of experiment designs are used respectively for discriminative features selection, i.e., factorial experiment design (FED) and orthogonal experiment design (OED), respectively. After feature extraction and selection, more consistent and discriminative features are selected for verification. For improving the efficiency of the system, an efficient method of signature matching based on DTW with SCC (signature curve constraint) is proposed for signature dissimilarity evaluation. Then, authenticity of test signatures can be judged accordingly. Contributions of our work can be mainly described from the following three facts. •

•

•

In order to reduce the interrupted noises and distortion, input signatures are fitted by cubic smoothing algorithm with fivepoint approximation. Signatures are aligned to their reference template based on Gaussian mixture model (GMM) to decrease influences caused by fluctuations of various size, location and rotation at different inputs. For improving robustness, more consistent features are selected as candidates for discriminative features selection. Then, discriminative features are selected among these consistent features by two methods of experiment design, respectively, i.e., full factorial experiment design (FED) and optimal orthogonal experiment design (OED). EER values are used directly to evaluate feature discriminative ability. Thus, features used in verification can be not only consistent within intra-class, but discriminative for inter-class to distinguish forgery signatures from genuine ones. For improving the efficiency, a modified dynamic time warping with signature curve constraint (DTW with SCC) is proposed. In DTW with SCC, features are not matched by DTW directly. Instead, features are matched with the location constraints, which are inherent in two matching signature curves. The dissimilarity of features between the test and reference signature is evaluated by DTW with SCC.

3. Preprocessing Prior to signature verification, the user should be familiar with the signature acquisition device and should be required to input signature skillfully. On-line signature is then captured and represented as dynamic time series through acquisition devices at fixed interval. There might be noises, distortion and fluctuations during the acquisition. To decrease influences caused by these noises and fluctuations, signatures should be preprocessed before verification. 3.1. Smoothing At first, signatures should be smoothed to reduce the interrupted noises and distortion. Set the signature curve ass(n ) = {(x(n ), y(n )), n = 1, 2, · · · , N}, each points included in signature can be fitted by cubic smoothing algorithm with five-point approximation, and the parameters can be estimated by the least square algorithm. Middle points of signature can be approximated

s (n1 ) = (−3s(n1 − 2 ) + 12s(n1 − 1 ) + 17s(n1 ) + 12s(n1 + 1 ) − 3s(n1 + 2 ))/35

(1)

where, n1 = 3, 4, · · · , N − 2. The first two points of signature can be approximated by

s (1 ) = (69s(1 ) + 4s(2 ) − 6s(3 ) + 4s(4 ) − s(5 ))/70

(2)

s (2 ) = (2s(1 ) + 27s(2 ) + 12s(3 ) − 8s(4 ) + 2s(5 ))/35

(3)

424

X. Xia et al. / Pattern Recognition 74 (2018) 422–433

Fig. 1. The framework of discriminative feature selection for on-line signature verification.

Fig. 2. Signature curve fitting by cubic smoothing, where the left signature is the original signature, and the right is the signature after smoothing.

The last two points of signature can be approximated by

s (N − 1 ) = (2s(N − 4 ) − 8s(N − 3 ) + 12s(N − 2 ) + 27s(N − 1 ) + 2s(N ))/35

(4)

s (N ) = (−s(N − 4 ) + 4s(N − 3 ) − 6s(N − 2 ) + 4s(N − 1 ) + 69s(N ))/70

of reference signature sref (n) through the transformation  (stest (m), θ ). At the optimum, the test signature stest (m) and reference signature sref (n) are aligned and the correspondence is obtained by the maximum of the GMM posterior probability. The GMM probability density function is given as,



(5)

An example of signature curve fitting by cubic smoothing algorithm with five-point approximation is shown in Fig. 2. 3.2. Signature alignment By reasons of internal or psychological and external environments changes, there are fluctuations of size, location and rotation angle of signatures within the same user at different inputs. These fluctuations might worsen the performance of verification. Thus, it is very important to effectively align test signatures to reference templates before verification. In our work, we propose a probabilistic method of signature alignment based on Gaussian mixture model (GMM-alignment). According to GMM-alignment, the size, location and rotation angle of test signature can be effectively aligned to reference templates. Let sre f (n ) = (xre f (n ), yre f (n )) and stest (m ) = (xtest (m ), ytest (m )) be reference template signature and test signature, respectively, where, n = 1, 2, · · · , N and m = 1, 2, · · · , M are sampled points included in sref (n) and stest (m). GMM-alignment considers two point sets of reference and test signatures as a probability density estimation problem, where, the reference is represented the centroid of GMM, and test signature is considered the data point set to be aligned. The test signature stest (m) becomes the GMM centroid



p sre f (n ) =

M 

    ωm p sre f (n )| stest (m ), θ , σ 2

(6)

m=1 2

−

sre f (n )−(stest (m ),θ )

2σ 2 where, p(sre f (n )|(stest (m ), θ ), σ 2 ) = 2π1σ 2 exp denotes the density distribution probability of Gaussian components. σ 2 is the equal isotropic covariance, assuming the equal membership probability for all GMM components. ωm denotes the  weight factors of uniform distribution, and M m=1 ωm = 1. For simplicity, set all points included in stest (m) to be equal weighted, i.e., ωm = 1/M. (stest (m ), θ ) = sc · R · stest (m ) + t is the transformation function applied to stest (m), where, θ = sc , R, t  is a set of transformation parameters, sc denotes the scaling factor, R is rotation matrix, and t denotes the translation vector. According to GMM-alignment, the test signature stest (m) can be maximally approximated to the reference sref (n) through the transformation for obtaining the maximum likelihood. The correspondence probability between stest (m) and sref (n) can be estimated by the joint probability distribution





P sre f (n )| stest (m ), θ



=

N 







p sre f (n )| stest (m ), θ , σ 2



(7)

n=1

where, P(sref (n)| (stest (m), θ )) is transformation parameter model with  (stest (m), θ ) and σ 2 . It can be solved by the maximum likelihood estimation or equivalently by minimizing the negative log-

X. Xia et al. / Pattern Recognition 74 (2018) 422–433 Table 1 Original features extraction. #

Description

f1 f2 f3 f4 f5 f6 f7 f8 f9 f10 f11 f12 f13 f14 f15

Position in x-direction: x(n) Position in y-direction: y(n) Absolute displacement: c(n) Absolute velocity: v(n) Linear velocity in x-direction: vx (n) Linear velocity in y-direction: vy (n) Absolute acceleration: a(n) Linear acceleration in x-direction: ax (n) Linear acceleration in y-direction: ay (n) Pressure: p(n) Centripetal acceleration: ac (n) Cosine of the angle between signature curve and x-axis: cos (α ) Cosine of the angle between absolute velocity and x-axis: cos (β ) Azimuth angle: Az(n) Altitude angle: Al(n)

signature, it is merely determined as genuine signature or not. The genuine signature is considered as belong to subject user, and the forgery signature is considered as belong to any other subject user. Actually, there is only one cluster of signature verification. To improve the performance of verification, more consistent features should be selected accordingly as the candidates for discriminative feature selection. 4.3. Discriminative features selection based on FED

likelihood function,

     stest (m ), θ , σ 2  =−

N 

log

n=1

M 







P (m ) p sre f (n )| stest (m ), θ , σ

  2

(8)

m=1

To obtain the best matching of test signature stest (m) being aligned to reference signature sref (n), it should find the optimal transformation  (stest (m), θ ) to minimize the ( (stest (m), θ ), σ 2 ). Obviously, it is a nonlinear function of ( (stest (m), θ ), σ 2 ) about  (stest (m), θ ), and the optimum solution can be estimated by Expectation Maximization (EM) algorithm. 4. Discriminative features selection 4.1. Original feature set extraction For discriminative features selection, 15 features are extracted subjectively as original feature set in our work, i.e., F _Base = { f1 , f2 , f3 , f4 , f5 , f6 , f7 , f8 , f9 , f10 , f11 , f12 , f13 , f14 , f15 }. Out of these, feature f1 ,f2 ,f10 ,f14 and f15 can be obtained directly by signature acquisition devices, and rest 10 features can be extracted by simple mathematical computation. Table 1 show original features in our work. 4.2. Consistent features selection Given a user subject, there are generally few reference signatures could be obtained, the consistency of feature becomes very important. There are many potential features to be chosen. Thus, we are facing an increasing demand for a consistency model. The consistency of feature can be simply measured by

consist fk =

μdiss (g, g) − μdiss (g, f ) k k

σ 2 dissk (g, g) + σ 2 dissk (g, f )

425

(9)

where, μdissk (g, g) and σ 2 dissk (g, g) are, respectively, the mean and variance dissimilarity of feature fk computed from genuine signatures of user, while μdissk (g, f ) and σ 2 dissk (g, f ) are, respectively, the mean and variance dissimilarity of feature fk computed from corresponding forgeries of the user. consist fk considers both intra-class and inter-class distribution, and the consistency of feature fk can be simply measured. The weight of inter-class distribution is considered the same as intraclass in above consistency model, which means that the importance of forgery signatures is the same as genuine signatures in training process. In fact, signature verification is a special twocategory classification problem, i.e., genuine or forgery. For the test

To improve the accuracy of on-line signature verification, discriminative features should be selected to distinguish genuine signatures from the forgeries. In our work, EER values are used directly to evaluate the performance of features effect on verification, and discriminative features are selected based on factorial experiment design (FED). According to FED [31], multiple input variables have different impacts on the output response for a given system. The variables, which have greater impacts on outputs, are called main effect factors. In our work, discriminative features are considered as main effect factors which can improve the performance of verification. Concisely, for a given on-line signature verification system, if the EER value was decreased with feature fk being used, the performance of the system is improved, and then the feature fk can be considered as a main effect factor or discriminative feature. Otherwise, if the EER value was increased with feature fk being used, the performance of the system is worsen and the feature fk should be discarded. As discussed above, consistent features are selected from the original feature set F_Base, which will be used as candidate features for discriminative features selection. We set these consistent features as F¯ _Base for convenience. For improving performance of on-line signature verification, discriminative features should be selected from F¯ _Base to distinguish the genuine from forgery signatures. Similarly, we set these discriminative features as F¯ _Set in subsequent discussions. In our work, the discriminative features are selected by sequential forward selection (SFS) [32]. Feature which satisfies the criterion is selected step by step until F¯ _Set is full filled. To implement SFS, the optimal feature number included in F¯ _Set should be determined at first. For on-line signature verification, lower EER values indicate higher accuracy and better performance. Different EER values are obtained by using different numbers of features. Thus, the optimal number of features can be determined depending on the one that outputs the lowest EER value. After the optimal feature number has been determined, the discriminative features should be selected according to the given criterion. In our work, discriminative features are selected depending on the contribution rate (CR). For analyzing the ability of feature impacts on the verification performance, two types of feature sets are employed, i.e., F¯ _PC fk _in and F¯ _PC fk _ex , which denote the feature sets with feature fk included and excluded, respectively. Then, values of EER fk _in and EER fk _ex can be obtained with F¯ _PC fk _in and F¯ _PC f _ex being used as input feature sets for verification. k

CR of feature fk is defined as

C R fk = EER fk _ex − EER fk _in

(10)

CR can indicate the ability of feature impacts on the performance of on-line signature verification. Concisely, the feature fk with higher C R fk value can provide higher contribution to improve the verification performance. Similarly, the feature fk with lower C R fk value can provide lower contribution. In worse cases, the feature fk with negative C R fk value will increase the EER value significantly and decrease the accuracy of on-line signature verification. Therefore, features with negative C R fk value should be discarded.

426

X. Xia et al. / Pattern Recognition 74 (2018) 422–433 Table 3 9 representative combinations of OA are selected for experimentation.

Table 2 Experimental design with three factors and four levels per factor. Levels

L1 L2 L3

Factors A: Circularity

B: Rectangularity

C: Sphericity

0.2 0.3 0.4

0.3 0.4 0.5

0.3 0.5 0.7

In practice, the discriminative features are selected by SFS depending on their C R fk values. Specifically, set F¯ _Set = φ at first. Then at each step, C R fk value of feature fk is calculated, and the feature with the maximum C R fk value is selected. Because the EER value of the system is obtained by the combined effects of all the features included in F¯ _Set, the value C R fk should be recomputed after each step. Feature sets F¯ _PC f _in and F¯ _PC f _ex should be rebuilt k

k

correspondingly according to (11) and (12) at each step in order to improve the robustness.

F¯ _PC fk _in = F¯ _Set ∪ F¯ 2_PC fk _in

(11)

F¯ _PC fk _ex = F¯ _Set ∪ F¯ 2_PC fk _ex

(12)

A

B

C

Results (similarity)

C1 C2 C3 C4 C5 C6 C7 C8 C9

0.2(1) 0.3(2) 0.4(3) 0.4(3) 0.2(1) 0.3(2) 0.3(2) 0.4(3) 0.2(1)

0.3(1) 0.4(2) 0.5(3) 0.3(1) 0.4(2) 0.5(3) 0.3(1) 0.4(2) 0.5(3)

0.3(1) 0.5(2) 0.7(3) 0.5(2) 0.7(3) 0.3(1) 0.7(3) 0.3(1) 0.5(2)

y1 = 0.56 y 2 = 0.78 y 3 = 0.52 y 4 = 0.85 y 5 = 0.72 y 6 = 0.65 y 7 = 0.88 y 8 = 0.75 y 9 = 0.92

designs. In general, when there are P factors and Q levels for each factor. There are in totalQP combinations. When P and Q become larger, it would be inefficient and even impossible to do all the combinations. Therefore, it is desirable to sample a small but representative set of combinations for experimentation. We will show, with the help of OED, one can predict the best combination by testing only few representative experimental cases. •

where, F¯ 2 denotes the feature set with F¯ _Set having been removed from F¯ _Base. F¯ 2_PC fk _in denotes the combination feature set with the feature fk included and with the feature set F¯ _Set having been removed from F¯ _Base simultaneously, i.e., fk ∈ F¯ 2_PC f _in k

and F¯ _Set ⊂ F¯ 2_PC fk _in . Similarly, F¯ 2_PC fk _ex denotes the combination feature set with the feature fk excluded and with the feature set F¯ _Set having been removed from F¯ _Base simultaneously, i.e., fk ∈ / F¯ 2_PC fk _ex and F¯ _Set ⊂ F¯ 2_PC fk _ex .

Orthogonal Array: The OED method works on a predefined matrix table called and orthogonal array (OA). An OA with P factors and Q levels for each factor is represented by LV (QP ), where L denotes the OA and V is the number of combinations of the test cases. It has V rows, where every row represents a combination of levels. The L9 (34 )OA is given in (13), in which, there are four factors, three levels for each factor, and nine combinations of test cases.

⎡

1 ⎢1 ⎢1 ⎢ ⎢2 ⎢ L9 (34 ) = ⎢2 ⎢2 ⎢ ⎢3 ⎣ 3 3

4.4. Discriminative feature selection based on OED As discussed above, discriminative features can be selected based on full FED, but it is inefficient because features included in F¯ _Base are combined in every possible permutation to obtain a series of combination feature sets for evaluating the CR at each step. For improving the efficiency of discriminative features selection, we use the orthogonal experimental design (OED) to form an orthogonal learning strategy for variant to discover discriminative features. Owing to OED’s good character of sampling a small number of well representative combinations for testing set, the orthogonal learning strategy can construct a more promising and efficient candidate solution [31,33,34]. The OED can offer an ability to discover the best combination levels for different factors with a reasonably small number of experimental samples. Therefore, the OED is used to form an orthogonal learning strategy for discriminative feature selection to improve the performance of on-line signature verification. Owing to the OED’s orthogonal test ability, the orthogonal learning strategy can bring better learning efficiency, which results in faster speed and better global optimization performance. We use an example to introduce the basic concept of OED methods. Assuming in one handwritten character recognition, the similarity of a given character is evaluated depending on features, such as circularity (A), rectangularity (B), and sphericity (C). These three quantities are called the factors of the experiment. Each factor could have three possible values which are shown in Table 2, thus, each factor has three levels. All values of features are normalized for convenience of illustration. The aim is to find the best level combination of the three factors to obtain the maximum similarity for this handwritten character recognition. We can do the full factorial experiment for every combination and then select the best one. In this way, there are 33 = 27 combinations of experimental

Combinations

1 2 3 1 2 3 1 2 3

1 2 3 2 3 1 3 1 2

⎤

1 2⎥ 3⎥ ⎥ 3⎥ ⎥ 1⎥ 2⎥ ⎥ 2⎥ ⎦ 3 1

(13)

The L9 (34 ) OA has four columns, which means that it is suitable for the problems with at most four factors. As any sub-columns of an OA is also an OA, any arbitrary three columns of L9 (34 )can be used for the OED experiment. A simple example is selected to be used for handwritten character recognition experiment as shown in Table 3. •

Factor Analysis: The ability of predicting the best combination of levels can be given by factor analysis (FA). The FA is based on the experimental results of all the V test cases of the OA.

Let yv denote the experimental result of the vth (k = 1,2,…,V) combination and Spq denote the effect of the qth level (q = 1,2,…,Q) in the pth factor (p = A,B,C). The calculation of Spq is to add up all the total count of zvpq , where, set zvpq = 1 when the vth experimental test is with the qth level of the pth factor, otherwise, zvpq = 0. Specifically,

V S pq =

v=1 yv × zv pq

V

v=1 zv pq

(14)

In this way, the effect of each level on each factor can be evaluated and compared, as shown in Table 4. With all the Spq calculated, the best combination of the levels can be determined by selecting the level of each factor that provides the maximum Spq value. For a maximization problem, the larger Spq value is, the better the qth level on factor p will be. From Table 4, the best result is the combination of A2, B1, and C2, which illustrates that it can obtain maximum similarity with combination [A2, B1, C2] being

X. Xia et al. / Pattern Recognition 74 (2018) 422–433 Table 4 Factor analysis of the OED method. Levels

Factors analysis

L1 L2 L3 FA results

SA1 = 0.73 SA2 = 0.77 SA3 = 0.71 A2

SB1 = 0.76 SB2 = 0.75 SB3 = 0.70 B1

The dissimilarity of feature fk between reference template and test signature can be evaluated by

dist fk =

SC1 = 0.65 SC2 = 0.85 SC3 = 0.71 C2

implemented for handwritten character recognition. Although the combination itself does not exist in the nine combinations tested, it can be discovered by the factor analysis process. For improving the performance of on-line signature verification, the orthogonal learning strategy is used for discriminative feature selection based on OED. The candidate features are regarded as factors, and thus, there are D factors in the OED. This results in V = 2log2 (D+1 ) orthogonal combinations because there are two levels per factor, i.e., feature included and feature excluded, respectively. Therefore, V is no larger than 2D, which is significantly smaller than the total number of combinations 2D . Then, CR of feature can be evaluated according to (10). With factors analysis, discriminative features can be selected accordingly. 5. Verification When template matching approaches are considered, DTW is commonly used for signature matching. In DTW algorithm, a full distance matrix, which includes all distances between sampled points contained in two signatures, should be calculated at first. Then the optimal path is planned by dynamic programming to obtain the minimum distance. It can result in heavy computation and system inefficiency when sampled points included in signatures become larger. We introduce a modified method based on DTW with SCC for signature distance dissimilarity evaluation. In DTW with SCC, features are not matched by DTW directly. Instead, features are matched with the location constraints, which are inherent in two matching signature curves. The location constraint can be obtained by the optimal path, which is planned during the DTW matching of the two signature curves through dynamic programming. For a given user, test signature curve stest (m) is matched with reference signature curve sref (n) by DTW at first. Optimal matching path is obtained by dynamic programming, i.e., Wpath = {w1 , · · · , wr , · · · , wR }, where,wr = (wsre f (r ), wstest (r )) denotes the matching pairs on the DTW path, which is called the SCC in our work. wsre f (r )andwstest (r ) are matching points on the path included in reference tempalte and test signature curves, respectively. r = 1, 2, · · · , Rwhere R is the length of optimal path of DTW. During the DTW matching, there might be one-to-one matching points and one-to-many matching points. Specifically, the one-toone matching points are called direct matching points (DMP). Generally, DMP indicates higher similarity and more consistent between the two matching signatures. Consistent penalty factor (SPF) of DTW matching can be defined as

KDMP =

NDMP max(N, M )

427

(15)

where, KDMP denotes the SPF of DTW matching, NDMP is number of DMP in DTW matching, N and M are lengths of two signatures, respectively, NDMP ≤ min (N, M). Greater value of KDMP indicates the higher similarity between two signatures. With the optimal path of DTW having been planned, matching pairs on DTW path are obtained as wr , the length of optimal path R. Matching points on the path included in reference and test signature curves can be given as wsre f (r ) and wstest (r ), respectively.

1 KDMP

·

R 

| fk (wsre f (r )) − fk (wstest (r ))|

(16)

r=1

According to the discussions above, not only the complete information of signature but also the DTW dynamic programming is included in the proposed method of signature distance dissimilarity evaluation. The efficiency of on-line signature verification can be improved by proposed method. The computational complexity of proposed method is given by O(N · M + K · (N + M − 1 )) ∼ = O(N · M ) as opposed to O(K · N · M) of DTW, where, K denotes feature numbers used in on-line signature verification. 6. Experiments 6.1. On-line signature verification database Two standard on-line signature databases MCYT_Subcorpus_100 (DB1) [35] and SVC2004 Task2 [36] are used in our work, and several experiments are implemented to demonstrate the effectiveness and efficiency of our proposed method. DB1 and SVC 2004 Task2 databases are acquired with WACOM graphic tablet. Dynamic information of each signature is collected and stored, i.e., position in x-axis, position in y-axis, pressure applied by the pen, azimuth and altitude angle of the pen, timestamps and button status. DB1 consists of 50 0 0 western signatures from 100 users. For each user, there are 50 signatures in all. Out of these, 25 signatures are genuine and 25 signatures are skilled forgeries. Since there are no references given directly from DB1, 5 genuine signatures are selected randomly to be used as references, and other 20 genuine signatures and 25 skilled forgeries are used for verification tests. SVC2004 Task2 database consists 1600 signatures from 40 users, including western signatures and Chinese signatures. For each user, there are 20 genuine signatures and 20 skilled forgeries. Out of these, 5 genuine signatures are selected randomly to be used as references, and the other 15 genuine signatures and 20 skilled forgeries are used as test signatures. Thus, in our work there are 5900 test signatures from 140 users to be verified in total. The details of on-line signature verification databases used in our work are described as bellow Detailed description of DB1 • • •

No. of users: 100 No. of references: 100 × 5 = 500 (genuines) No. of test signatures: 100 × 20 = 2000 (genuines); 100 × 25 = 2500 (skilled forgeries) Detail description of SVC2004 Task2

• • •

No. of users: 40 No. of references: 40 × 5 = 200 (genuines) No. of test signatures: 40 × 15 = 600 (genuines); 40 × 20 = 800 (skilled forgeries)

6.2. Signature alignment 4 users from DB1 and SVC2004 Task2 are chosen randomly to illustrate signature alignment based on GMM. For each user, one reference signature is chosen as the template, while 10 genuine signatures and 10 skilled forgery signatures are also chosen randomly to be aligned to the template. Results of signature alignment

428

X. Xia et al. / Pattern Recognition 74 (2018) 422–433

Fig. 3. Signatures from 4 users are aligned to their reference template based on GMM-alignment, respectively. Genuine signatures alignment is demonstrated in the first row and skilled forgery signatures alignment is demonstrated in the second row.

are given in Fig. 3, where, signature in dash line denotes the reference template, and signatures in solid line are test signatures being aligned to the template. From the results, genuine signatures can be aligned to template completely, which will improve the accuracy of the verification. Besides, the skilled forgery signatures can be also aligned to template. The experimental results illustrate that test signatures can be matched well with the reference template by GMM-alignment. To improve the accuracy of genuine signature verification and reduce false accept error rate, discriminative features should be selected to distinguish genuine signatures from skilled forgery signatures in verification.

6.3. Discriminative feature selection 6.3.1. Consistent feature selection To improve the robustness of the system, consistent features should be selected and used for verification. In our work, the consistency of feature fk can be simply measured and evaluated considering both intra-class and inter-class distributions which depend on genuine signatures and skilled forged signatures, respectively. For a given user, the consistency of feature are measured and evaluated. In order to evaluate and compare the feature consistency validly, features should be normalized before dissimilarity measurement. In our experiments, values of features are normalized into (0,1). 5-fold cross validation is used to analyze and select high consistent features. Experiments are implemented on databases of DB1 and SVC2004 Task2, respectively, and experimental results are given in Table 5. From the results, the consistencies of the different features are various. To improve the system performance, higher consistent features should be selected and used. There are several candidate methods that can be chosen. In our work, for the ultimate purpose of improving the performance of on-line signature verification, discriminative features are selected among consistent features. Thus, 11 features with higher consistency are reserved as candidate features for discriminative feature selection in subsequent experiments. From the experimental results, consistent feature sets F¯ _BaseDB1 = { f1 , f2 , f3 , f4 , f5 , f6 , f7 , f9 , f10 , f11 , f14 } and F¯ _BaseSVC2004 = { f1 , f2 , f3 , f4 , f5 , f6 , f7 , f9 , f10 , f11 , f15 } can be se-

Table 5 Consistent feature selection. Feature

f1 f2 f3 f4 f5 f6 f7 f8 f9 f10 f11 f12 f13 f14 f15

DB1

SVC2004 Task2

Mean consist

Std consist

Meanconsist

Std consist

0.14 0.19 0.14 0.20 0.20 0.19 0.13 0.08 0.11 0.21 0.13 0.01 0.02 0.11 0.03

0.13 0.15 0.12 0.16 0.18 0.18 0.09 0.07 0.08 0.13 0.09 0.04 0.08 0.15 0.03

0.11 0.15 0.12 0.17 0.21 0.19 0.12 0.07 0.22 0.23 0.19 0.02 0.04 0.09 0.10

0.16 0.11 0.15 0.13 0.16 0.19 0.15 0.09 0.10 0.08 0.15 0.08 0.12 0.17 0.08

lected, which will be used as candidate features for discriminative feature selection. 6.3.2. Discriminative feature selection based on FED For discriminative feature selection, two methods are presented in our work, i.e., method 1: discriminative feature selection based on FED and method 2: discriminative feature selection based on OED. In this section, experiments are implemented on method 1 which is the full factorial experiment. To implement method 1 for discriminative feature selection which is based on SFS, the optimal feature number included in discriminative feature set F¯ _Set should be determined at first. For the input training set, features included in F¯ _Base are combined in every possible way to obtain a series of combination feature sets F¯ _PCNC , where, NC denotes the number of features included in F¯ _PCNC , and NC = 2, 3, · · · , 10. Thus, different values of EERF¯ _PC NC are obtained based on different feature sets F¯ _PCNC . The optimal feature number NF of F¯ _Setis determined by the number of features included in F¯ _PCNC with lowest EER ¯ . 5-fold cross valiF _PCNC

dation is adopted in experiments, and results are shown in Fig. 4. The boxplot illustrates that it will obtain lowest EER value when

X. Xia et al. / Pattern Recognition 74 (2018) 422–433

429

Fig. 4. The optimal feature number determination based on EER values, where the left boxplot is implemented on DB1 and right boxplot is implemented on SVC2004 Task2.

Table 6 5-Fold cross validation of discriminative feature selection based on FED (DB1).

Table 8 5-Fold cross validation of discriminative feature selection based on OED (DB1).

Fold No.

Feature set

μEER

σ EER

Fold No.

Feature set

μEER

σ EER

1 2 3 4 5

f2 , f2 , f2 , f2 , f4 ,

2.05 2.22 2.45 2.16 2.66

3.71 3.54 4.45 3.76 5.62

1 2 3 4 5

f4 , f4 , f4 , f4 , f2 ,

2.16 2.34 2.58 2.23 2.86

3.52 3.34 4.56 3.86 4.85

f5 , f5 , f6 , f5 , f5 ,

f6 , f6 , f8 , f6 , f6 ,

f10 , f11 f10 , f11 f10 , f11 f10 , f11 f9 , f10

Table 7 5-fold cross validation of discriminative feature selection based on FED (SVC2004 Task2).

f5 , f5 , f6 , f5 , f5 ,

f6 , f6 , f9 , f6 , f6 ,

f10 , f11 f10 , f11 f10 , f11 f10 , f11 f9 , f10

Table 9 5-fold cross validation of discriminative feature selection based on OED (SVC2004 Task2).

Fold No.

Feature set

μEER

σ EER

Fold No.

Feature set

μEER

σ EER

1 2 3 4 5

f5 , f2 , f5 , f2 , f5 ,

2.56 3.15 2.68 3.04 3.12

3.12 4.23 3.62 4.15 5.18

1 2 3 4 5

f5 , f5 , f4 , f2 , f2 ,

2.68 2.93 3.28 3.34 3.22

3.45 4.38 3.78 4.35 5.64

f6 , f5 , f6 , f6 , f6 ,

f9 , f6 , f9 , f9 , f8 ,

f10 , f11 f10 , f11 f10 , f11 f10 , f11 f9 , f10

5 features combination were used as input feature set for verification. The results demonstrate that it can obtain the highest performance with 5 features included in the input feature set of the on-line signature verification. Thus, we set feature number NF = 5 in our work. After the optimal feature number having been determined, the optimal features are selected according to CR values. For training set, CR values of features are evaluated based on FED. Experiment will be executed step by step to select discriminative features, and the feature with max C R fk value is selected at each step until F¯ _Set is fulfilled. Experimental results based on 5-fold cross validation are given in Tables 6 and 7. By counting the frequency of discriminative features occurred at each fold experiment, feature sets F¯ _SetDB1 = { f2 , f5 , f6 , f10 , f11 }and F¯ _SetSVC2004 = { f5 , f6 , f9 , f10 , f11 }are selected as discriminative features. 6.3.3. Discriminative feature selection based on OED Discriminative features can be selected based on full FED, but it is inefficient since features are combined in every possible permutation for feature selection based on their CV values. For improving the efficiency, an alternative method based on OED is presented for discriminative feature selection in our work. The OED can offer an ability to discover the best combination levels for different factors with a reasonably small number of experimental samples. According to OED, OA should be constructed at first since experiments are implemented on OA. For discriminative feature selection, the candidate features are regarded as factors, and there are two levels of each factors, i.e., feature being used and feature be-

f6 , f6 , f6 , f5 , f5 ,

f9 , f9 , f9 , f6 , f6 ,

f10 , f11 f10 , f11 f10 , f11 f9 , f10 f10 , f11

ing unused. As discussion above, consistent feature sets have been selected as F¯ _BaseDB1 = { f1 , f2 , f3 , f4 , f5 , f6 , f7 , f9 , f10 , f11 , f14 }and F¯ _BaseSVC2004 = { f1 , f2 , f3 , f4 , f5 , f6 , f7 , f9 , f10 , f11 , f15 }, which are used as candidate features. Thus, the OA can be constructed as 11 factors with two levels per factor, i.e., L12 (211 ).

⎡1

⎢1 ⎢1 ⎢ ⎢1 ⎢1 ⎢ ⎢1 L12 (211 ) = ⎢ ⎢2 ⎢ ⎢2 ⎢2 ⎢ ⎢2 ⎣ 2 2

1 1 1 2 2 2 1 1 1 2 2 2

1 1 2 1 2 2 2 2 1 2 1 1

1 1 2 2 1 2 2 1 2 1 2 1

1 1 2 2 2 1 1 2 2 1 1 2

1 2 1 1 2 2 2 2 1 1 1 2

1 2 1 2 1 2 2 1 2 2 1 1

1 2 1 2 2 1 2 1 2 2 1 1

1 2 2 1 1 2 1 1 2 2 1 2

1 2 2 1 2 1 2 1 1 1 2 2

⎤

1 2⎥ 2⎥ ⎥ 2⎥ ⎥ 1⎥ 1⎥ ⎥ (17) 1⎥ ⎥ 2⎥ 1⎥ ⎥ 2⎥ ⎦ 2 1

where, “1” denotes the feature being used, and “2” denotes the feature being unused. The feature CR value can be measured based on OED, and discriminative features are selected accordingly. 5-fold cross validation is also adopted, and experimental results are shown in Tables 8 and 9, which are implemented on databases DB1 and SVC2004 Task2, respectively. 5 most discriminative features are selected at each fold experiment, since optimal feature number has been determined as NF = 5 in our work. Similarly, by counting the frequency of discriminative features occurred at each fold experiment, feature sets F¯ _SetDB1 = { f4 , f5 , f6 , f10 , f11 }and F¯ _SetSVC2004 =

430

X. Xia et al. / Pattern Recognition 74 (2018) 422–433

{ f5 , f6 , f9 , f10 , f11 } are selected as discriminative features. For database DB1, featuref4 is selected as discriminative feature by OED while featuref2 is selected by FED. Other discriminative features f5 , f6 , f10 , f11 selected by OED are the same as the discriminative features that are selected by FED. For database SVC 2004 Task2, discriminative features selected by OED are completely the same as features that are selected by FED. From the results, the most discriminative features can also be selected by OED other than FED. But the efficiency of OED is much higher than FED, because OED can discover the best combination levels for different features with a reasonably small number of experimental samples. 6.3.4. Efficiency analysis As discussed above, discriminative features can be selected based on full FED, but it is inefficient because features included in candidate feature set are combined in every possible permutation to obtain a series of combination feature sets for evaluating the CR at each step. For example, there are 211 = 2048 combinations for discriminative feature selection at the first step. Then, at the second step, there are 210 = 1024 combinations since the most discriminative feature has been selected and removed from candidate feature set. In our experiment, for full steps of discriminative feature selection with full FED method being implemented, there are 3968 candidate feature combination sets being used and they need 943.54 s time consumption in total. Owing to OED’s good character of sampling a small number of well representative combinations for testing set, the orthogonal learning strategy can construct a more promising and efficient candidate solution. In our work, there are 12 candidate feature combination sets in total for 5 discriminative features selecting, which are significantly smaller than the total number of candidate feature sets used in full FED method. Our experiment shows, it needs only 2.86 s time consumption when OED method is used, which is much faster than full FED method. Experimental results demonstrate that the orthogonal learning strategy of OED can bring better learning efficiency, which results in faster speed and better global optimization performance than full FED. 6.4. Verification To demonstrate the effectiveness of our proposed method, the authenticity of 5900 test signatures from 140 users are determined. Error rates of equal error rate (EER), false reject rate (FRR) and false accept rate (FAR) are adopted to evaluate the performance of online signature verification. Out of these, EER could indicate the security level of a given biometrics system. In our work, discriminative features are selected and used for on-line signature verification. Two methods are presented to select discriminative features among consistent feature sets. Discriminative features were selected by FED and OED, respectively, and features selected by OED are mostly the same as the features that are selected by FED. For convenience of experiments implementing on databases DB1 and SVC 2004 Task2, feature sets F¯ _SetDB1 = { f4 , f5 , f6 , f10 , f11 } and F¯ _SetSVC2004 = { f5 , f6 , f9 , f10 , f11 } are used for verification in our work. As discussed above, it would result in heavy computation if sampled points included in signatures were increased when DTW is used for dissimilarity evaluation, and the efficiency of system will be deteriorated. For improving the efficiency, features will not be matched by DTW. Instead, features are matched by DTW with SCC which is inherent in signature curves in our work. The optimal matching path of signature curves can be obtained by two signature curves matching during the process of DTW matching, as shown in Fig. 5.

Fig. 5. Signature curve matching path of DTW.

Fig. 6. Comparison of time consumption between DTW and DTW with SCC. Table 10 Performances of the system with different methods. Database

Methods

FRR(%)

FAR(%)

EER(%)

DB1

DTW DTW with SCC DTW DTW with SCC

2.19 2.16 2.43 2.32

2.33 2.18 2.87 2.88

2.26 2.17 2.65 2.60

SVC2004

Time consumption of DTW and our proposed modified DTW with SCC are shown in Fig. 6. When DTW with SCC is used, SCC inherent in signature curves should be extracted which is optimized by 2-dimension signature curves matching by DTW, and then features are matched by DTW with SCC. When one feature is used in verification, time consumption of DTW with SCC is 0.025 s which is a little more than 0.021 s of DTW. But as feature number being increased, time consumption of DTW is increased dramatically, but time consumption of modified DTW with SCC is increased slowly. As discussed above, five features are used in our on-line signature verification, time consumption of DTW is 0.088 s as opposed to 0.026 s of modified DTW with SCC. Performance of the system with FRR, FAR and EER measured in percentage for different methods of dissimilarity evaluation are given in Table 10. From the results, when experiments are implemented on DB1, EER = 2.17% can be provided by DTW with SCC as opposed to 2.26% which is provided by DTW. As for SVC2004

X. Xia et al. / Pattern Recognition 74 (2018) 422–433 Table 11 Performances of different alignment methods implemented on DB1 and SVC2004 Task2. Database

Methods

FRR (%)

FAR (%)

EER (%)

DB1

Size-alignment Center-alignment Centroid-alignment GMM-alignment Size-alignment Center-alignment Centroid-alignment GMM-alignment

4.62 3.00 3.22 2.16 4.88 4.02 3.57 2.32

5.22 4.08 3.08 2.18 5.34 4.88 3.15 2.88

4.92 3.54 3.15 2.17 5.11 4.45 3.36 2.60

SVC2004

431

Table 13 Comparative studies of state-of-the-art methods implemented on DB1. Literatures

Features

Method

EER (%)

Pascual-Gaspar [28] Sae-Bae [37] Lopez-Garcia [38] Fishcher [39] Hafs [40] Diaz [41] Proposed method

6 features x, y, r, θ , p 25 features 8 features IMFs ofx, y 6 features v, vx , vy , p, ac

VQ Statistical Distance DTW + GMM DTW EMD DTW DTW with SCC

2.46 4.02 2.74 3.94 2.23 13.56 2.17

Table 14 Comparative studies of state-of-the-art methods implemented on SVC2004 Task2. Table 12 Performances of different feature sets implemented on DB1 and SVC2004 Task2. Database

Feature sets

FRR (%)

FAR (%)

EER (%)

DB1

f2 , f1 , f8 , f4 , f2 , f1 , f8 , f5 ,

3.35 5.56 6.36 2.16 3.56 5.68 6.85 2.32

3.86 6.12 7.58 2.18 4.02 7.22 8.21 2.88

3.61 5.84 6.97 2.17 3.79 6.45 7.53 2.60

SVC2004

f4 , f5 , f6 , f10 f7 , f8 , f12 , f14 f12 , f13 , f14 , f15 f5 , f6 , f10 , f11 f4 , f5 , f9 , f10 f3 , f7 , f8 , f12 f12 , f13 , f14 , f15 f6 , f9 , f10 , f11

Task2, it can be provided EER = 2.60% by DTW with SCC as opposed to 2.65% by DTW. The results illustrate that our proposed modified DTW with SCC can provide higher accuracy for on-line signature verification.

Literatures

Features

Method

EER (%)

Porwik [9] Sharma [30] Diaz [41] Fallah [42] Khoh [43] Liu [44] Proposed method

discriminating features 5 features 6 features MFCCs y, v, vy , φ Coefficients of DCT vx , vy , ay , p, ac

PSO + PNN DTW + VQ DTW Neural Network DWT-DFT + SVM Sparse representation DTW with SCC

0.71 2.73 18.25 3.00 7.17 3.98 2.60

[41] seem higher than other methods, it should be noted that method proposed by [41] is based on only one real reference template signature. But, most other methods, including our proposed method, are mainly implemented based on 5 reference templates. Signature reference template analysis and selection is one of our major researches in the future. Comparative studies illustrate that our proposed method can obtain higher accuracy with discriminative features for the on-line signature verification.

6.5. Comparative studies To demonstrate the effectiveness of our proposed method of GMM-alignment, other state-of-the-art methods, including Sizealignment [13], Center-alignment [22] and Centroid-alignment [15], are implemented on DB1 and SVC2004 Task2 for comparison. Performances of different alignment methods are shown in Table 11 with FRR, FAR and EER in percentage values. Experimental results demonstrate that it can obtain the max similarity between the test signatures and reference templates after GMM-alignment. The system can provide lowest EER values of 2.17% and 2.60% by GMMalignment method implemented on DB1 and SVC2004 Task2, respectively. Performances of different feature sets implemented on DB1 and SVC2004 Task2 are shown in Table 12. In this comparative study, other three feature sets, including most consistent features, lowest consistent features, and random features, are selected as input features for verification. From Table 12, lowest EER values can obtain by discriminative feature set. Highest EER values are given by lowest consistent features. Experimental results demonstrate that the effectiveness of our proposed method and performance of online signature verification can be improved by discriminative feature selection. In order to demonstrate the effectiveness of our proposed method, we compare it with other state-of-the-art methods. We compare the performance of systems which are carried out on DB1 and SVC2004 Task2. The best-result methods carried out on DB1 and SVC2004 Task2 are taken for comparative studies. Performances of state-of-the-art methods implemented on DB1 and SVC2004 Task2 are given in Tables 13 and 14. From the results, our proposed method can provide lowest EER value among stateof-the-art methods that executed on DB1. Moreover, it can provide lower EER value among state-of-the-art methods executed on SVC2004 Task2 except [9]. From biometric point of view, it should be affirmative that the novel method provided by [9] and experimental results are inspiring. While, EER values provided by

6. Conclusions In this work, we presented a method of discriminative feature selection for on-line signature verification. To reduce the influences of fluctuations caused by internal and external writing environment changes, signatures were effectively aligned to their reference templates before verification. To enhance the accuracy of the on-line signature verification, discriminative features were selected and used in verification. Differing from other state-of-the-art methods, discriminative features were selected though two steps in our work. In the first step, more consistent features were selected which were used as candidates for discriminative features selection. In the second step, discriminative features were selected among consistent features based on FED and OED. To improve the algorithm efficiency, a modified DTW with SCC was presented. Open access signature databases MCYT-100 and SVC2004 Task2 were used in our work, and comprehensive experiments were implemented. Experimental results illustrated that the best matching could be obtained by our proposed GMM-alignment method. The best error rates EERDB1 = 2.17% and EERSVC2004 = 2.60% were provided, respectively, which demonstrated the effectiveness and robustness of our proposed method. Acknowledgements This work was supported in part by the National Natural Science Foundation of China (Grant no. 61602322) and Liaoning Province Education Department of China (Grant nos. L2015437; L2015455). References [1] R. Plamondon, S. Srihari, On-line and off-line handwriting recognition: a comprehensive survey, IEEE Trans. Pattern Anal. and Machine Intelligence 22 (1) (20 0 0) 63–84.

432

X. Xia et al. / Pattern Recognition 74 (2018) 422–433

[2] D. Impedovo, G. Pirlo, Automatic signature verification: the state of the art, IEEE Trans. Syst. Man Cybern. Part C 38 (5) (2008) 609–635. [3] D. Guru, H. Prakash, Online signature verification and recognition: an approach based on symbolic representation, IEEE Trans. Pattern Anal. Mach. Intell. 31 (6) (2009) 1059–1073. [4] J. Fierrez-Aguilar, L. Nanni, J. Lopez-Peñalba, An on-line signature verification system based on fusion of local and global information, in: Audio-and Video-Based Biometric Person Authentication, 2005, pp. 523–532. [5] L. Nanni, A. Lumini, A novel local on-line signature verification system, in: Pattern Recognit. Lett., 29, 2008, pp. 559–568. [6] L. Lee, T. Berger, E. Aviczer, Reliable online human signature verification systems, IEEE Trans. Pattern Anal. Mach. Intell. 18 (6) (1996) 643–647. [7] J. Mei, M. Liu, F. Wang, Learning a Mahalanobis distance-based dynamic time warping measure for multivariate time series classification, IEEE Trans. Cybern. 46 (6) (2016) 1363–1374. [8] M. Faundez-Zanuy, On-line signature recognition based on VQ-DTW, Pattern Recognit. 40 (3) (2007) 981–992. [9] P. Porwik, R. Doroz, T. Orczyk, Signatures verification based on PNN classifier optimised by PSO algorithm, Pattern Recognit. 60 (2016) 998–1014. [10] C. Gruber, T. Gruber, S. Krinninger, Online signature verification with support vector machines based on LCSS kernel functions, IEEE Trans. Syst. Man Cybern. Part B 40 (4) (2010) 1088–1100. [11] R. Doroz, P. Porwik, T. Orczyk, Dynamic signature verification method based on association of features with similarity measures, Neurocomputing 171 (2015) 921–931. [12] J. Lee, H. Yoon, J. Soh, Using geometric extrema for segment-to-segment characteristics comparison in online signature verification, Pattern Recognit. 37 (1) (2004) 93–103. [13] P. Thumwarin, J. Pernwong, T. Matsuura, FIR signature verification system characterizing dynamics of handwriting features, EURASIP J. Adv. Signal Process. 2013 (1) (2013) 1–15. [14] M. Fahmy, Online handwritten signature verification system based on DWT features extraction and neural network classification, Ain Shams Eng. J. 1 (1) (2010) 59–70. [15] M. Parodi, J. Gómez, Legendre polynomials based feature extraction for online signature verification. Consistency analysis of feature combinations, Pattern Recognit. 47 (1) (2014) 128–140. [16] J. Igarza, L. Gómez, I. Hernáez, Searching for an optimal reference system for on-line signature verification based on (x,y) alignment, in: Biometric Authentication, in: Lecture Notes in Computer Science, 3072, 2004, pp. 519–525. [17] M. Munich, P. Perona, Continuous dynamic time warping for translation-invariant curve alignment with applications to signature verification, in: The Proceedings of the Seventh IEEE International Conference on Computer Vision, 1, 1999, pp. 108–115. [18] A. Al-Shoshan, Handwritten signature verification using image invariants and dynamic features, in: Computer Graphics, Imaging and Visualisation, International Conference on, IEEE, 2006, pp. 173–176. [19] A. Myronenko, X. Song, Point set registration: coherent point drift, IEEE Trans. Pattern Anal. Mach. Intell. 32 (12) (2010) 2262–2275. [20] Y. Zheng, D. Doermann, Robust point matching for nonrigid shapes by preserving local neighborhood structures, IEEE Trans. Pattern Anal. Mach. Intell. 28 (4) (2006) 643–649. [21] B. Jian, B.C. Vemuri, Robust point set registration using Gaussian mixture models, IEEE Trans. Pattern Anal. Mach. Intell. 33 (8) (2011) 1633–1645. [22] X. Xia, Z. Chen, F. Luan, X. Song, Signature alignment based on GMM for on– line signature verification, Pattern Recognit. 65 (2017) 188–196. [23] K. Manjunatha, S. Manjunath, D. Guru, Online signature verification based on writer dependent features and classifiers, Pattern Recognit. Lett. 80 (2016) 129–136.

[24] S. Rashidi, A. Fallah, F. Towhidkhah, Similarity evaluation of online signatures based on modified dynamic time warping, Appl. Artif. Intell. 27 (7) (2013) 599–617. [25] A. Shakil, S. Ahmad, R. Anwar, M. Balbed, Analysis of the effect of different features’ performance on hidden Markov modeling based online and offline signature verification systems, in: Digital Image Computing: Techniques and Applications, IEEE Computer Society, 2008, pp. 572–577. [26] X. Song, X. Xia, F. Luan, On-line signature verification based on stable features extracted dynamically, IEEE Trans. Syst. Man Cybern. (2016), doi:10.1109/TSMC. 2016.2597240. [27] S. Rashidi, A. Fallah, F. Towhidkhah, Local features analysis of on-line signature using modified distance of DTW, Int. J. Comput. Methods 12 (3) (2015) 1550016. [28] J. Pascual-Gaspar, M. Faundez-Zanuy, C. Vivaracho, Fast on-line signature recognition based on VQ with time modeling, Eng. Appl. Artif. Intell. 24 (2) (2011) 368–377. [29] G. Gupta, R. Joyce, Using position extrema points to capture shape in on-line handwritten signature verification, Pattern Recognit. 40 (10) (2007) 2811–2817. [30] A. Sharma, S. Sundaram, An enhanced contextual DTW based system for online signature verification using vector quantization, Pattern Recognit. Lett. 84 (2016) 22–28. [31] D. Montgomery, Design and Analysis of Experiments, eigth ed., 2013, pp. 8–10. Environmental Progress & Sustainable Energy 32(1). [32] G. Chandrashekar, F. Sahin, A survey on feature selection methods, Comput. Electrical Eng. 40 (1) (2014) 16–28. [33] H. Zhan, J. Zhang, Y. Li, Orthogonal learning particle swarm optimization, IEEE Trans. Evol. Comput. 15 (6) (2011) 832–847. [34] F. Gao, Y. Liu, L. Huang, A novel artificial bee colony algorithm based on modified search equation and orthogonal learning, IEEE Trans. Cybern. 43 (3) (2013) 1011–1024. [35] J. Ortega-Garcia, J. Fierrez-Aguilar, D. Simon, M. Gonzalez, V. Espinosa, A. Satue, I. Hernaez, J. Igarza, C. Vivaracho, D. Escudero, Q. Moro, MCYT baseline corpus: a bimodal biometric database, IEE Proc. 150 (6) (2003) 395–401. [36] Y. Yeung, H. Chang, Y. Xiong, S. George, R. Kashi, T. Matsumoto, G. Rigoll, SVC2004: first international signature verification competition, in: Biometric Authentication, in: Lecture Notes in Computer Science, 3072, 2004, pp. 16–22. [37] N. Sae-Bae, N. Memon, Online signature verification on mobile devices, IEEE Trans. Inf. Forensics Secur. 9 (6) (2014) 933–947. [38] M. López-García, R. Ramos-Lara, O. Miguel-Hurtado, E. Cantó-Navarro, Embedded system for biometric online signature verification, IEEE Trans. Ind. Inf. 10 (1) (2014) 491–501. [39] A. Fischer, M. Diaz, R. Plamondon, Robust score normalization for DTW-based on-line signature verification, document analysis and recognition (ICDAR), in: 2015 13th International Conference on. IEEE, 2015, pp. 241–245. [40] T. Hafs, L. Bennacer, M. Boughazi, A. Nait-Ali, Empirical mode decomposition for online handwritten signature verification, IET Biometrics 5 (3) (2016) 190–199. [41] M. Diaz, A. Fischer, M. Ferrer, R. Plamondon, Dynamic signature verification system based on one real signature, IEEE Trans. Cybern. 99 (2016) 1–12. [42] A. Fallah, M. Jamaati, A. Soleamani, A new online signature verification system based on combining mellin transform, MFCC and neural network, Digital Signal Process 21 (2) (2011) 404–416. [43] W. Khoh, S. Thian, P. Ying, B. Andrew, Score level fusion approach in dynamic signature verification based on hybrid wavelet-Fourier transform, Secur. Commun. Netw. 7 (7) (2014) 1067–1078. [44] Y. Liu, Z. Yang, L. Yang, Online signature verification based on DCT and sparse representation, IEEE Trans. Cybern. 45 (11) (2015) 2498–2511.

X. Xia et al. / Pattern Recognition 74 (2018) 422–433

433

Xinghua Xia received the B.S. and M.Sc. degrees in Institute of Information and Control Engineering from Shenyang JianZhu University in 20 0 0 and 20 03, respectively. He is currently an associate professor with the Faculty of Information and Control Engineering, Shenyang Jianzhu University, Shenyang, China. His main researching interests include machine learning and pattern recognition. Xiaoyu Song received the B.S. degree in Harbin University of Science and Technology in 1983, and the M.Sc. degree in Harbin Institute of technology in 1988, the Ph.D. degree in University of Chinese Academy of Sciences in 2007. He is currently a professor with the Faculty of Information and Control Engineering, Shenyang Jianzhu University, Shenyang, China. His main research interests include pattern recognition and image processing, artificial intelligence optimistic algorithms and data mining. He is now a member of ACM. Fangjun Luan received the B.S. degree in department of computer science and application of Shenyang JianZhu University. He received the M.Sc. and Ph.D. degrees in department of Computational Mathematics from University of Jilin in 2003 and 2007, respectively. He is currently a professor with the Faculty of Information and Control Engineering, Shenyang Jianzhu University, Shenyang, China. His main researching interests include pattern recognition and intelligent building. He is now a member of Liaoning Association of Automation. Jungang Zheng received the B.S. and M.Sc. degrees in Institute of Information and Control Engineering from Shenyang JianZhu University in 20 0 0 and 2006, respectively. He is currently an associate professor with the Faculty of Information and Control Engineering, Shenyang Jianzhu University, Shenyang, China. His main researching interests include wireless sensor networks and intelligent building. Zhili Chen received the M.Sc. degree in Communication and Information Systems from Northeastern University, Shenyang, China, in 2007, and the Ph.D. degree in Computer Science from Aberystwyth University, Aberystwyth, U.K., in 2013. She is currently an associate professor with the Faculty of Information and Control Engineering, Shenyang Jianzhu University, Shenyang, China. Her current research interests include computer vision and image processing. Xiaofu Ma received the B.S. degree in electronics from Northwest University, Xi’an, China, in 2008, and the M.S. degree in computer science from Tongji University, Shanghai, in 2011, and the Ph.D. degree in electrical and computer engineering from Virginia Tech, Blacksburg, USA, in 2016. He is currently with Ruckus Wireless, Sunnyvale, CA, USA. His research interests include pattern recognition and machine learning algorithms, signal and image processing, advanced wireless technologies, and network protocol and optimization.