UTSig: A Persian Offline Signature Dataset

Transcription

1 UTSig: A Persian Offline Signature Dataset Amir Soleimani 1*, Kazim Fouladi 2, Babak N. Araabi 1, 3 1 Control and Intelligent Processing Center of Excellence, School of Electrical and Computer Engineering, University of Tehran, Tehran, Iran 2 School of Electrical and Computer Engineering, University of Tehran, Tehran, Iran 3 School of Cognitive Sciences, Institute for Research in Fundamental Sciences (IPM), Tehran, Iran * A.Soleimani.B@gmail.com This paper is a preprint of a paper submitted to IET Biometrics and is subject to Institution of Engineering and Technology Copyright. If accepted, the copy of record will be available at IET Digital Library Abstract: In this paper a Persian offline signature dataset is introduced. It has totally 8280 images, which consist of 27 genuine signatures, 3 opposite-hand signed and 42 simple forged samples for 115 authentic persons. We review main characteristics of popular offline signature datasets in the literature. Then we statistically show that Persian signatures in comparison with other signatures have straight and cursive lines which are less probable to cross themselves. Next we train 4 different setups of writer-dependent verification systems with SVM on our dataset and test 1 setup (one-class SVM) for some others. We used decision and likelihood based criteria to check verification systems. For the former, false rejection ratio (FRR) and false acceptance ratio (FAR) are employed. For the latter, we used cost of log likelihood ratio (c llr ) and its minimum value (c llr min ), which are information-theoretic measures. 1. Introduction Signature is one of the most widespread methods among other authentication approaches. It is simple, cheap and accepted by ordinary people, official organization and courts. However it has a number of drawbacks. For example: variations due to writing instrument, paper and physical conditions of its writer, possibility of being forged and necessity of existence of considerable samples to be authenticate. Researchers by automatic signature verification systems have tried to help Forensic Handwritten Experts (FHEs) in decision making or even to play in their roles. In the literature automatic signature verification systems, based on data acquisition approaches, are divided into two categories: offline and online. In offline, signatures are known as images while in online, in addition to the shape of the signature, temporal, pen orientation and pressure information are available too [1] [5]. It is obvious that the abundance and uniqueness of information in online mode, provide more accurate results [2] but online approach is relatively away from the nature of the signature, its simplicity and availability. Therefore although offline approach has had less accuracy, it will be more applicable if it reaches desirable results. Signatures are divided into three kinds based on authenticity: genuine (authentic); forgery and disguised (when author tries to show its signature as forgery). In the literature also different terms were used to distinguish forgery types. For instance in [6], authors divide it to, simple forgery: when forger has no 1

2 attempt to mimic a signature; random forgery: when forger use his signature instead of genuine one; freehand or skilled forgery: when forger tries to practice and simulate genuine signature as closely as possible. Also traced forgery is used in some papers with the meaning of exactly following a signature with an instrument [7], [8]. Moreover, authors in [9] know simple forgery as forging by ordinary people and skilled forgery as expert efforts. Altogether, in this paper we divide forgery into: 1) Random forgery: genuine signatures of other authors or signing without seeing the genuine sample(s); 2) Simple forgery: forging by ordinary people with relatively remarkable efforts; 3) Skilled forgery: the result of professional forgers. Signatures look different in distinct cultures [10], for instance, English signatures usually consist of reshaped writer name while Persian ones are cursive independent of signers name [11]. Therefore, in each area for automatic signature verification systems, which are data-driven, existence of a rich dataset relating to its culture is mandatory. The term of rich for signature datasets refers to their number of samples and participants, and variables involving in sample collecting procedures such as signing period, writing instruments, paper, provided space for signing, samples showed to forgers, forgers efforts, meta-data and etc. In this manner we tried to enter different variables as much as possible. An important challenge, which hinders development of automatic signature verification, is lack of common datasets. Hence there is a need to standardize signature datasets and test systems on all relevant datasets. In this paper we compare our dataset with other datasets which have required conditions to provide reasonable comparison. It must mentioned that the aim of this paper is not introducing new verification method, we just want to introduce a Persian offline dataset, show its characteristics and uniqueness, test 4 approach on our data and one approach on other datasets. For more information and details about signature verification background we address to surveys done in this field. Authors in [6] and [12] survey the literature up to 1988 and 1993, respectively. In [13], signature verification is descripted as an application of handwritten recognition. Authors in [14], survey the literature up to 2007 and more recent developments can be found in [10], [15] and [16]. The rest of this paper is organized as follows: in section 2, we review popular offline signature datasets, then in section 3 we introduce new Persian offline signature dataset. We allocate section 4 to show distinct characteristics of Persian signatures. In section 5, we train and test verification systems on our data and others. Conclusion follows in section 5 and finally we appreciate participants in collecting the dataset. 2. Offline Signature Datasets There are several offline signature datasets for a few languages or cultures. In this section we review main characteristics of more popular datasets in the literature. 2

3 Spanish dataset MCYT-75 [17], a sub-corpus of MCYT bimodal database [18], has 75 classes containing 15 genuine signatures and 15 forgeries contributed by 3 user-specific forgers. They acquired signatures by inking pen and paper over a pen tablet. To make forgeries, they gave forgers genuine images and asked them to imitate the shape and natural dynamics. GPDSsignature [19] has 160 classes with 24 genuine samples gathered in a single day and 30 forged samples imitated by 10 forgers form 10 genuine specimens. Forgers had a random genuine sample and enough time. The collection was built by black or blue ink on white papers with 2 different size boxes. GPDS-960 [9] contains 960 classes with 24 genuine signatures and 30 forgeries for each one. Genuine samples signed in a single day on papers with 2 different size boxes. Totally, forgeries were produced by 1920 people apart from genuine persons and each forger had 5 genuine samples from 5 specimens (one sample per specimen). ICDAR2009 signature competition [2] offline dataset, contains training and evaluations sets. For training they used NISDCC signature collection acquired in WANDA project [20]. It has 12 classes, for each one 5 genuine and 5 forged samples. Forged signatures were written by 31 forgers. Evaluation set collected in the Netherlands Forensic Institute (NFI) contains 100 classes. Each class has 12 genuine samples and 6 forgeries created by 4 forgers from 33 writers. A Persian signature dataset (FUM) was used in [21]. It has 20 different classes; each class contains 20 genuine and 10 forged signatures. In 4NSigComp2010 [22] La Trobe signature collection was used. Its training set is composed of 9 reference signatures by one author and 200 questioned signatures, including: 76 genuine; 104 simulated (forged) signatures written by 27 freehand forgers and 20 disguised samples by its own author. Genuine and disguised samples were signed over a week. In addition, the author wrote another 81 genuine signatures to be used in forgery process. To build forgeries, 3 samples from 81 samples were given to forgers and asked them to imitate without tracing in 2 ways: forging 3 times without practice and simulating 3 times after 15 times practice. The test set has 25 signatures of another person written during 5 days and 100 questioned samples, including: 3 genuine; 7 disguise and 90 simulated (forged) signatures written by 34 freehand forgers from lay persons and calligraphers. Both sets were written by ball-point pen and same paper. SigComp2011 [3] offline dataset contains Chinese and Dutch signature. Chinese and Dutch training set have: 235 and 240 genuine; 340 and 123 forged signatures totally for 10 authors, respectively. Furthermore respectively, 116 and 648 references; 120 and 648 questioned; 367 and 638 forged samples are provided for 10 and 54 different Chinese and Dutch authors in test sets. 3

4 In 4NSigComp2012 [23] a new dataset was used. It contains 3 authentic authors (classes) with 15 to 20 signatures as references and 100 to 250 questioned signatures, including: 20 to 50 genuine; 8 to 47 disguised and 42 to 160 forged samples. Genuine and disguised samples were collected during 10 to 15 days. The number of forgers varied from 2 to 31 and they had 3 to 6 authentic samples. Forgers were asked to imitate signature (without tracing) by pen, pencil, with and without practice. SigWiComp2013 [5] introduced new Dutch and Japanese offline signature dataset. Dutch one has 27 authentic persons who signed 10 times with arbitrary writing instruments during 5 days. For forgeries, 9 persons could use any to all of the supplied specimen signatures as model(s). In average, there are 36 forgeries for each class. Japanese dataset images were converted from online signatures on a tablet PC. It contains 30 classes that each one has 42 genuine samples generated in 4 days and 36 forgeries by 4 forgers. Some datasets with few changes in number of samples are now publicly available; we provided their statistics including our dataset (UTSig [24]) in Table 1. Table 1 Statistics of datasets which are available for us Dataset Authentic Authors Average Genuine Per Author Average Simple Forgery Per Author Average Disguised (Opposite-hand) Per Author Total Samples MCYT ICDAR FUM NSigComp SigComp2011 Dutch SigComp2011 Chinese NSigComp SigWiComp2013 Dutch SigWiComp2013 Japanese UTSig UTSig Dataset University of Tehran Persian offline signature dataset (UTSig) [24] consists of 8280 images from 115 sets. Each set belongs to one specific authentic person and has 27 genuine and 45 forged samples of his signatures. Fig.1 shows samples from 4 sets. Participants for collecting signatures were randomly selected from undergraduate and graduate students of University of Tehran and Sharif University of Technology. Their age was between 18 and 31 (average 24.14) and 90% of them were right-hand writers. 4

5 Participants wrote with arbitrary pens on A4-sized white forms in predetermined boxes (Fig.2). Then scanning procedure was done with 600 dpi resolution and forms were stored as 8 bit grayscale TIF files. Two preprocessing were done on images: manually removing considerably large artifacts (e.g. artifacts from bad printing) and a simple threshold noise removal that converted pixels with intensity higher than a threshold to pure white (i.e. pixels brighter than 237 were assigned to 255). Fig. 1. Four genuine samples from UTSig, their opposite-hand and forgeries (similarity scores were determined by forgers) 5

6 Fig. 2. Forms for data collecting a Genuine and opposite-hand form b Forgery form 3.1. Genuine Signatures a b To obtain genuine signatures, 115 participants were asked to sign 10 times on a form and repeat it for 3 days. In each day, first 9 signatures were genuine and the last one was an opposite-hand signed signature of its writer that can be used as forgery or disguise. Therefore totally 3105 genuine and 345 opposite-hand signatures were collected. The forms (Fig.2) used for this procedure, contained 10 boxes (6 different sizes): 9 for genuine and 1 for opposite-hand signatures. The reason for using different sizes is to provide natural condition for authentic changes. Also to keep the nature of signature; writes were free to sign either vertically or horizontally. Furthermore, the boxes sizes were large enough for Persian signatures but if a signature crossed the boxes boundaries another form was given to its author to repeat Forged Signatures UTSig [24] has 5175 forgeries divided into 3 categories. The first one is 3 opposite-hand signed signatures of who wrote their genuine ones. These are 345 samples. 6

7 The second category is simple forged samples that were obtained from 230 people apart from authentic ones. Each person forged 3 different signatures (from different classes) 6 times in forms with 6 same size boxes (Fig.2). Their observable genuine samples varied from 1 to 3 signatures of same class. In other word each set or genuine class was forged by 6 persons in a way that so-called 2 simple forgers saw 1 randomly selected sample of genuine signatures; 2 persons saw 2 and 2 persons saw 3 randomly selected samples. For each forged sample, the forger was asked to determine its similarity to genuine sample(s) (Fig.1) from very low to very high (1, 2, 3, 4 or 5). The average of scores is 2.72 (2.67, 2.71 and 2.79 when respectively 1, 2 or 3 genuine sample were observable).totally this category has 4140 simple forged samples and 91% of them (74% of total forgeries) were ranked by its so-called forger. The third category is also a simple forgery but its samples were forged by a more skilful person than two pervious categories. The form used was such as the second category and the observable sample for this category was only one random genuine sample. Altogether, 690 samples were built in this category. In all categories writers had enough time to practice and they were ask to be cautious about boxes boundaries but we manually refined samples crossed the boundaries (Fig.3). Fig. 3. An example for manual refinement a Original sample which crossed the boundaries b Refined sample a b 4. Persian Signatures Characteristic By observations we see that signatures in distinct cultures have different shapes. So, it seems clear to use different feature extraction and classification methods with respect to the different signatures. For instance in [25] and [26], authors use a two-stage approach to improve verification accuracy for multiscripts signatures. They first use an identification system to see whether questioned signature is Hindi or English then different verification systems are employed. To statistically compare signatures differences in available datasets we use morphological operations to count branch points and end points in each sample. First genuine samples are binarized with appropriate manual thresholds. Second after removing small objects from binary images, dilation operator is used. Third 7

8 we set a pixel to black if five or more pixels in its 3-by-3 neighbours are black (majority operation). Forth, skeletonization is done and finally branch points and end points are extracted by relating morphological operators. To remove unreal points caused by skeletonization operation, we just selected one point in small areas which had a number of points. Fig. 4 shows two branch points and two end points for a signature. Fig. 4. A signature with two branch points and two end points Analysis on the number of branch points and end points shows that Persian signature in UTSig [24] and FUM [21] datasets have less branch points and end points in comparison with other datasets, including: Dutch, Spanish, Chinese and Japanese signatures (see Fig. 5). This phenomenon can be used to state that Persian signatures contain straight or cursive lines which are less probable to cross themselves. In other word, if we use lines or curves to connect branch points and end points to rebuild a signature, Persian ones have less lines or curves in comparison with other datasets. Probability Density Fucntion UTSig FUM MCYT ICDAR2009 4NSigComp2010 SigComp2011 Dutch SigComp2011 Chinese 4NSigComp2012 SigWiComp2013 Japanese SigWiComp2013 Dutch Number of branch points and end points Fig. 5. Estimated probability density function (PDF) of number of branch points and end points in offline signature datasets 8

9 5. Verification Systems Writer-Independent (WI) and Writer-dependent (WD) are two major approaches to employ verification systems. In WI, all genuine samples and all forgeries, regardless of their authors, are compared pair-to-pair in training phase to compute distribution of within-writer and between-writer similarities (dissimilarities). Next, questioned signature is checked in both distributions to determine its authenticity. In contrast, WD approach needs to use author-based samples, in other word, for each authentic person, training phase is done separately by its author samples alone or versus other samples [27] Training Setups In this paper we focus on WD approach and use 4 setups for training on UTSig [24] dataset while only one setup is used for other datasets. 4 setups are: Genuine vs. random forgery (setup 1): training verification system for each author based on his genuine samples and random forgeries. Genuine vs. random forgery and opposite-hand (setup 2): training verification system for each author based on his genuine and opposite-hand samples and random forgeries. Genuine vs. opposite-hand (setup 3): training verification system for each author based on his genuine and opposite-hand samples. Genuine alone (setup 4): training verification system for each author with only his genuine samples. It must be mentioned that any use of simple forgery during training phase for WD approach is not applicable, since in real conditions, for instance in banks, collecting simple or skilled forgeries is exhausting procedure for both consumers and bankers Classifier We used Support Vector Machine (SVM) classifier with linear kernel for the first, second and third setup. For forth setup, which is a one-class problem, we used one-class SVM with radial basis functions (RBF) kernel. We also calculated probabilistic outputs [28] or scores for SVMs to determine output labels and their likelihood ratio Evaluation Methods In this work we follow the paradigm shift introduced in SigComp2011 [3] to use both decision and likelihood based criteria in signature verification systems. For the former we deal with false acceptance rate (FAR) and false rejection rate (FRR). For the latter, we refer to two information-theoretic measures, cost of 9

10 log-likelihood-ratio (c llr ) and its minimal possible value (c llr min ) [29]. c llr is a positive unbounded measure and c llr min is always between 0 and 1. The less these two criteria are the better performance, verification system has. These two information-theoretic were used in 4NSigComp2012 [23] and SigWiComp2013 [5] too. In SigWiComp2013 was shown that good equal error rate (EER) does not always result good c llr min [5] Feature Extraction We used fixed-point arithmetic which was described in [19] as: description of the signature envelope and the interior stroke distribution in polar and Cartesian coordinates. In fixed-point arithmetic feature extraction, using sample geometric centre, three parameters are calculated in polar coordinate: derivative of radius of signature envelope; its angle and the number of black pixels that the radiuses cross when rotating from one point to next point. Moreover, in Cartesian coordinates, height, width and the number of transitions form black to white or white to black pixels of signatures are calculated respect to their geometric centre too [19] Numerical Results To find numerical results of UTSig [24] dataset, as a default we randomly selected 10 genuine samples for an author to train a verification system relating to him. In each setup if it contained forged or opposite-hand signatures we respectively selected 1 genuine sample from each authors as random forgeries (totally 114 samples) and all 3 samples from his opposite-hand signatures. We repeated this procedure for each author 10 times and averaged their results. In test phase, other genuine and forged samples (simple forgery) of its author and all samples of other authors (random forgery) were checked by the verification system. Results are available in Table 2. Among setups, first and second one, are more accurate, forth setup also is in competition with them but third setup has disastrous simple forgery FAR. Note in forth setup due to the nature of one-class SVM, a threshold must be selected. Here we selected one similar ideal threshold for all authors but better results may be obtained by user-based threshold estimation. 10

11 Table 2 Setups results for UTSig Setup Genuine FRR Simple Forgery Random Forgery Gen vs. Simple Gen vs. Simple FAR FAR Forgery c llr min Forgery c llr % 31.7% 0.5% % 27.9% 0.4% % 88.6% 0.754% % 31.7% 5.9% The behaviour of the trained systems due to participant self-scores in forgery process, has a trend toward distinguishing low score samples more than high score ones. The average score of true negative samples is 2.65 while for false positive it is To check results with respect to number of samples or simple forgeries in training phase we repeated first setup (see Fig. 6). Percentage FRR FAR Percentage FRR FAR Number of Genuine Samples in Training Number of Random Forged Samples in Training a Fig. 6. FRR and FAR variation by a Number of genuine samples in training b Number of forged samples in training b For other datasets it seems reasonable to perform only the fourth setup, since other setups need a variety of random forgeries (from distinct authors) which there are not in all datasets. Moreover oppositehand samples are not available too. Note that some datasets due to unavailability, missed or low genuine sample were not used. Results are provided in Table 3. 11

12 Table 3 Datasets results for setup 4 (Genuine alone) Setup Genuine FRR Simple Forgery Random Forgery Gen vs. Simple Gen vs. Simple FAR FAR Forgery c llr min Forgery c llr MCYT % 25.9% 5.9% FUM 25.9% 25.8% 0.02% NSigComp % 29.0% 0% SigComp2011 Dutch 31.2% 31.4% 6.5% NSigComp % 22.5% 4.4% SigWiComp2013 Japanese 35.7% 33.7% 9.5% UTSig 29.8% 31.7% 5.9% Conclusion In this paper we introduced a Persian offline signature dataset, University of Tehran signature dataset (UTSig) [24]. It has 115 classes, each one consists of 27 genuine, 3 opposite-hand and 42 simple forged samples. We also presented conditions for collecting samples (e.g. collecting genuine samples in 3 days, different size boxes, arbitrary pens, opposite-hand signatures, self-scores, different observable samples for forgers and etc.). Then by counting branch points and end points of signatures in UTSig and other datasets we found that Persian signatures generally have straight and cursive lines which are less probable to cross themselves. Next, we checked 4 different training setups of writer-dependent (WD) verification systems on UTSig and 1 setup for some other datasets. We employed fixed-point arithmetic features and the best results for UTSig were: FAR=27.4%; FAR=31.7% for simple forgery; FAR=0.5% for random forgery; c llr=0.926 and c llr min =0.821 for genuine vs. simple forgery. Furthermore we found that if the number of genuine samples in training phased is increased, the genuine FRR drops and simple forgery FAR is increased. In contrast, increase in the number of forged samples has a reverse effect. 7. Acknowledgments We appreciate participants in sample collection procedure for their cooperation and permission to use and publish their signatures. 12

13 8. References [1] D.-Y. Yeung, H. Chang, Y. Xiong, S. George, R. Kashi, T. Matsumoto, and G. Rigoll, SVC2004: First international signature verification competition, in Biometric Authentication, Springer, 2004, pp [2] V. L. Blankers, C. E. van den Heuvel, K. Y. Franke, and L. G. Vuurpijl, ICDAR 2009 Signature Verification Competition, in th International Conference on Document Analysis and Recognition, 2009, pp [3] M. Liwicki, M. I. Malik, C. E. van den Heuvel, X. Chen, C. Berger, R. Stoel, M. Blumenstein, and B. Found, Signature Verification Competition for Online and Offline Skilled Forgeries (SigComp2011), in 2011 International Conference on Document Analysis and Recognition, 2011, pp [4] N. Houmani, a. Mayoue, S. Garcia-Salicetti, B. Dorizzi, M. I. Khalil, M. N. Moustafa, H. Abbas, D. Muramatsu, B. Yanikoglu, a. Kholmatov, M. Martinez-Diaz, J. Fierrez, J. Ortega-Garcia, J. Roure Alcobé, J. Fabregas, M. Faundez-Zanuy, J. M. Pascual-Gaspar, V. Cardeñoso-Payo, and C. Vivaracho-Pascual, BioSecure signature evaluation campaign (BSEC 2009): Evaluating online signature algorithms depending on the quality of signatures, Pattern Recognit., vol. 45, no. 3, pp , Mar [5] M. I. Malik, M. Liwicki, L. Alewijnse, W. Ohyama, M. Blumenstein, and B. Found, ICDAR 2013 Competitions on Signature Verification and Writer Identification for On- and Offline Skilled Forgeries (SigWiComp 2013), in th International Conference on Document Analysis and Recognition, 2013, pp [6] R. Plamondon and G. Lorette, Automatic signature verification and writer identification the state of the art, Pattern Recognit., vol. 22, no. 2, [7] J. K. Guo, D. Doermann, and A. Rosenfield, Off-line skilled forgery detection using stroke and sub-stroke properties, in Pattern Recognition, Proceedings. 15th International Conference on, 2000, vol. 2, pp [8] A. Mitra, P. K. Banerjee, and C. Ardil, Automatic Authentication of handwritten documents via low density pixel measurements, Int. J. Comput. Intell., vol. 2, no. 4, pp , [9] F. Vargas, M. Ferrer, C. Travieso, and J. Alonso, Off-line Handwritten Signature GPDS-960 Corpus., ICDAR, [10] S. Pal, M. Blumenstein, and U. Pal, Non-English and Non-Latin Signature Verification Systems A Survey., in AFHA, 2011, pp [11] A. Chalechale and A. Mertins, Line segment distribution of sketches for Persian signature recognition, in TENCON Conference on Convergent Technologies for the Asia-Pacific Region, 2003, vol. 1, pp [12] F. Leclerc and R. Plamondon, Automatic Signature Verification: The State of the Art , Int. J. Pattern Recognit. Artif. Intell., vol. 08, pp , [13] R. Plamondon and S. N. Srihari, Online and off-line handwriting recognition: a comprehensive survey, Pattern Anal. Mach. Intell. IEEE Trans., vol. 22, no. 1, pp ,

14 [14] D. Impedovo and G. Pirlo, Automatic Signature Verification: The State of the Art, IEEE Trans. Syst. Man, Cybern. Part C (Applications Rev., vol. 38, no. 5, pp , Sep [15] S. Pal, M. Blumenstein, and U. Pal, Off-line signature verification systems: a survey, in Proceedings of the International Conference & Workshop on Emerging Trends in Technology, 2011, pp [16] D. Impedovo, G. Pirlo, and R. Plamondon, Handwritten Signature Verification: New Advancements and Open Issues., in ICFHR, 2012, pp [17] J. Fiérrez-Aguilar, N. Alonso-Hermira, G. Moreno-Marquez, J. Ortega-Garcia, B. Schafer, and S. Viriri, An off-line signature verification system based on fusion of local and global information, in Biometric Authentication, Springer, 2004, pp [18] J. Ortega-Garcia, J. Fierrez-Aguilar, D. Simon, J. Gonzalez, M. Faundez-Zanuy, V. Espinosa, A. Satue, I. Hernaez, J.-J. Igarza, C. Vivaracho, and others, MCYT baseline corpus: a bimodal biometric database, IEE Proceedings-Vision, Image Signal Process., vol. 150, no. 6, pp , [19] M. a Ferrer, J. B. Alonso, and C. M. Travieso, Offline geometric parameters for automatic signature verification using fixed-point arithmetic., IEEE Trans. Pattern Anal. Mach. Intell., vol. 27, no. 6, pp , Jun [20] K. Franke, L. Schomaker, C. Veenhuis, C. Taubenheim, I. Guyon, L. Vuurpijl, M. Van Erp, G. Zwarts, M. van Erp, and G. Zwarts, WANDA: A generic Framework applied in Forensic Handwriting Analysis and Writer Identification., Des. Appl. hybrid Intell. Syst. Proc. 3rd Int. Conf. Hybrid Intell. Syst., pp , [21] M. R. Pourshahabi, M. H. Sigari, and H. R. Pourreza, Offline Handwritten Signature Identification and Verification Using Contourlet Transform, in 2009 International Conference of Soft Computing and Pattern Recognition, 2009, pp [22] M. Liwicki, C. E. Van Den Heuvel, B. Found, and M. I. Malik, Forensic Signature Verification Competition 4NSigComp Detection of Simulated and Disguised Signatures, th Int. Conf. Front. Handwrit. Recognit., no. Svc 2004, pp , Nov [23] M. Liwicki, M. I. Malik, L. Alewijnse, E. Van Den Heuvel, and B. Found, ICFHR 2012 Competition on Automatic Forensic Signature Verification (4NsigComp 2012), 2012 Int. Conf. Front. Handwrit. Recognit., no. 4NsigComp, pp , Sep [24] UTSig Dataset. [Online]. Available: [25] M. B. Srikanta Pal, Umapada Pal, A Two-Stage Approach for English and Hindi Off-line Signature Verification, in New Trends in Image Analysis and Processing ICIAP 2013, Springer Berlin Heidelberg, 2013, pp [26] S. Pal, M. Blumenstein, U. Pal, and M. Blumenstein, Multi-script Off-line Signature Verification: A Two Stage Approach., in AFHA, 2013, pp [27] S. N. Srihari, A. Xu, and M. K. Kalera, Learning strategies and classification methods for off-line signature verification, in Frontiers in Handwriting Recognition, IWFHR Ninth International Workshop on, 2004, pp

15 [28] J. Platt and others, Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods, Adv. large margin Classif., vol. 10, no. 3, pp , [29] N. Brümmer and J. du Preez, Application-independent evaluation of speaker detection, Comput. Speech Lang., vol. 20, no. 2, pp ,