Off-line Bangla Signature Verification: An Empirical Study

Transcription

1 Off-line Bangla Signatue Veification: An Empiical Study Autho Pal, Sikanta, Alieza, Alaei, Pal, Umapada, Blumenstein, Michael Published 2013 Confeence Title The 2013 Intenational Joint Confeence on Neual Netwoks (IJCNN), DOI Copyight Statement 2013 IEEE. Pesonal use of this mateial is pemitted. Pemission fom IEEE must be obtained fo all othe uses, in any cuent o futue media, including epinting/epublishing this mateial fo advetising o pomotional puposes, ceating new collective woks, fo esale o edistibution to seves o lists, o euse of any copyighted component of this wok in othe woks. Downloaded fom Link to published vesion Giffith Reseach Online

2 Off-line Bangla Signatue Veification: An Empiical Study Sikanta Pal 1, Alieza Alaei 2, Umapada Pal 3, Michael Blumenstein 1 1 School of Infomation and Communication Technology, Giffith Univesity, Austalia, 2 Laboatoie d Infomatique, Polytech Tous, Univesité Fançois-Rabelais, Tous, Fance 3 Compute Vision and Patten Recognition Unit, Indian Statistical Institute, Kolkata, India. sikanta.pal@giffithuni.edu.au Abstact Among all of the biometic authentication systems, handwitten signatues ae consideed as the most legally and socially accepted attibutes fo pesonal veification. The objective of this pape is to pesent an empiical contibution towads the undestanding of a theshold-based signatue veification technique involving off-line Bangla (Bengali) signatues. Expeiments on signatue veification involving non-english signatues ae an impotant consideation in the signatue veification aea. Only vey few eseach woks employing signatues of Indian scipt have been consideed in the field of non-english signatue veification. To fill this gap, a theshold-based scheme fo veification consideing off-line Bangla signatues is poposed. Some techniques such as undesampled bitmap, intesection/endpoint and diectional chain code ae employed fo featue extaction. The Neaest Neighbou method is consideed fo classification. Futhemoe, a Bangla signatue database, which consists of 2400 (100x24) genuine signatues and 3000 (100x30) fogeies has been ceated and is employed fo expeimentation. We obtained a 15.57% Aveage Eo Rate (AER) as the best veification esult using diectional chain code featues employed in this eseach wok. T I. INTRODUCTION ODAY, biometic technologies ae inceasingly being used to ensue identity veification o the authoisation of access to sensitive data. Among all of the biometic authentication systems, handwitten signatues have been accepted as an official means to veify pesonal identity fo legal puposes on such documents as cheques, cedit cads and wills [1]. The development of compute aided handwitten signatue veification systems has been ongoing fo decades [1]. In [1, 13], a thoough suvey of automatic handwitten signatue veification and wite identification techniques was povided. Justino et al. [2] poposed an off-line signatue veification system based on Hidden Makov Models (HMMs) to detect andom, casual, and skilled fogeies. Jing Wen et al. [14] pesented a featue extaction method based on the intensity of the coefficients of the Gabo tansfom. The image is subjected to a multichannel Gabo Tansfom, then the tansfomed Gabo image is divided equally into N1 * N2 non-ovelapping boxes. The angle featues of the position of the maxima intensity of the Gabo Tansfom Coefficients ae extacted. Ramachanda et al. [15] also poposed an offline signatue authentication system using coss-validated gaph matching. A bipatite gaph is constucted fom which a minimum cost-complete match is obtained. The Euclidean distance is calculated by the Hungaian method. A coss-validation pinciple is used to select efeence signatues. Anothe signatue veification method was pesented by Maha et al. [16]. Thee kinds of featues such as: Gid, Global, and Textue Featue Compaison ae used fo signatue veification. When employing gid-based featues, a signatue image is divided into ectangula egions, and the ink distibution in each egion is evaluated. In the global featue compaison, a numbe of featues extacted globally fom the whole signatue ae compaed. The textue-based featue compaison is based on the co-occuence matices of the signatue image. The Euclidian distance is used fo offline signatue veification. In some ecent wok, Pal et al. [20] pesented a signatue veification system employing Hindi Signatues. The objective of this pape was to pesent an investigation of the pefomance of a signatue veification system involving Hindi offline signatues. Hindi is a vey popula language in India, witten in Devnagai scipt. Encouaging esults wee obtained in this investigation. Recently, Pal et al. [8] pesented a signatue veification system based on a local language in India. In thei pape, the pefomance of an offline signatue veification system involving Bangla (Bengali) signatues, whose style is distinct fom Westen scipts, was investigated. The Gaussian Gid featue extaction technique was employed fo featue extaction and Suppot Vecto Machines (SVMs) wee consideed fo classification. The Bangla signatue database employed in those expeiments consisted of 3000 fogeies and 2400 genuine signatues. Howeve, the techniques pesented in the liteatue fo signatue veification of othe languages such as English, Japanese, Chinese and Pesian may not be suitable fo the numeous Indic scipts and languages, which ae officially in use in India and Bangladesh. To deal with the poblem of multi-scipt signatue veification in such an envionment, one possible solution may be the identification of signatues in the fist stage and then signatue veification in the next stage. A few methods ae poposed fo signatue identification of Bangla, Hindi, Chinese and English signatues [17] and a piece of wok is also pesented on Bangla signatue veification [18]. Since Bangla signatues ae mainly composed of a text potion, in this pape, an empiical study is pefomed fo Bangla signatue

3 veification using diffeent featues with a paticula focus on a theshold-based technique. Some Bangla signatue samples ae shown in Table I. The emainde of this pape is oganized as follows. The signatue veification concept is descibed in Section II. Section III descibes the impotance of Bangla signatue veification. The Bangla signatue database developed fo the cuent eseach is descibed in Section IV. Section V biefly descibes the featue extaction technique employed. Details of the classifies used ae pesented in Section VI. The expeimental settings ae pesented in Section VII. Results and discussion ae given in Section VIII. Finally, conclusions and futue wok ae discussed in Section IX. TABLE I SIGNATURE SAMPLES OF BANGLA SCRIPT has no access to a sample of the signatue. C. Skilled fogey The last type is the skilled fogey, pefomed by an expet peson. The foge has a good knowledge about the oiginal signatue and signs with sufficient pactice. Natually it is moe difficult to detect skilled fogeies than othe fogeies. Genuine Signatues/ Random Fogeies Signatue Pe-pocessing Featue Extaction Questi- oned Signatues Taining Testing Signatue Mod el II. SIGNATURE VERIFICATION CONCEPT In the aea of pesonal authentication, signatue veification is consideed a vey popula biometic method. Geneally, off-line/on-line signatue veification is fomulated as a two-class classification poblem whee the classes ae associated with a genuine signatue set and a foged signatue set. Usually two types of eos [21] ae consideed in a signatue veification system: The False Rejection o Type-I eo and the False Acceptance o Type- common types II eo. These eo types associate with two of eo ates: False Rejection Rate (FRR), which is the pecentage of genuine signatues misclassified as fogeies, and False Acceptance Rate (FAR) which is the pecentage of foged signatues misclassified as genuine. A geneal signatue veification appoach is shown in Fig. 1. The fogeies involved in handwittenn signatues have been categoized based on thei chaacteistic featues. Thee diffeent types of fogeies ae taken into account in the field of signatue veification. Accoding to Coetze et al. [9], the thee basic types of foged signatues ae: A. Random fogey The signe uses the name of the autho in his own style to ceate a fogey known as a simple o andom fogey. It is done by a peson who doesn t know the shape and stuctue of the oiginal signatue. This type of fogey is vey easy to detect. In many cases, the fogeies ae the foges own genuine signatue. B. Simple fogey The foge knows the autho s name and the scipt, but Type of input Signatue Fig. 1. Signatue veification appoach III. IMPORTANCE OF BANGLA SIGNATURE VERIFICATION A county having two o moe scipts and languages is known as a multi-scipt and multi-lingual county. Most counties have only a single language but vey few counties have moe than one scipt fo eading and witing puposes. Howeve in India, thee ae officially 23 (Indian constitution accepted) languages and 11 diffeent scipts. Geneally, scipts ae employed fo witing languages down in a eadable fom. In such a multi-scipt and multi-lingual county like India, Bangla scipt, as one of the official languages of India, is not only used fo witing/eading puposes but also fo signing puposes. Bangla signatue veification is quite diffeent with espect to the shape and stuctue of Bangla signatue images. Most of the chaactes of Bangla hand-witten wods ae touching. A distinctive hoizontal line unning along the top of the lettes that links them togethe is pesent. Bengali scipt is less blocky and pesents a moe sinuous shape. Bangla signatues always contain textual components. Fo Bangla signatues, thee is no such concept of an initial signatue (fist lette of the fist name followed by a family name) which is seen fequently in English signatues. To get the desied accuacy when dealing with patten ecognition poblems, featue set selection takes on a vital ole. It is sometimes difficult to choose the appopiate

4 featues because of the diffeent chaacteistics of vaious scipts. The featues used hee (unde-sampleintesection/endpoint and diectional chain code) ae suitable bitmap, fo Bangla scipt with the chaacteistics peviously descibed. IV. SIGNATURE DATABASE Some notable popeties of Bangla scipt and the Bangla signatue dataset used fo these expeiments ae descibed hee. A. Popeties of Bangla Scipt Bangla is the second most popula language in India. Bangla scipt is widely used fo official easons thoughout West Bengal and Assam (two states of India).. Othe counties using Bangla include Bangladesh. It is estimated that thee ae about 300 million people using Banglaa scipt. Bangla is an oiental scipt descended fom Bahmi scipt [19]. In Bangla scipt, the witing diection is fom left to ight and does not have the concept of distinct lette cases. It is identifiable by a distinct hoizontal line called the head line that uns along the top of full lettes and it connect all the lettes togethe in a wod. Bangla scipt has about fifty basic chaactes. These chaactes ae pesented in Fig. 2. pactice thei fogeies as long as they wished with static images of genuine specimens. A total numbe of 3000 (30 signatues 100 individuals) foged signatues wee collected fom the wites. In summay, the Bangla signatue database consists of 5400 signatues of which 2400 (24 signatues 100 individuals) signatues ae genuine signatues and 3000 (30 signatues 100 individuals) signatues ae foged signatues. Some genuine signatue samples with thei coesponding fogeies ae shown in Table II. Fom Table II, it can be noted that in the Bangla signatue database povided in this eseach wok, the genuine and foged signatues ae vey simila. This popety of the Bangla signatue database makes the veification task vey challenging. TABLE II EXAMPLES OF GENUINE AND FORGED SIGNATURES OF BANGLA SCRIPT Genuine Signatues Foged Signatues Fig. 2. Basic chaactes of Bangla scipt B. Data Collection and Pepaation Although automatic signatue veification has been an active eseach aea fo seveal decades, and a few English language signatue databases ae available fo eseach [11], thee has been no publicly available signatue database fo Bangla, the second most popula Indian scipt. Moeove, eseach on Indian-based scipt signatues has gadually inceased in ecent times. Theefoe, a Bangla signatue copus was ceated fo this wok. The signatues wee collected fom diffeent pats of the West Bengal State of India. The majoity of the signatues wee contibuted by students. The emaining signatues wee collected fom senios of the locality. This Bangla signatue database consists of 100 sets (classes). In ode to collect the genuine signatues coesponding to each individual, a collection fom was designed. The fom contained 24 boxes whee the signatues could be witten. Fom each individual, 24 genuine signatues wee collected. A total numbe of 2400 genuine signatues wee collected fom 100 individuals. Fo each contibuto, all genuine specimens wee collected in a single day's witing session. In addition, skilled foged signatues wee collected fo this wok. In ode to poduce the fogeies, the imitatos wee allowed to V. FEATURE EXTRACTION Featue extaction is a vital step in any patten ecognition system. Diffeent methods have been poposed in the offto pefom the extaction line signatue veification liteatue of featues fom signatue images [12]. In this eseach wok, thee diffeent featue extaction methods, used in the liteatue fo document ecognition, ae studied. These featues ae chosen, since, they have shown thei stength on text pattens [3-6] and Bangla signatues ae mainly textual pattens. In the following, bief desciptions of thee diffeent featue sets used in this pape ae discussed. It is woth mentioning that vey small components (mainly dots o noise) ae eliminated fom evey signatue image using a theshold based on the aveage size of connected components in the image to have a pope bounding box of the signatue (Fig. 3(b)). A. Undesampled Bitmap Featues Featue extaction based on undesampled bitmaps is a simple technique used in the liteatue by many eseaches [3, 4] fo ecognition puposes. The undesampled bitmaps ae fomed by dividing each input image into a numbe of non-ovelapping blocks of simila size. Then, the numbe of black pixels is counted in each block. This geneates an input matix with each element being an intege in the ange 0 to the size of the non-ovelapping block. Dividing these values by the size of the block, the values ae nomalized between 0 and 1. The undesampling pocess educes dimensionality of

5 the featues compaed to the whole image size and povides invaiance to small distotions and slant [4]. In ou study, afte binaizing the input image, a minimum- (Fig. 3(b)). bounding box of the input image is obtained Then, fo a bette esult and independence of featues to size and position (invaiant to scale and tanslation), the minimum-bounding box of the image is conveted into a nomal size of pixels (Fig. 3(c) )). This value is detemined expeimentally. To compute the undesampled bitmap featues, the nomalized image ( ) is divided into 80 non-ovelapping blocks of size pixels (Fig. 3(c)). Then, the numbe of black pixels is counted in each block. This geneates an input matix of 8 10 with each element being an intege in the ange 0 to 2000 (25 80). Dividing these values by the size of the block (2000), the values ae nomalized between 0 and 1. Since, the nomalized image is divided into 80 blocks, 80 featues ae obtained fo each input signatue. (a) (b) (c) Fig. 3. (a) A Bangla signatue sample, (b) Bounding box of the Bangla signatue(afte emoving the small components), (c) Nomalized image ( ) and its non-ovelapping window-map of size on a nomalized image. B. Intesection/Junction/End Points Featues An intesection point is defined as a pixel point which has moe than two neighbouing pixels with 8-connectivity, while an endpoint has exactly one neighbouing pixel. Intesection featues ae extacted fom the thinned signatue image, which is fist nomalized into pixels. Fo the thinning pocess, the algoithm pesented in [5] is utilized. The thinned signatue image is then divided into 20 blocks each of size pixels. Fo each block, the numbe of endpoints and intesection/junctions ae found and counted sepaately. Thus, 40 (20 2) featues fom 20 constituent blocks of the signatue image ae computed, out of which the fist 20 featues epesent the numbe of end points and the emaining 20 featues epesent the numbe of intesection/junction points. Intesection points and endpoints of a Bangla signatue image ae shown in Fig. 4. Fig. 4. Intesection points and endpoints ae shown by ed maks on a thinned signatue image (fo bette visibility of the figue see the soft copy of the pape). C. Diectional Chain Code Featues The chain code diection infomation of the contou points of an input image have been used as featues fo diffeent puposes, including numeal and chaacte ecognition [6]. Contou infomation can povide a good epesentation of signatue shape. To compute diectional chain code featues, simila to the undesampled and intesection/end point featue extaction techniques, at fist, the minimum bounding box of the input image is obtained and the minimum bounding box is nomalized into pixels. Using the nomalized binay image, the contou points of the signatue image ae found based on the 8-connectivity fomula (Fig. 6). The image contou is scanned hoizontally by keeping an ovelapping window-map of size (Fig. 6) on the image fom the top-left most point to the down-ight most point (80 ovelapping blocks). Fo each ovelapping block, the chain code fequencies fo all 8 diections (Fig. 7) ae computed. Instead of expessing the featues in tems of 8 diections, we simplify the featues into 4 diections (Fig. 8): (i) Hoizontal diection code (diections 0 and 4), (ii) Vetical diection code (diections 2 and 6), (iii) Diagonal diection code (diections 1 and 5) and (iv) Off-diagonal diection code (diections 3 and 7). Thus, in each block, fou featues epesenting the fequencies of these fou diections ae obtained. As a esult, fo each image we obtain 320 (80 4) featues fom 80 blocks. The eason fo choosing an ovelapping window-map (one pixel fom each side) instead of a non-ovelapping window-map is to peseve the infomation between a window-map and its neighbouing blocks. Moeove, based on an expeimental study we have extacted featue sets fom non-ovelapping as well as ovelapping blocks, and ovelapping blocks povided bette pefomance. Fig. 5. Inne and oute contou of the Bangla signatue shown in Fig. 3(c). Fig. 6. An ovelapping window-map of size contou image is shown in ed on the

6 Fig. 7. Point P and its 8-diection codes. Fig. 8. Fou diections obtained fom 8 diections. VI. CLASSIFIER DETAILS Amongst the vaious methods of supevised statistical patten ecognition, the Neaest Neighbou (NN) ule achieves consistently high pefomance, without a pioi assumptions about the distibutions fom which the taining examples ae dawn. A new sample is classified by calculating the distance to the neaest taining case; the sign of that point then detemines the classification of the sample. The distances can be calculated using one of the distance measues such as Euclidian, Mahalanobis and City-block. In this pape, the Euclidian distance measue is used fo expeimentation. TABLE III VERIFICATION PERFORMANCE OF CONTOUR FEATURES ON THE BANGLA DATASET BASED ON DIFFERENT THRESHOLD VALUES. Theshold FAR FRR AER 0.8*Aveage distance * Aveage distance *Aveage distance *Aveage distance *Aveage distance VII. EXPERIMENTAL SETTINGS To evaluate the poposed systems, thee well-known eo measues, specifically AER (Aveage Eo Rate), FAR (False Acceptance Rate) and FRR (False Rejection Rate) used in many papes in the liteatue [13] ae utilized. AER is the aveage of FAR and FRR. The AER is geneally adopted as a unique measue fo chaacteizing the pefomance level of biometic systems, and it indicates the secuity level povided by the biometic system. Fo the expeiments, ou own Bangla database descibed in Section IV was used. We tained the system with a set of 12 genuine signatues of each individual (class). Then a distance map using the Euclidian distances between all 12 genuine signatues of each class (121 distances) wee calculated. The aveage distance of all 121 distances fo each class was found. The aveage distance of each class was consideed as the acceptance/ejection theshold of that class. To get the minimum AER, we empiically found that when FAR and FRR wee equal, then the AER was at a minimum. So, we adjusted the acceptance/ejection theshold based on the aveage distance of each efeence class. The minimum AER was desied in the expeimental method and it is achieved at the level whee the othe two eo ates (FAR and FRR) wee the same. Theefoe, the equal values of FAR and FRR shown in Table IV and Table V ae consideed to find the least value of AER. The vaiations of FAR and FRR using contou featues fo vaious thesholds on the Bangla database is shown in Table III. The test set consists of the emaining samples of genuine signatues and all the foged signatues. If the minimum distance between a test sample and the tained samples of a efeence class is less than the pedefined acceptance/ejection theshold of the efeence class, then the test signatue is consideed as authentic, othewise it is consideed as a fogey. VIII. RESULTS AND DISCUSSION As mentioned ealie, the vaiations of FAR and FRR using contou featues fo vaious thesholds on the Bangla database ae shown in Table III. Hee values of FAR, FRR and AER ae shown fo five diffeent theshold values. FAR, FRR and AER values obtained fom thee diffeent featues on the Bangla signatue dataset ae povided in Table IV. The gaphical epesentation of diffeent values of AER obtained fom diffeent featues on the Bangla database is also shown in Fig. 9. We measued the veification pefomance in tems of the commonly used aveage eo ate (AER). Fom the esults shown in Table IV, it can be noted that the best esults wee obtained when employing the contou featues of signatue images. Howeve, undesampled bitmap featues also povided good esults, with only a vey small numbe of featues as compaed to contou featues. TABLE IV VERIFICATION PERFORMANCE ON THE BANGLA DATASET FOR 3 DIFFERENT FEATURE SETS Featue set 80 dimensional featues undesampled bitmaps 40 dimensional featues intesection and end points 320 dimensional featues contou infomation FAR FRR AER

7 Fo a compaative study about the pefomance of these thee featues, we also computed FAR, FRR and AER values of the poposed model using 100 sets of the GPDS [7] dataset, and the esults ae povided in Table V. Fom Table IV and Table V it can be seen that contoubased featues outpefom the othe two featues fo both Bangla and the GPDS signatue datasets. Howeve, to get a compaative idea between Bangla and English signatue databases, the AER values in the GPDS dataset (Table V) ae highe than the AER values in the Bangla dataset. This is because, most of the signatue samples in Bangla dataset ae textual signatues wheeas in the GPDS dataset, most of the samples ae composed of lines and cuved shape pattens, and the featues studied in this eseach wok ae mainly dedicated to textual pattens. It may be noted that the FAR, FRR and AER ae equal in ou expeimental esults because of ou theshold selection as discussed in Section IV. The gaphical epesentation of diffeent values of AER obtained fom diffeent featues using the GPDS database is also shown in Fig. 10. E o R a t e Fig. 9. Repesentation of aveage eo ate obtained using thee diffeent featues on the Bangla dataset. TABLE V VERIFICATION PERFORMANCE ON THE GPDS DATASET FOR 3 DIFFERENT FEATURE SETS Featue set Intesection and End points Undesampled Bitmaps Diectional Chain Code Featues Used fo Bangla Datasets 80 dimensional featues undesampled bitmaps 40 dimensional featues intesection and end points 320 dimensional featues contou infomation FAR FRR AER E o R a t e Diectional Chain Code Undesampled Bitmaps Intesection and End points Featues Used fo the English (GPDS) Dataset Fig. 10. Repesentation of aveage eo ate obtained using thee diffeent featues using the GPDS dataset. IX. CONCLUSIONS AND FUTURE WORK This pape demonstates an investigation of the excellent pefomance of theshold-based signatue veification technique involving Bangla off-line signatues. This novel appoach fo signatue veification employed a lage Bangla off-line signatue dataset, poviding a substantial contibution to the field of signatue veification. Thee diffeent featue sets (unde-sampled bitmap featue, intesection/endpoint featue and diectional chain code featues) ae used hee to pefom a compaative study on the Bangla signatue dataset. To the best of ou knowledge, the abovementioned methods have neve been used fo the task of Bangla signatue veification. Moeove, this scheme of Bangla off-line signatue veification is also a new contibution to the field of signatue veification. In the nea futue, we plan to extend ou wok consideing moe samples of Bangla off-line signatues. We also plan to make this extended dataset available publicly to eseaches. REFERENCES [1] R. Plamondon and G. Loette, Automatic Signatue Veification and Wite Identification The State of the At, Patten Recognition, vol. 4, no. 2, pp , [2] E. Justino, E. Botolozzi, R Sabouin, Off-line Signatue Veification Using HMM fo Random, Simple and Skilled Fogeies, Poceedings of 7th in Intenational Confeence on Document Analysis and Recognition, pp , [3] M.D. Gais, NIST Fom-Based Handpint Recognition System, NISTIR 5469, [4] E. Alpaydin, C. Kaynak, Cascading Classifies, Kybenetika, 34(4), pp , [5] C. Rafael, R. Gonzalez, E. Woods, Digital Image Pocessing, Second Edition, Pentice Hall India, [6] F. Kimua, T. Wakabayashi, S.Tsuuoka and Y. Miyake, Impovement of handwitten Japanese chaacte ecognition using weighted diection code histogam, Patten ecognition, vol. 30, no. 8, pp , [7] M. A. Fee, J. B. Alonso, and C. M. Tavieso, Offline geometic paametes fo automatic signatue veification using fixed-point

8 aithmetic, IEEE tans. on Patten Analysis and Machine Intelligence, vol. 27, no. 6, pp , [8] S. Pal, V. Nguyen, U. Pal and M. Blumenstein, Off-line bangla signatue veification, In: Intenational Wokshop on Document Analysis Systems, pp , [9] J. Coetze, B. Hebst, and J. D. Peez, Off-line signatue veification using the discete adon tansfom and a hidden makov model, EURASIP Jounal on Applied Signal Pocessing, vol. 4, pp , [10] B. B. Chaudhui and U. Pal, An OCR system to ead two Indian language scipts: Bangla and Devnagai (Hindi), Poceedings of 4 th Intenational Confeence on Document Analysis and Recognition, pp , [11] B. Fang, C.H. Leung, Y.Y. Tang, K.W. Tse, P.C.K. Kwok and Y.K. Wong, "Off-line signatue veification by the tacking of featue and stoke positions", Patten Recognition, vol. 36, pp , [12] S. Chen, and S. Sihai, Use of Exteio Contou and Shape Featues in Off-line Signatue Veification, Poceedings of 8 th Intenational Confeence on Document Analysis and Recognition, pp , [13] D. Impedovo, G. Pilo, Automatic signatue veification: The state of the at, IEEE Tansactions on Systems, Man, and Cybenetics pat-c, vol. 38, no. 5, pp , [14] J. Wen, B. Fang, Y. Y. Tang, T. P. Zhang, and H. X. Chen, Offline Signatue Veification Based on the Gabo Tansfom, Poceedings of the Intenational Confeence on Wavelet Analysis and Patten Recognition, pp , [15] A. C. Ramachanda, K. Pavitha, K. Yashasvini, K.B.Raja, K.R. Venugopal and L. M. Patnaik, Offline Signatue Authentication using Coss validated Gaph Matching, Poceedings of the second Bangaloe Annual Compute Confeence, ACM, Aticle no. 7, [16] J. A. Maha, M. H. Maha and M. K. Khan, Compaative Study of Featue Extaction fo Off-Line Signatue Veification, Second Intenational Confeence on Emeging Technologies, pp , [17] S. Pal, U. Pal and M. Blumenstein, Off-line English and Chinese Signatue Identification Using Foegound and Backgound Featues, In Poc. of IJCNN Special Session on Machine Leaning fo Compute Vision at 2012 IEEE Wold Congess on Computational Intelligence, WCCI 2012, pp. 1-7, [18] S. Pal, A. Alaei, U. Pal and M. Blumenstein, Off-line Signatue Veification based on Foegound and Backgound infomation, Poceedings of Intenational Confeence on Digital Image Computing Techniques and Applications, pp , [19] B. B. Chaudhui and U. Pal, An OCR system to ead two Indian language scipts: Bangla and Devnagai (Hindi), Poceedings of 4 th Intenational Confeence on Document Analysis and Recognition, pp , [20] S. Pal, M. Blumenstein and U. Pal, Hindi off-line signatue veification, In: Intenational Confeence on Fonties in Handwiting Recognition, pp , 2012.