Hindi Off-line Signature Verification

Srikanta Pal

School of Information and Communication Technology,

Griffith University, Gold Coast Australia, Email:

srikanta.pal@griffithuni.edu.au

Michael Blumenstein School of Information and

Communication Technology, Griffith University, Gold Coast,

Australia, Email: m.blumenstein@griffith.edu.au

Umapada Pal Computer Vision and Pattern

Recognition Unit, Indian Statistical Institute, Kolkata-700108, India.

Email: umapada@isical.ac.in

Abstract—Handwritten Signatures are one of the widely used biometrics for document authentication as well as human authorization. The purpose of this paper is to present an off-line signature verification system involving Hindi signatures. Signature verification is a process by which the questioned signature is examined in detail in order to determine whether it belongs to the claimed person or not. Despite of substantial research in the field of signature verification involving Western signatures, very little attention has been dedicated to non-Western signatures such as Chinese, Japanese, Arabic, Persian etc. In this paper, the performance of an off-line signature verification system involving Hindi signatures, whose style is distinct from Western scripts, has been investigated. The gradient and Zernike moment features were employed and Support Vector Machines (SVMs) were considered for verification. To the best of the authors’ knowledge, Hindi signatures have never been used for the task of signature verification and this is the first report of using Hindi signatures in this area. The Hindi signature database employed for experimentation consisted of 840 (35x24) genuine signatures and 1050 (35x30) forgeries. An encouraging accuracy of 7.42% FRR and 4.28% FAR were obtained following experimentation when the gradient features were employed.

Keywords- Signature verification, Indian script, Hindi signatures, Document security.

I. INTRODUCTION

Handwritten signatures are one of the most widely accepted personal attributes for identity verification. Signature verification has been a topic of renewed intensive research over the past several years [1, 2] due to the important role it plays in numerous areas, including in financial applications.

Automatic signature verification systems can be classified into two categories: on-line and off-line [3]. In an on-line technique, signatures are signed on a digitizer and dynamic information such as speed and pressure is captured in addition to a static image of the signature [4, 5]. In an off-line technique, signatures are signed on a piece of paper

and then scanned to digitally store the signature image [6]. Hence, off-line signature verification deals with the verification of signatures, which appear in a static format [7]. Verification decisions are usually based on local or global features extracted from the signature being processed. Excellent verification results can be achieved by comparing the robust features of the test signature with that of the user’s signature using an appropriate classifier [20].

Signatures are considered as a complete image with a special distribution of pixels, and a particular writing style. They are not considered as a collection of letters and words [8]. A person’s signature may change radically during their lifetime. Great inconsistency can even be observed in signatures according to country, habits, psychological or mental state, physical and practical conditions [9].

There has been substantial work in the area involving off-line verification of Western signatures. Armand et al. [10] presented an effective method to perform off-line signature verification and identification. Unique structural features were extracted from the signature's contour. Using a publicly available database of 2106 signatures containing 936 genuine and 1170 forgeries, the verification rate of 91.12% was obtained. Ramachandra et al. [11] proposed an off-line signature verification system based on a Cross-Validation principle and graph matching. Schafer and Viriri [12] presented an off-line signature verification system based on the combination of feature sets. Some extracted features were: Aspect ratio, centroid feature, four surface features, six surface features, number of edge points, transition features etc. The verification of signatures was accomplished using the Euclidean distance classifier.

Signatures may be written in different languages and there is a need to undertake a systematic study in this area. Many published works are available for Western signatures and only a few studies have been undertaken for signatures written in Chinese, Japanese, Persian, Arabic etc. [19]. To the best of the authors’ knowledge there is no published work on Hindi signature verification and this paper deals with Hindi signature verification. The present work of Hindi signature verification would be considered as a novel

2012 International Conference on Frontiers in Handwriting Recognition

978-0-7695-4774-9/12 $26.00 © 2012 IEEE

DOI 10.1109/ICFHR.2012.212

371

2012 International Conference on Frontiers in Handwriting Recognition

978-0-7695-4774-9/12 $26.00 © 2012 IEEE

DOI 10.1109/ICFHR.2012.212

373

contribution to the field of signature versignature samples of Hindi script are shown

The remainder of this paper is organizeddifferent types of forgeries are described inHindi signature database developed for the is described in Section III. Some notablDevnagari script are introduced in Sectionbriefly describes the feature extraction technin the work. Details of the classifiers used Section VI. The experimental settings arSection VII, and results and discussion are gVIII. Error analyses are described in Secticonclusions and future work are discussed in

Figure 1. Hindi signature sam

Figure 1. Hindi signature samples

II. TYPES OF FORGERI

In general, off-line/on-line signature ver

considered as a two-class classification profirst class represents the genuine signatusecond class represents the forged signatutwo types of errors are considered in a signasystem: The False Rejection or Type-I erroAcceptance or Type-II error. These eassociated with two common types of errorRejection Rate (FRR) which is the percensignatures misclassified as forgeries, and FaRate (FAR) which is the percentage of fomisclassified as genuine. According to Coethree basic types of forged signatures, whichinto account, are:

1. Random forgery. The forger has nogenuine signature (not even the authorreproduces a random one. In many cases, tthe forger’s own genuine signature.

2. Simple forgery. The forger knows theand the script, but has no access to a signature.

3. Skilled forgery. The forger has accesssamples of the genuine signature and is able

rification. Some in Figure 1.

d as follows. The n Section II. The current research le properties of n IV. Section V niques employed are presented in re presented in given in Section ion IX. Finally, n Section X.

mples

IES

rification can be oblem. Here the

ure set, and the ure set. Usually ature verification or and the False rror types are

r rates: the False ntage of genuine alse Acceptance orged signatures etzer et al. [13], h are often taken

o access to the r’s name) and the forgeries are

e author’s name sample of the

s to one or more to reproduce it.

III. HINDI SIGNAT

Although automatic signature veactive research area for several depublicly available signature databpopular official Indian script. Thedatabase was created for the purporesearch in automatic signature veconstrained by the unavailability of

TABLE 1. GENUINE AND FORG

A. Data collection and database p

The signatures of Hindi script wsignature verification approach. Assignature corpus available for Hindto create a database of Hindi sigwere collected from West Bengalthe signatures were contributed bsignature database consists of 35 number ranges from H-S-Set-001 Signature-Set). In order to colleccorresponding to each individualdesigned. The form contained signatures could be written. Frogenuine signatures were collectedgenuine signatures from 35 individeach contributor, all genuine specisingle day's writing session. In addsignatures were collected for this prproduce the forgeries, the imitatorstheir forgeries for as long as they wof genuine specimens. A total n

Hindi SignatGenuine Signatures

TURE DATABASE

erification has been an ecades, there has been no base for Hindi, the most erefore, a Hindi signature ose of this work. So, the erification has long been f a standard database.

GED SIGNATURES

preparation

were considered for this s there has been no public di script, it was necessary gnatures. The signatures , India. The majority of by students. This Hindi sets whereby the writer to H-S-Set-035 (Hindi -

ct the genuine signatures , a collection form was

24 boxes where the om each individual, 24

d. A total number of 840 duals were collected. For mens were collected in a

dition, only skilled forged roposed work. In order to

s were allowed to practice wished with static images number of 1050 forged

ures Forged Signatures

372374

signatures were collected from the writerssignature samples with their correspondindisplayed in Table 1.

B. Pre-processing

The signatures to be processed by the systein a digital image format. Each signature won a rectangular space of fixed size on apaper. It was necessary to scan all signpages. At the very beginning, the images w256 level grey scale at 300 dpi and stored (Tagged Image File Format) for the purprocessing. In the pre-processing step, a hthreshold technique was applied for binaristep, the digitized grey-level image is convetone image. Then the signature images werethe signature-collection document forms. collection form containing 24 genuine siglevel are shown in Figure 4. The extracted bimages were stored in TIFF format. A tsignature and its corresponding binary imagFigure 2 and Figure 3, respectively.

Figure 2. Scanned signature image

Figure 3. Binary signature image

IV. PROPERTIES OF DEVNAGA

Devnagari is an oriental script descendedscript [14]. It is the most popular official scrlanguage of India. In Hindi script, the writfrom left to right and there is no concept case. Hindi script has about fifty basic chcharacters are presented in Figure 5. Vowels in this script generally take a mmost words and are called modifiers Modifiers generally do not disturb the characters in the middle zone of a line. disturbed in the middle zone, we call the recompound character. Vowel modifiers of Dare shown in Figure 6.

. Some genuine ng forgeries are

em needed to be was handwritten a white sheet of ature document

were captured in in TIFF format

rpose of future histogram-based ization. In this erted into a two-e extracted from

The signature gnatures in grey binary signature typical scanned ge are shown in

e

ARI SCRIPTS

d from Brahmi ript and national ting direction is of upper/lower

haracters. These

modified shape in or allographs.

shape of basic If the shape is esultant shape a

Devnagari scripts

Figure 4. Signature-collection form

Figure 5. Basic characters of

A text line in such scripts canzones (upper, middle and lower). the portion above the headline, theportion between the headline andzone is the portion below the baseseparating the middle and lower zo

Figure 6. Vowel modifiers

V. FEATURE EXTRACTION

Feature extraction is a crucialrecognition system. The Zernike feature extraction technique are des

A. Zernike moments feature

Zernike polynomials are an orthogovalued polynomials:

������ � ������ �� �where

with genuine signatures

f Devnagari script

n be partitioned into three The upper zone denotes

e middle zone denotes the d baseline and the lower eline. The imaginary line

ones is called the baseline.

s of Devnagari

N

l step in any pattern feature and the gradient scribed below.

onal set of complex-

�������� ��������

373375

�� � �� � ������ ���� � � �!�! � �

and � � !�! is even and Radial polynomials "���# are defined as:

������ � $ %�!�!&��� � �����&����!�!

�

&'(

where

%�!�!& ���&�) � �*

)* +� � !�!, � )- * +� � !�!

, � )- *

The complex Zernike moments of order n and repetition m are given by:

.�� � � �/ $$0��� ����1 ��� �

23

Where �� � �� � � and the symbol 1 denotes the complex conjugate operator [15]. The Zernike moments can be computed by the scale invariant Central moments as follows:

.�� � � �/ $ $$���4

!�!

4'(

5

6'(

�

7�!�!��7'898�

� :;�< ;= :>?=%�!�!&@7��6�4��6A4

where

> � � !�!, � �)����?�<������ ������

B. Computation of 576-dimensional gradient features

The grey-scale local-orientation histogram of the handwriting component is used for 576-dimensional feature extraction. To obtain a 576-dimensional gradient-based feature vector, the following steps were executed. Step 1: 2 x 2 mean filtering is applied 5 times on the input image. Step 2: The grey-scale image obtained in Step 1 is normalized so that the mean grey scale becomes zero with a maximum value of 1.

Step 3: The normalized image is then segmented into 17x7 blocks. Compromising trade-off between accuracy and complexity, this block size is decided experimentally. To get the bounding box of the grey-scale image, it is converted into a two-tone image using Otsu’s thresholding algorithm [16]. This will exclude unnecessary background information from the image.

Step 4: A Roberts filter is then applied on the image to obtain the gradient image. The arc tangent of the gradient (direction of gradient) is quantized into 32 directions and the strength of the gradient is accumulated with each of the quantized directions. The strength of the Gradient ��B�C� D is defined as follows:

B�C� D E�FG� � �FH� (1) and the direction of gradient �I��� J is: I��� J ����� FK

FL (2) where FG M�� � �� J � � � M��� J (3) and FH M�� � �� J � M��� J � � (4) and M��� J is the grey level of point

��� J. Step 5: Histograms of the values of 32 quantized directions are computed for each of the 17 x 7 blocks.

Step 6: The directional histogram of the 17 x 7 blocks is down-sampled into 9 x 4 blocks and 16 directions using Gaussian filters. Finally, a 9 x 4 x 16 = 576-dimensional feature vector is obtained.

VI. CLASSIFIER DETAILS

In our experiments, we have used Support Vector Machines (SVMs) as classifiers. SVMs have been originally defined for two-class problems and they look for the optimal hyper plane, which maximizes the distance and the margin between the nearest examples of both classes, namely support vectors (SVs). Given a training database of M data: {xm| m=1,..., M}, the linear SVM classifier is then defined as:

where {xj} are the set of support vectors and the parameters αj and b have been determined by solving a quadratic problem [17]. The linear SVM can be extended to various non-linear variants; details can be found in [17, 18]. In our experiments, the RBF kernel SVM outperformed other non-linear SVM kernels, hence we are reporting our verification results based on the RBF kernel only. Different parameters of the kernel are chosen experimentally.

bxxxf jj

j +⋅=�α)(

374376

VII. EXPERIMENTAL SETT In the field of signature verification, t

publicly available signature databases. Tavailable databases also varies, as therestandard collection protocol. Besides, it iscreate a large corpus with different typeespecially skilled forgeries.

For the experiments in the proposed resdatabase described in Section 4 was usedwas split into two parts to perform the train

phases. For each signature set, an SVM w14 randomly chosen genuine signatures.samples for training were the 20 skillesignatures. For testing, the remaining gen(10 signatures) and the remaining skilledsignatures) of the signatures being coemployed. The number of samples for trainfor the purpose of experimentation with eacare shown in Table 2

TABLE 2. NUMBER OF SIGNATURE SAMPLES USEDEXPERIMENTATION

VIII. RESULTS AND DISCUS

For training and testing of the systemsignatures were employed. Among these s(14x35+20x35) samples were considered phase and 700 (10x35+10x35) samples werthe testing phase. Using the gradient featumoment feature, an FRR (False Rejection (False Acceptance Rate) were obtained. At point, the Support Vector Machines produ7.42% and an FAR of 4.28% using the gwhereas an FRR of 8.57% and an FAR obtained using the Zernike moment Featurbetter results compared to our previous Bverification work [21].

The FRR and the FAR rates changed rvalue of gamma used in the SVM classifiersobtained from these experiments are showTable 6. These encouraging results drobustness of the gradient feature and Zfeature extraction techniques for signature ve

From the experimental results obtainobserved that the performance of the gravery encouraging, compared to the resultsthe Zernike moment feature. The results obexperiment using our Hindi dataset are compthan the state-of-the-art methodologies propo

Genuine Signature Skill

Training 14

Testing 10

TINGS

there is lack of The quality of e has been no s very costly to es of forgeries,

search, our own d. The database

ning and testing

was trained with . The negative ed forgeries of nuine signatures d forgeries (10 onsidered were ning and testing ch signature set,

D PER SET FOR

SSION

m, 1890 Hindi signatures, 1190

in the training re considered in ure and Zernike Rate) and FAR this operational

uced an FRR of gradient feature of 5.14% were

re. We obtained Bangla signature

rapidly with the s. The error rates

wn in Table 3 to demonstrate the Zernike moment erification. ned, it is also adient feature is obtained using tained from this paratively better osed [9].

TABLE 3. VALUE OF FRR USING TH

Gamma

0.000005 0.000003 0.000001

TABLE 4. VALUE OF FAR USING TH

Gamma

0.000005 0.000003 0.000001

TABLE 5. VALUE OF FAR USING

Gamma

0.000003 0.000002 0.000001

TABLE 6. VALUE OF FRR USING

Gamma

0.000003 0.000002 0.000001

IX. ERROR

Confusing samples obtained u

based on the gradient feature anfeature are shown in Figure 7, Figu10, respectively. Two categoriewere obtained by the classifier. Ththe genuine signature samples tresamples. The second one illustrasamples treated as genuine signatur

Figure 7. Genuine signature sample treate

on the gradient fea

led Forgeries

20

10

HE GRADIENT FEATURE

FRR (%)

8.57 8.00 7.42

HE GRADIENT FEATURE

FAR (%)

5.14 4.87 4.28

G ZERNIKE MOMENTS

FRR (%)

28.57 17.14 8.57

ZERNIKE MOMENTS

FAR (%)

13.71 8.28 5.14

ANALYSIS

using the SVM classifier nd the Zernike moment ure 8 and Figure 9, Figure es of confusing samples e first category illustrates

eated as forged signature ates the forged signature re samples.

ed as a forged signature (based ature).

375377

Figure 8. Forged signature sample treated as a genuinon the gradient feature).

Figure 9. Genuine signature treated as a forged signaZernike moment feature).

Figure 10. Forged signature treated as a genuine signaZernike moment feature).

X. CONCLUSIONS AND FUTU

This paper presents an investigation of tof a signature verification system involvingsignatures. The gradient feature, Zernike mand SVM classifiers were employed, anresults were obtained. To the best of ouHindi signature database has not previouslythe task of signature verification and this isof using Hindi signatures in this area. The pverification scheme is the first investigasignatures in the field of off-line signaturethe near future, we plan to extend our wmore samples of Hindi signatures with thedifferent feature extraction techniques and also plan to prepare a Hindi signature databpublicly available.

REFERENCES [1] S. Chen, and S. Srihari, “Use of Exter

Shape Features in Off-line SignaturProceedings of 8th ICDAR, pp. 1280-1284

[2] M.A. Ferrer, J.B. Alonso and C. M. TrGeometric Parameters for Automatic SignUsing Fixed-Point Arithmetic" IEEE PAMpp. 993-997, 2005.

[3] S. Madabusi, V. Srinivas, S. BhasBalasubramanian, “On-line and offverification using relative slope algorithWorkshop on Measurement Systems Security, pp. 11-15, 2005.

[4] R. Plamondon and G. Lorette, “Autoverification and writer identification - thePattern Recognition, 22, pp.107–131, 1989

ne signature (based

ature (based on the

ature (based on the

URE WORK

the performance g Hindi off-line moment feature nd encouraging

ur knowledge, a y been used for s the first report

proposed off-line ation for Hindi e verification. In

work considering e exploration of classifiers. We

base and make it

rior Contour and re Verification”, 4, 2005. ravieso, "Off-line natureVerification MI, Vol.27, no.6,

skaran and M. ff-line signature hm”, International

for Homeland

omatic signature e state of the art”, 9.

[5] A. Kholmatov, and B. Yanikoglusing improved online signaturepp. 2400-2408, 2005.

[6] M. Kalera, S. Srihari, and Averification and identification IJPRAI, pp.1339-1360, 2004.

[7] I. Pottier and G. Burel, ‘‘Identiof Handwritten Signatures with In Proc. IEEE Conf. On Neural1994.

[8] B. Fang, C.H. Leung, Y.Y. Tanand Y.K. Wong, "Off-line sigtracking of feature and stRecognition, pp. 91-101, 2003.

[9] K. Franke, “Analysis of AForgeries” In Proc. IWCF, pp 15

[10] S.Armand, M. Blumenstein an“Off-line Signature VerificatioDirection Feature,” Proceedings2006.

[11] A. C. Ramachandra, K. PavithraK. R. Venugopal and L. M. PVerification based on CroMatching,” IEEE Internationaland Electronics (INDICON-2008

[12] B.Schafer and S.Viriri, “An Off-System”, ICSIPA, pp.95-100, 20

[13] J. Coetzer, B. Herbst and J. D.verification using the discrete ramarkov model”, EURASIP JoProcessing, vol 4, pp.559–571, 2

[14] B. B. Chaudhuri and U. Pal, “AIndian language scripts: BanglProceedings of 4th ICDAR, pp. 1

[15] C. H. Teh and R. T. Chin, “OMethods Of Moments,” IEEE Intell., Vol. 10, No. 4, pp. 496-5

[16] N. Otsu. A threshold selectionhistogram. SMC, IEEE Trans. on

[17] V.Vapnik, “The Nature of StaSpringer Verlang, 1995.

[18] C. Burges, “A Tutorial on suppattern recognition”, Data discovery, pp.1-43, 1998.

[19] H. Lv, W. Wang, C. Wang and signature verification based on Pattern Recognition Letters, pp.

[20] R. Plamondon and S.N. SrihHandwriting Recognition: A cotran. On Pattern Analysis and Mno.1, pp. 63-84, 2000.

[21] S. Pal, V. Nguyen, M. BlumenBangla Signature Verification”Queensland, Australia, 27-29 Ma

lu, “Identity Authentication verification method”, PRL,

A. Xu. “Offline signature using distance statistics”,

fication and Authentication a Connectionist Approach”, l Networks, pp. 2948-2951,

ng, K.W. Tse, P.C.K. Kwok gnature verification by the troke positions", Pattern

Authentic Signatures and 50-164, 2009. nd V. Muthukkumarasamy, on based on the Modified of 18th ICPR, pp. 509-512,

a, K. Yashasvini, K. B. Raja, Patnaik, “Off-line Signature ss-Validation for Graph Conference on Electrical 8), pp. 17-22, 2008. -Line Signature Verification 009. . Preez, “Off-line signature

adon transform and a hidden ournal on Applied Signal 2004. An OCR system to read two

a and Devnagari (Hindi)”, 1011–1015, 1997. On Image Analysis By The

Trans. Patt. Anal. Mach. 13, 1988. n method from gray-level n, 9, pp. 62–66, 1979. atistical Learning Theory”,

pport Vector machines for mining and knowledge

Q. Zhuo, “Off-line chinese support vector machines”,

2390–2399, 2005. ari, “On-line and off-line

omprehensive Survey”, IEE Machine Intelligence. Vol. 22

nstein and U. Pal “Off-line ”, 10th DAS, Gold Coast, arch, pp. 282-286, 2012.

376378