Global Features for the Off-Line Signature Verification Problem

Please download to get full document.

View again

of 5
9 views
PDF
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Document Description
Global Features for the Off-Line Signature Verification Problem
Document Share
Document Tags
Document Transcript
  Global Features for the Off-Line Signature Verification Problem Vu Nguyen, Michael Blumenstein School of Information & Communication TechnologyGriffith University, Australia { Vu.Nguyen2, M.Blumenstein } @griffith.edu.au Graham Leedham School of Science & TechnologyUniversity of New England, AustraliaGraham.Leedham@une.edu.au Abstract Global features based on the boundary of a signatureand its projections are described for enhancing the processof automated signature verification. The first global fea-ture is derived from the total ’energy’ a writer uses to cre-ate their signature. The second feature employs informa-tion from the vertical and horizontal projections of a signa-ture, focusing on the proportion of the distance between keystrokes in the image, and the height/width of the signature.The combination of these features with the Modified Direc-tion Feature (MDF) and the ratio feature showed promis-ing results for the off-line signature verification problem.When being trained using 12 genuine specimens and 400random forgeries taken from a publicly available database,the Support Vector Machine (SVM) classifier obtained anaverage error rate (AER) of 17.25%. The false acceptancerate (FAR) for random forgeries was also kept as low as0.08%. 1. Introduction Handwritten signatures have long been accepted as anofficial means to verify personal identity for legal purposeson such documents as cheques, credit cards, contracts andwills. The handwritten signature is therefore well estab-lished and accepted as a behavioural biometric. Consid-ering the large number of signatures verified daily throughvisual inspection by people, the construction of a robust andaccurate automatic signature verification system has manypotential benefits for ensuring authenticity of signatures andreducing fraud and other crimes. As a consequence, re-search into signature verification has been vigorously pur-sued for several decades, particularly with reference to off-line verification [1,2].Off-line verification refers to when the signature is onlyavailable as a static image, typically obtained after it hasbeenwrittenonpaperusingavarietyofwritinginstruments,with no reference to the sequence and timing of the pen-strokes, which created the signature. When the sequenceof the pen-strokes is available the process is referred to ason-line signature verification.The off-line signature verification problem is more chal-lenging due to the lack of valuable behavioural informationabout how the person created the signature in terms of pen-point velocity and accelerations, writing pressure and strokesequence. Access to this on-line information during thetraining phase, has been demonstrated to result in improvedperformance in off-line signature verification systems [3].Like any other pattern recognition scheme, one crucialaspect of a signature verification system includes appropri-ate feature extraction procedures. Thus, new feature extrac-tion techniques are being extensively explored. The surveyby Weiping  et al.  [1] summarises some previously investi-gated features and approaches. In this paper we report ourlatest results in the pursuit of new global and local featuresfor addressing the off-line signature verification problem. Figure 1. Automatic Offline Signature Verifi-cation System 2009 10th International Conference on Document Analysis and Recognition 978-0-7695-3725-2/09 $25.00 © 2009 IEEEDOI 10.1109/ICDAR.2009.1231300  2. Methodology A general automated off-line signature verification sys-tem that employs only genuine specimens and other ran-dom signatures in the training phase is depicted in Figure 1.After being preprocessed through conversion to a portablebitmap (PBM) format, the boundary of the signature imageis extracted. The Modified Direction Feature (MDF) andother global features are then extracted from the boundaryin the feature extraction stage.To perform learning and classification tasks for signatureverification, machine learning techniques such as ArtificialNeural Networks, Support Vector Machines, or a thresholddecision method can be employed. The signature model isconstructed in the learning phase using predefined param-eters, which are invoked in the classification phase. It ishypothesised that there exist a set of parameters that is uni-versal for every writer. 3. Feature Extraction Features extraction plays an important role in a signa-ture verification system, and the features extracted can becategorized as global or local features. Global features treatpatterns as a whole whilst local features are extracted froma portion or a limited area of the pattern. Global featurestend to be less sensitive to variation or noise whilst localfeatures provide more detailed information.A suitable combination of global and local features haspreviously been found to improve a classifier’s ability torecognize forgeries and to tolerate intrapersonal variances[1]. In this research, the combination of the Modified Di-rection Feature with three other global features is reported. 3.1. Modified Direction Feature The Modified Direction Feature (MDF) feature extrac-tion technique [4] employs the location of transitions frombackground to foreground pixels and the direction at transi-tions in the vertical and horizontal directions of the bound-ary representation of an object. At each transition, the Lo-cation of the Transition (LT) and the Direction Transition(DT) values are recorded.The LT value is calculated by dividing the position valuewhere a transition occurs by the distance across the entireimage in a particular direction, whilst the DT is obtainedby examining the stroke direction of an object’s boundaryat the position where the transition occurred. Finally, theLT and DT values are locally averaged to reduce the fea-ture vector size. The local averaging process also makesthe MDF feature extraction technique more tolerant to localnoise and intrapersonal variations. 3.2. Features from Energy Information The first global feature investigated in this work was de-rived from the total energy a writer uses to create their sig-nature. It is hypothesised that the planned execution of thesignature uses the same amount of energy, and this infor-mation can be extracted from the trajectory of a signatureusing its digital image. Although it is difficult to preciselyestimate the amount of energy used between any two pointson the signature’s trajectory, it is of interest to determinewhether information extracted from the energy of specialsegments (e.g. cusp to cusp, end to end) could increase ver-ification accuracy. Recent work of Qiao  et al.  [3] showedthat the signature trajectory could effectively be recoveredusing the information from on-line registration. As a con-sequence, it is expected that an energy-based feature couldenhance this approach.For simplicity, in this work, we only extract energy in-formation from the boundary of the whole signature image.This energy information was decomposed into horizontaland vertical components. The features extracted are thevalues from the signature width divided by horizontalenergy ( e h ), signature height divided by vertical energy( e v ), and  min ( e h ,e v ) /max ( e h ,e v ) . The following is thepseudo code for the extraction of these values: function  Energy(binary image im) 1:  Sum h  ←  0  { horizontal energy } 2:  Sum v  ←  0  { vertical energy } 3:  Sum L  ←  0  { energy from the left diagonal } 4:  Sum R  ←  0  { energy from the right diagonal } 5:  h  ←  im height 6:  w  ←  im width 7:  Mark all black pixels of im as unvisited 8:  for  each black pixel p of im  do 9:  if   p is not visited  then 10:  track(p) 11:  end if  12:  end for 13:  e v  ←  h/Sum v 14:  e h  ←  w/Sum h 15:  e hv  ←  min ( Sum v ,Sum h ) /max ( Sum v ,Sum h ) 16:  e lr  ←  min ( Sum L ,Sum R ) /max ( Sum L ,Sum R ) 17:  e hv  ←  Sum v  +  Sum h 18:  e LR  ←  Sum L  +  Sum R 19:  e r  ←  min ( e hv ,e LR ) /max ( e hv ,e LR ) 20:  return  { e v ,e h ,e hv ,e lr ,e r } end functionprocedure  track(pixel p) 1:  mark p as visited 2:  for  each neighbour pixel pNb of p  do 3:  if   pNb is not visited  then 1301  4:  { subscripts y and x mean the row and column of apixel } 5:  Sum v  ←  Sum v  +  |  pNb y  −  p y | 6:  Sum h  ←  Sum h  +  |  pNb x  −  p x | 7:  if   (  pNb y  −  p y )  ∗  (  pNb x  −  p x ) = 1  then 8:  Sum L  ←  Sum L  + 1 9:  end if  10:  if   (  pNb y  −  p y )  ∗  (  pNb x  −  p x ) =  − 1  then 11:  Sum R  ←  Sum R  + 1 12:  end if  13:  track(pNb) 14:  break  15:  end if  16:  end forend procedure 3.3. The Maxima Feature Information from horizontal and vertical projections hasbeen adopted by researchers for image alignment [5], andprofiling signatures for off-line signature verification [6,7].From our observations, handwriting boundary projectionscontain valuable information about key strokes, and nearstraight curves, which are horizontal or vertical. It is sug-gested that the proportion of the distance between two mainstrokes, either both horizontally, or both vertically, and theheight and width of a writer’s signature remain relativelystable among signature specimens and thus can be used as aglobal feature in the offline signature verification problem.The formula of the ’maxima’ feature is defined as: (  col 1 max  −  col 2 max  W  ,  row 1 max  −  row 2 max  H   )  (1)Where W and H are the width and height of the signatureimage,  row 1 max ,  row 2 max ,  col 1 max , and  col 2 max  are the in-dexes of the columns and rows that have the highest numberof black pixels among the rows or columns. 3.4. Ratio Feature The next global feature we used in conjunction with theMDF, the energy-based feature and the Maxima feature wasthe Ratio feature, which employed the width and height in-formation of the rectangular box that encloses the signa-ture pattern. The Ratio feature has been widely used by re-searchers in cursive handwritten character recognition andsignature verification [4,8]. In our previous work with neu-ral networks [9], the Ratio feature (R1) was calculated usingthe following formula in order to generate a feature valueranging from 0 to 1: Ratio  = arctan( width/height ) π/ 2  (2) Table 1. Experimental Settings Phase Genuine Random TargetedSetting I  Training 12 400 -Testing 12 59 15 Setting II  Training 20 400 -Testing 4 59 15However, our experiments showed that a more straightfor-ward calculation of the Ratio feature would increase the ac-curacy. The alternative Ratio feature value is calculated bydividing the minimum by the maximum of the width andheight values. The following is the alternative formula forthe Ratio feature (R2): Ratio  =  min ( width,height ) max ( width,height )  (3) 4. Experimental Settings To enable result comparison with other work, the pub-licly available gpdsSIGNATURE [10] handwritten signa-ture database was employed in our research. This large cor-pus consists of 160 signature sets with 24 genuine and 30targeted forgeries in each set.It is essential to select a reasonable number of genuinesamples for the training process to construct a signaturemodel that is tolerant to intrapersonal variances and effec-tively rejects imitations [11]. In each test, 12 genuine sig-natures and 400 random forgeries were employed for train-ing. Whilst the random forgeries are easier to collect, therandom forgeries were chosen from 100 randomly selectedwriters, four genuine signatures from each. To representrandom forgeries for testing, 59 genuine signatures werechosen from the remaining 59 writers. Table 1 summarizesthe sample configuration for the training and testing phases.For each signature set, the experiment is repeated 30 timesin order to obtain a more stable result.Support Vector Machines (SVMs) [12] were employedto construct the signature models. The choice of a suitablekernel for SVM training varies between classification prob-lems and feature extraction techniques. Within the area of off-line signature verification, Ferrer  et al.  [10] and Lv  et al.  [13] obtained better results with the Radial Basis Func-tion (RBF) kernel. Meanwhile, Justino  et al.  [14] achievedtheir best results with the linear kernel. Results from ourprevious research [9] suggested that the MDF and its vari-ants performed best with the RBF kernel. Consequently,the RBF kernel was chosen and the experiments were con-ducted using SVM light v6.01 [15]. 1302  5. Experimental Results A signature verification system can be challenged byforgeries with varied levels of skill. It is much easier tocompare the performance of different approaches againstrandom forgeries than against targeted forgeries due to theunavailability of a standard signature corpus. As can beseen in Table 2, a false acceptance rate (FAR) of 0.08%for random forgeries was obtained with the proposed fea-ture combination. This figure is comparable to the resultsof other researchers [5,10,14,16,17] and can be explainedby the LT values of the MDF, which were effectively gener-alized by the local averaging process.It is natural that a system has a substantially higher FARfor targeted forgeries than for simple forgeries, or randomforgeries. However, some researchers observed an oppositetrend regarding the proportion of random forgery errors asopposed to targeted forgery errors on other databases withdifferent approaches [14,17]. Figure 2. The FRR and FAR values obtainedwith different values of gamma when the RBFkernel was employed When the proposed system was challenged by targetedforgeries, the lowest average error rate of 17.25% was ob-tained with the combination of MDF and the global fea-tures. Figure 2 shows the false rejection rate (FRR) andfalse acceptance rate (FAR) values for different settings of sigmausedin theRBFkernel of SVM. Thisresult iscompa-rable to the AER of 17.17% as reported by Zhang  et al.  [18](although they employed a smaller database with 80 sets of signatures). It is worth mentioning that, to obtain our result,only 12 genuine signatures were used for the training pro-cess as compared to 20 genuine signatures in Zhang  et al. ’sresearch. In experiments using Setting II, our approach ob-tained a superior AER of 15.77%. However, training with alarger number of genuine signatures reduces the feasibilityof many systems. Figure 3. The distribution of score for gen-uine and targeted forgeries of varied skill If it is hypothesised that the score for the genuiness of genuine signatures would follow a normal distribution re-gardless of which approach was employed, the distributionof the score curve for forged signatures would likely take ona similar shape. This is due to the fact that the genuine spec-imens were randomly selected for imitation. Consideringthe above assumptions, different approaches should eitherhave their best AER values approximating their equal errorrate (EER, where FAR equals FRR) or the database em-ployed is less representative (Figure 3 - Less Skilled curve).Using the above arguments, the FRR and the targeted forg-eries’ FAR values obtained in this work support the premisethat the quality of the gpdsSIGNATURE database is supe-rior to other signature corpuses (Zhang  et al.  [18] exhibitinga 2.74% difference, Bansal  et al.  [19] with a 10.38% differ-ence). 6. Conclusions The performance of the MDF feature extraction tech-nique in conjunction with three other simple global featureshas been investigated. With the proposed feature set, an au-tomated off-line signature verification system obtained anAER of 17.25% when 12 genuine signatures were used inthe training process.The results obtained using the energy-based featuresdescribed in this paper (in conjunction with other globaland local features) encourage further investigation in au-tomatic signature verification using the energy informationextracted from the signature trajectory. Although it is not 1303  Table 2. The error rates of feature combi-nations obtained using Setting I (FAR1 andFAR2 are false acceptance rates for randomand targeted forgeries respectively) Features FRR FAR1 FAR2 AER MDF 18.14% 0.11% 18.32% 18.23%MDF R1 18.42% 0.13% 18.29% 18.36%MDF R1 M 17.99% 0.13% 18.82% 18.41%MDF R2 18.56% 0.10% 17.49% 18.03%MDF E 18.01% 0.09% 16.98% 17.50%MDF E R1 18.09% 0.09% 16.90% 17.50%MDF E R2 17.77% 0.08% 16.96% 17.37%MDF E R2 M 17.25% 0.08% 17.25%  17.25% easy to recover dynamic information directly from off-lineimages [3], energy information extracted from certain partsof a signature may help increase verification accuracy. AsSVMs can still perform well when some of the featuresare missing [15], energy-based features could be extractedat the micro-level from an uncertain trajectory, rearranged,and could assist in lowering the verification error rate.Future work will include (1) the automation of off-linehandwritten signature trajectory recovery; (2) the extractionof energy information from different parts of the signature;(3) appropriately organizing the extracted information foruse with SVMs. References [1] H. Weiping, Y. Xiufen, and W. Kejun, “A survey of off-linesignature verification,” in  Intl. Conf. on Intelligent Mecha-tronics and Automation , pp. 536–541, Aug. 26-31 2004.[2] R. Plamondon and S. Srihari, “Online and off-line hand-writing recognition: a comprehensive survey,”  PAMI, IEEE Trans. on , vol. 22, pp. 63–84, Jan 2000.[3] Y. Qiao, J. Liu, and X. Tang, “Offline signature verificationusingonlinehandwritingregistration,” in CVPR,IEEEConf.on , pp. 1–8, June 2007.[4] M. Blumenstein, X. Y. Liu, and B. Verma, “A modified di-rection feature for cursive character recognition,” in  Proc. IEEE Intl. Joint Conf. on Neural Networks IJCNN  , vol. 4,pp. 2983–2987, 25–29 July 2004.[5] A. Piyush Shanker and A. N. Rajagopalan, “Off-line signa-ture verification using DTW,”  Pattern Recognition Letters ,vol. 28, no. 12, pp. 1407–1414, 2007.[6] B. Fang, C. H. Leung, Y. Y. Tang, K. W. Tse, P. C. K.Kwok, and Y. K. Wong, “Off-line signature verification bythe tracking of feature and stroke positions,”  Pattern Recog-nition , vol. 36, no. 1, pp. 91–101, 2003.[7] T. Wei and Q. Yizheng, “Off-line chinese signature verifi-cation based on optimal matching of projection profiles,”in  Proc. of the 6th World Congress on Intelligent Controland Automation WCICA , vol. 2, (Dalian, China), pp. 10240–10244, June 21 - 23, 2006.[8] F. Camastra, “A SVM-based cursive character recognizer,” Pattern Recognition , vol. 40, pp. 3721–3727, 2007.[9] V. Nguyen, M. Blumenstein, V. Muthukkumarasamy, andG. Leedham, “Off-linesignatureverification using enhancedmodified direction features in conjunction with neural clas-sifiers and support vector machines,” in  Proc. 9th Intl Conf on Document Analysis and Recognition ICDAR , vol. 2,pp. 734–738, 23–26 Sept. 2007.[10] M. A. Ferrer, J. B. Alonso, and C. M. Travieso, “Offline ge-ometric parameters for automatic signature verification us-ing fixed-point arithmetic,”  PAMI, IEEE Trans. on , vol. 27,no. 6, pp. 993–997, 2005.[11] R. Plamondon and G. Lorette, “Automatic signature verifi-cation and writer identification - the state of the art,”  Pattern Recognition , vol. 22, no. 2, pp. 107–131, 1989.[12] V. N. Vapnik,  Statistical Learning Theory . John Wiley &Sons, 1998.[13] H. Lv, W. Wang, C. Wang, and Q. Zhuo, “Off-line chinesesignature verification based on support vector machines,” Pattern Recognition Letters , vol. 26, no. 15, pp. 2390–2399,2005.[14] E. J. R. Justino, F. Bortolozzi, and R. Sabourin, “A com-parison of SVM and HMM classifiers in the off-line signa-ture verification,”  Pattern Recognition Letters , vol. 26, no. 9,pp. 1377–1385, 2005.[15] T. Joachims,  Advances in Kernel Methods - Support Vector  Learning, , ch. Making large-Scale SVM Learning Practical.MIT-Press, 1999.[16] J. F. Vargas, M. A. Ferrer, C. M. Travieso, and J. B. Alonso,“Off-line handwritten signature GPDS-960 corpus,” in  Proc. Ninth International Conference on Document Analysis and  Recognition ICDAR , vol. 2, pp. 764–768, 23–26 Sept. 2007.[17] L. S. Oliveira, E. Justino, and R. Sabourin, “Off-line sig-nature verification using writer-independent approach,” in Proc. International Joint Conference on Neural Networks IJCNN  , pp. 2539–2544, 12–17 Aug. 2007.[18] Z. Tai-Ping, F. Bin, X. Bin, H.-X. Chen, M. Chen, and Y.-Y.Tang, “Signature envelope curvature descriptor for offlinesignature verification,” in  Intl. Conf. on Wavelet Analysisand Pattern Recognition ICWAPR , vol. 3, pp. 1262–1266,2007.[19] A. Bansal, D. Garg, and A. Gupta, “A pattern matching clas-sifier for offline signature verification,” in  Proc. First In-ternational Conference on Emerging Trends in Engineeringand Technology ICETET  , pp. 1160–1163, 16–18 July 2008. 1304
Similar documents
View more...
Search Related
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks