Fuzzy Logic Based Off-line Signature Verification and Forgery Detection System

Automatic Signature Verification and Forgery Detection has numerous applications in various fields like Bank-Cheque processing, ATM access, Document Authentications etc. Handwritten Signatures have proved pivotal in Authentication of the identity of the person signing the document. This being very important area the problem of signature verification and forgery detection becomes crucial. Handwritten signatures differ largely in their shapes and sizes and the variations are so much that it is difficult to find out a forged signature just by having a glance at the signature. Hence when it is impossible to authenticate a signature we need alternative techniques like the use of computers. In this project we have implemented an offline Signature Verification and Forgery Detection System based on Fuzzy Logic. Handwritten signatures are scanned and they undergo a series of Image preprocessing techniques so as to facilitate feature extraction. Initially, the signatures are binarized, in which the image is converted into Black and White. After that, noise is removed from the picture to get a clear image. This is followed by Thinning in which the basic pattern of the signature is extracted. This is used to take care of the thickness variations. Different types of features are possible for a signature like Distance features, Angle features, Contour measures etc. In our project, Angle features are used since they are found to better than other features in distinguishing the variations. The signature is divided into 3 rows and 20 columns and for each block the average angle is calculated with the bottom-left point as the origin. This is taken as the angle feature for that block. Thus we obtain 60 such features for each signature. We have used Takagi-Sugeno (TS) model for fuzzy modeling. The system utilizes angle features extracted from the signature. Each feature is fuzzified by an exponential Membership Function (MF). Structural parameters of the signatures are also used in order to account for the variations resulting from different handwriting styles of the user. The parameters (for the MFs and Structural parameters) are obtained by training the system with the genuine signatures of the user. During training, the parameters are tuned iteratively in order to minimize the mean square error of the output of the TS model. This training is carried out using Gradient Descent Algorithm. Authentication of any new signature is done using these parameters. We used 25 signatures for training. By training we obtain the structural parameters for this set of signatures for the user. Thus for each feature, Membership Function is determined. The testing is done for 20 genuine signatures, and the threshold is chosen based on the maximum variations in the TS model output, such that optimum performance in signature verification and forgery detection is achieved. The system as described is implemented using MATLAB and Visual Basic and the results are observed. With this implementation, we could achieve an accuracy of 85 percent in detecting the genuine signatures and Forgery Detection was accomplished with the accuracy of 74 percent. The results obtained indicate that computer based signature authentication systems outperform human judgments.