Offline signature verification based on improved extracted features using neural network

Abstract

The verification of handwritten signatures is one of the oldest and the most popular biometric authentication methods in our society. A history which spans several hundred years has ensured that it also has a wide legal acceptance all around the world. As technology improved, the different ways of comparing and analyzing signatures became more and more sophisticated. Since the early seventies, people have been exploring how computers may aid and maybe one day fully take over the task of signature verification. Based on the acquisition process, the field is divided into on-line and off-line parts. In on-line signature verification, the whole process of signing is captured using some kind of an acquisition device, while the off-line approach relies merely on the scanned images of signatures. This thesis addresses some of the many open questions in the off-line field. In this thesis, we present off line signature recognition and verification system which is based on image processing, New improved method for features extraction is proposed and artificial neural network are both used to attend the objective designed for this thesis, Two separate sequential neural networks are designed, one for signature recognition, and another for verification (i.e. for detecting forgery). Verification network parameters which are produced individually for every signature are controlled by a recognition network. The System overall performs is enough to signature recognition and verification signature standard and popular dataset, In order to demonstrate the practical applications of the results, a complete signature verification framework has been developed, Which incorporates all the previously introduced algorithms. The result was very good comparing with other work, sensitivity was more than 0.94% and 0.80% for training and testing data respectively, and for specificity it was more than 0.78% and 0.74 for training and testing data respectively, and for specificity. The results provided in this thesis aim to present a deeper analytical insight into the behavior of the verification system.