Study On Key Problems In Off-line Signature Verification

Handwritten signature verification technology is the most socially and legally accepted means of personal identification, and is one of the most important ways of biometric identification. Automatic signature verification aims at endowing computers with the ability to the detection of one or more category of forged signatures. As a scientific issue, Automatic signature verification is a typical pattern analysis, understanding and classification problem, closely related to many disciplines such as Pattern Recognition, Computer Vision, Intelligent Human-Computer Interaction, Computer Graphics, and Cognitive Psychology etc. Its research achievements would greatly contribute to the development of these disciplines. While as one of the key technologies in Biometrics, automatic signature verification technology is believed having a great deal of potential applications in public security, law enforcement, information security, and financial security.After nearly 30 years'development, automatic signature verification technology has made great progress, on-line signature verification system has been performed in practical applications. However, as one of important components among handwritten signature verification, off-line signature verification techniques are currently far from mature. Automatic off-line signature verification techniques, especially, the detection of skilled forgeries have a great number of challenges to be solved.In this dissertation, some key issues on feature extraction method, the building of stable signature model and multi-classifier combination problem in signature verification have been studied, and the main contributions are listed below:(1). Non-linear rotation of signature patterns is one of the major difficulties to solve in off-line signature verification. This paper presents two models utilizing rotation invariant structure features to tackle the problem. In principle, the elaborately extracted ring-peripheral features are able to describe internal and external structure changes of signatures periodically. In order to evaluate match score quantitatively, discrete fast Fourier transform (FFT) is employed to eliminate phase shift and verification is conducted based on a distance model. In addition, the ring-hidden Markov model (HMM) is constructed to directly evaluate similar between test signature and training samples. Experimental results demonstrated that the proposed methods were effective in self-made database and a subcorpus of MCYT. (2). A training sample quality evaluation and selection method is proposed. There are inevitable variations in the signature patterns written by the same person, some of them show a lot of change due to various bad writing conditions. If such samples are used for training to establish verification models, statistics of the models will not be stable. With respect to the side effect of outlier training samples for stable statistical model and threshold estimation, we propose a selection strategy to improve the performance of system. Experimental results demonstrated that the proposed method were effective to improve verification accuracy.(3). Edge Orientation Distance Histogram (EODH) descriptor is proposed. This feature not only represents the shape of signature, but also describes the direction and smoothness of the writing, and the relations among edge points in the same orientation possess lots of information (e.g. the characteristic geometry and topology of a signature) which may be neglected by forgers. Therefore, this feature is adapted to the detection of skilled forgeries.(4). Multi-level weighted fuzzy classifier is proposed. According to the characteristic of EODH, Multi-level weighted fuzzy classifier is the developed base on the weighted fuzzy classifier. Experiment results shows that the performance of multi-level fuzzy classifier is better than the weighted fuzzy classifier for EODH features.(5). A multi-classifier combination approach based on serial architecture and decision level for off-line signature verification was introduced. The combination system is composed of two subsystems, which one subsystem is based on EODH and multi-level weighted fuzzy classifier, and another subsystem based on directional gradient density feature and weighted fuzzy classifier. Experiment shows that the proposed approach were effective to fuse the advantages of two subsystems and improve verification accuracy.