Study Of The Off-line Signature Verification Based On The Decision Fusion Thought

The handwritten signature verification is a hot issue in the field of pattern recognition and artificial intelligence. It owns good values on both sides of society and economy. It can be applied in many areas such as finance, communication and security. Currently, the on-line signature verification has applied in some areas, but the off-line signature verification has still stay in the laboratory. Since the basic hardware configurations of the on-line verification can't be satisfied easily, the research of the off-line verification techniques has a much important realistic meaning. The main purpose of this dissertation is to do some beneficial researches in this aspect. Based on the existing techniques, the characteristics of the Chinese signatures, the neural network technique and the thought of decision fusion, a novel method of off-line signature verification is proposed and applied. During our study, we did deeply argumentation in each step of the system. (1) During the collection of signatures, not only the authentic signatures of legal users but also a large number of simple and skilled forgery signatures are collected which will provide dependable experimental data for the efficiency validation of the signature verification method proposed in this dissertation. (2) In the preprocessing step, we remove the color distortion firstly, then the denoise method is discussed. Whereas they can remove some isolated noise, the normal smoothing filters are not inapplicable to the signature verification because they will blur the details especially the boundary of the signature image. So the projection denoise method is designed which can remedy the shortage of the normal smoothing filter. (3) Since the unitary operation usually applied in most of the existing feature abstraction methods will destroy the personal writing style, a new kind of features named component feature is used to reserve the personal writing style such as the size and the location of each component in the signature. (4) Combined with the statistical verification method and the fusion conception, the outstanding classification characteristic of the neural network is also utilized to achieve the last result with much higher conference level. Some concrete problems such as the computing method of the conference level, the definition of the signature pattern, and the design of the neural network are discussed. A series of experiments and evaluating methods are designed for drawing the credible conclusions of our system. By the analysis of these experiments, the fine examination capability of our system is proved.