Research Of The Key Techniques On Off-line Chinese Signature Verification System

Signature is one of the biological features used abroad to verify identity, because of the significant advantages of signature verification as a way of personal identification, it has a wide perspective in applications in electronic-bank, military affairs, electronic-business, communication, office automation and engineering files, the research into signature verification is indeed significant both to practical application and the development of science. First, the application background and history of development of handwritten signature verification techniques are briefly introduced in this thesis. The state of the handwritten signature verification researches at home and abroad is also surveyed. The nature of the off-line signature verification problem and difficulties are analyzed, some key techniques and main algorithms of off-line Chinese signature verification have been studied, and an off-line Chinese signature verification system has been developed.The main research work including:In the process of preprocessing and feature extraction, after consulting a fund of information, a deep study of existing algorithm has been made. In this thesis, most of the methods used for preprocessing of the signature image are ripe algorithms for off-line Chinese signature verification. The signature's static shape feature and pseudo dynamic features have been extracted in feature extraction, and a new pseudo dynamic feature, the high gray stable aera feature of the signature has been proposed.After feature extraction, feature selection has been made. A method combining Bhattacharyya distance which is a distance in probability method and the feature's self characteristic has been proposed and used, eight features have been abtained. In this way, the verification time is shortened, expecially the verification time of the neural network.Five verification methods have been used in the process of verification. First, the K-NN method and the method of weighted Euclidean distance have been used. Then,a deep analysis of the theory and the key algorithm of the neural network have been made, the designing, comparison and implementation of the BP (Back-Propogation) and RBF (Radial Basis Function) Neural Network have been made. Experimental results prove that the RBF NN performs better than BP NN in terms of classification accuracy. Last, a Lagrangian Support Vector Machine (LS VM) classifier is proposed in terms of deeply study of support vector machine (SVM) ,which not only runs faster than standard support vector machine classifiers but also is easy to implement with satisfactory result.