Research On Off-line Chinese Signature Vetiifcation Based On Support Vector Machine

The signature is universally accepted by the public of an authentication method. Becauseof its easy imitation, thus become a fake object. Signature is a kind of method which decides theidentity of writers through the analysis of handwriting verification. In daily life, signatureverification has played a large role. If the identification accuracy is high, it will play a key rolein the various sectors of society, also in the national development. Signature verification hasbecome the hot pot in the field of the computer and pattern recognition, and has a goodapplication prospect.This dissertation studied a series of effective off-line Chinese signature verificationmethod for directed signature, firstly, introduce the overview of Chinese signature verificationand research status at home and abroad; secondly, expound the off-line Chinese signatureverification system flow; finally, made a deep study of the off-line signature image collection,preprocessing, feature extraction, selection and classification identification technology, andmake a series of experiments.According to the characteristics of Chinese signature, the signature image undertook aseries of pretreatment, including smoothing, images of the two values as well as the skeletonextraction. In preprocessing, using genetic algorithm binarazation, and commonly used Otsumethod are compared by simulation, the effect is better. The feature extraction, in order to makeup for the deficiency of single feature, which can accurately reflect the characteristic signature.In this dissertation, the static characteristics are extracted, pseudo dynamic features, and usingmultiple feature fusion method.At present, the support vector machine (Support Vector Machine, SVM) learning methodwas used by many scholars in the signature verification field, and made some achievements.However, for the high dimensionality of the signature data, SVM signature verification haslong training time, test speed and other deficiencies.In view of the above, in the feature extraction, this dissertation uses rough set attributefeature selection method to remove the characteristics of redundancy, in order to achieve thepurpose of reducing feature dimension. This method can not only avoid the dimension disasterproblem number feature extraction, but also improve the training time.Classification decision stage, this dissertation is suitable for finite samples, bettergeneralization ability of support vector machine classification. But the choice of the kernelfunction that directly affect the SVM classification performance. To solve this problem, selectglobal optimization capability, and it’s easy to fall into local optimal genetic algorithm tooptimize the parameters, to obtain the optimal classification parameters. To reduce the feature vectors for training by learning, and establish a maximal margin classification classifier.Experimental results show that, the algorithm can obtain better identification effect.