Study On Signature Verification

Signature, as an important method for identification verifying, has been widely applied in the field of finance and politics in the society. On the other hand, more and more law cases which involve the illicit forgery of the signature are bringing about tremendous financial loss to the nation and numerous individuals. Presently, the verification of the signature is mainly carried out manually by experts, thus this critical procedure is usually omitted in most financial activities that need to be handled in real time. The research of the automatic handwritten signature verification technologies is aimed to provide a scientific and precise, highly automatic verification system that can effectively prevent relevant financial crimes.The thesis aims at the automatic verification techniques towards signature images, and is mainly divided into three parts, which are Preprocessing, Feature Extraction and Classification. The thesis first analyzed the influence image background may exert on the verification result, and proposed the method of highlighting the signature in order to remove background noise; Based upon this, considering that the size and position of the signature is always uncertain, proposed the method that can normalize the margin between characters and the height of the characters. After the preprocessing, ideal signature images which have approximately the same size and position can be obtained.As for feature extraction, the thesis divided signature features into shape and sham dynamic features. The thesis proposed a feature extraction method based on strokes, which extracts stroke texture information through Gabor filter and stroke direction information through decomposing strokes. The paper also proposed the method which extracts statistical gray level feature and low gray level feature to represent the actual dynamic information (Sham Dynamic Feature).After the extraction of effective features, the thesis used Support Vector Machine and Euclidean Distance Classifiers to make the decision, according to the characteristics of different features. Then, the decisions made by different classifiers are merged through the method of Information Fusion. To improve the precision and effectiveness of the system, the paper divided forged signatures into simple forgeries and skillful ones, and utilized two levels of classification systems to make the final decision.The thesis used the Image Processing and Pattern Recognition theories to partition the signature image, extract features and classify through merging multiple classifiers in multi-level expert systems. The experiment result indicates that the system satisfied anticipated requirements in both robustness for signature in the same category and sensitivity for ones in different category.