Research On Chinese Offline Signature Verification Algorithm

As one of the necessary security measures,signature is easy to be imitated by criminals in social life.Therefore,an effective signature verification system has great significance for social security and personal property security,which is worth to study further.Few methods have been proposed for Chinese signature verification due to the lack of datasets required for Chinese offline signature verification.In view of the existing offline signature verification methods at home and abroad,two signature verification systems has been designed in this thesis based on Chinese offline signature verification.Although the first system has a good identification accuracy,it needs to retrain the sample when adding the signature class,which is not ideal in terms of efficiency.And it is suitable for the verification environment where the test sample type exists in the original verification signature database.Therefore,another system(called second system)with ideal efficiency and decision-making is designed,which allows a model to be designed and integrated for a new individual without re-adjusting all parameters.This system has a wide range of practical applications.The main contents of this thesis are as follows:Firstly,the existing mature technology has been used to preprocess the collected offline signature,and the pixel redundancy problem has been solved by cutting the boundary and reducing the image.Furthermore,the binarized signature image has been skeletonized to obtain the boundary shape of the signature image,which lays a good foundation for subsequent signature feature extraction in different ways.Secondly,a signature verification system based on two-level classifier similarity matching has been designed.The GLOH matching system is integrated with the improved shape context authentication system by using the string structure.Only one-to-one matching is supported in the original shape context,and the constraint will cause mismatching.Therefore,a method of increasing virtual point matching has been proposed to improve the performance of the shape context in this thesis by allowing each sample point of one signature to be in another signature.And a new method of measuring the similarity value between Test and real signature has also been proposed to make the mapping function H(?)satisfy the minimum of global matching cost.Both the two matching algorithms of the verification system adopt RANSAC to eliminate mismatched points and achieve reasonable signature similarity judgment.Thirdly,a fuzzy similarity metric signature verification system based on interval symbol model has been designed.Considering the feature vector of the existing texture features is too large,the improved local binary pattern(LBP)features has been merged with the gray level co-occurrence matrix texture features to construct an interval symbol model for each feature of the real sample.A new fuzzy similarity measure algorithm has been proposed to calculate the similarity between test samples and corresponding interval value symbol models.A relatively complete offline signature verification system has been constructed through training experiments and adjusting parameters.Finally,experimental design and data analysis has been carried out.The average error acceptance rate and error rejection rate of different samples in the two signature verification systems are obtained by testing the accuracy of the two systems when the training samples of different categories were set to 20.The samples used in the system test were 20 skilled forged signatures of different signature categories and 20 random forged signatures.The results show that the first system has a better recognition rate than the second system,whether it is skilled forgery or random forgery.In conclusion,the first system exhibits better signature authenticity identification capability than the second system.At the same time,the second system not only solves the problem of inter-class variability between features,but also does not need to be retrained when adding new classes to the system.The second system is cheaper than the first one in terms of memory usage and computation time.