Research On Offline Chinese Signature Verification Method Based On Transformer

Offline handwritten signature,as a relatively stable and easy-to-obtain behavioral biometric traits,plays an important role and is widely applied in many fields such as economics,politics and social activities.As a symbol of personal identity,signature involves huge interests,and forged signature not only damages personal rights,but also seriously disrupts social order.Therefore,the study of offline signature verification is of great significance.Currently,convolutional neural networks are widely used in signature verification systems,but there are certain limitations.Due to the outstanding performance of Transformer network in the field of natural language processing,this thesis studies the application method of Transformer on offline Chinese signature verification.In the field of offline Chinese signature verification,there are very few publicly available datasets,which greatly constrains the development of Chinese signature verification tasks.In order to solve this problem,this thesis presents a convenient method to collect offline handwritten signature samples.The genuine signatures of the students in the four years of university were obtained by collecting the registration sheet of examinations in a college over the years,and used the automatic method to generate forged signature samples to construct a large offline Chinese signature dataset.Other than that,this thesis preprocesses the signature image to eliminate noise and other interference factors,reducing the difficulty of handwritten signature verification.The signature samples collected in this thesis satisfy the complexity of the actual situation and lay the foundation for the research of offline signature verification technology.Offline Chinese signature has the characteristics of fewer words,strong personality and easy change.In order to extract valid information from handwritten signature strokes,this thesis proposes an offline Chinese signature verification technology based on improved Transformer network.This method splits the signature image into fixed-size patches so that subsequent network can focus on the details of the strokes of the signature.In addition,by adding depthwise separable convolution,the Transformer network is able to focus not only on the global features of the signature image,but also on the local features,which enhances the feature extraction ability of the network for valid information.Hence it provides a new idea for the research in the field of offline Chinese signature verification.This thesis conducts a large number of comparative experiments based on the self-built offline Chinese signature dataset.The experimental results show that the improved Transformer model proposed in this thesis has a better verification effect on the offline Chinese signature datasets than the compared network models.