Research And Application Of Handwritten Signature Verification Based On Deep Generative Adversarial Networks

Biometrics technology is used in a wide variety of security applications.Handwritten signature is playing a very special role in the wide series of biological characteristics,because of non-invasive and easy to obtain.But the past methods more or less there are some problems and the insufficiency,this paper puts forward a new idea and method to solve the problem of off-line handwritten signature verification.Offline handwritten signature verification is a classic pattern recognition classification problem,the same as other classification problem,to solve these problems mainly is divided into two parts,feature extraction and classifier design.Thus,this paper concentrates on how to efficiently extract reflect and qualitative characteristics,as well as how to design the classifiers that allows the verification problem to get a good solution.Besides,this model is deployed to live on actual project.This paper focus on off-line handwritten signature verification,the process can be divided into three phases:sample pre-processing,feature extraction and identification of decision-making.Innovation points summarized as follows:(1)In terms of feature extraction,this paper presents a method to extract features from off-line handwritten signatures using discriminator of confrontation network with Deep Convolutionl Generative Adversarial Networks.In this paper,improved methods such as complete batch standardization,avoiding sparse gradients,adding attenuation-type noise and soft label are adopted to make the model has many advantages than previous methods.Such as the model in this paper is more convenient,it can self-learning sample features without manual intervention,besides,the model is more stable and effective.(2)In the sample collection stage,as we all know,artificial intelligence is booming today,machine forgery also needs attention.In this paper,for the first time,a machine forged signature sample is used.Forged signatures are generated using Deep Convolutionl Generative Adversarial Networks Generator and used as skillde forged test sample.(3)In classifier design,this paper proposes a strong classifier-AdaBoostS VMRBF.The classifier is composed of the parameter can dynamically update AdaBoost enhancement method which is proposed in this paper and the SVMRBF.The classifier is used to classify instead of the full-connection layer of the deep network.(4)In the aspect of verification pattern design,this paper obtains a weighted voting verification pattern with dynamically updated weights by combining Writer-Depended and Writer-Independde verification patterns,which combines both convenience and accuracy.Accuracy of model is improved through the weighted voting mechanism to draw the final score and judge.Experimental results show that the accuracy of the model is 92.57%(5)The trained off-line handwritten signature verification model is deployed on a system to run a test.Experimental results and system tests show that the offline signature verification system designed and implemented in this paper has good performance and is stable and reliable.Compared with other existing methods,it has the characteristics of high degree of automation,generalization ability and high accuracy.