Research On Contract Handwritten Signature Authentication Based On Siamese Neural Network

There are many occasions and fields where signatures are needed in our daily work and life.Although more and more signatures are transferred from offline to online,and from paper to electronic,offline handwriteen signatures still have a wide application prospect in today’s society,especially involving large interest transations,such as paper purchase contracts,paper private fund contracts,paper IT purchase contracts,etc.,and forged signatures emerge one after another.Therefore,establishing a fast and efficient handwriteen signature authentication tool is still a hot issue at present.With the rapid rise of the field of neural network,offline handwriteen signature,as a kind of biometric identification,has made great achivevments in its technical research,but some shortcomings restrict the development of handwriteen signature authentication application of paper contracts,such as the high similarity between skilled pseudo-signatures and genuine signatures,the small sample set of handwritten Chinese signatures and the difficulty in collecting them.Therefore,this paper makes research and application on the basis of predecessors’ research,using the generation model of Deep Convolution Generative Adversarial Networks(DCGAN)to generate pseudo-signature,using the classical Convolutional Neural Network(CNN)to extract features,and using the Siamese Neural Network(SNN)to verify the handwritten signature of the contract.The main work of this paper is as follows,and all the work in this paper includes signature data collection,signature data preprocessing,DCGAN training and generation,SNN training and verification,and all of them,the corresponding visual interfaces are designed.1.Preprocessing of signature sample data set.Through the visualization platform designed in this paper,paper signatures are scanned into electronic files and collected into the database to become structrued data samples,and the generation of DCGAN is used to generate skilled pseudo-signatures,thus increasing the sample set.Secondly,common techniques are used to preprocess the located signature,including background removal,gray foreground extraction,edge blank clipping,gray normalization and size normaliztion,and the effects of different preprocessing processes on the performance of the model are studied and analyzed.2.Feature extraction and authentication algorithm.In this paper,the signature image is extracted by classical CNN,and the reference signature and the signature to be authenticated are authenticated based on the SNN model.Through experiments,the final choice of CNN structure(ResNet)and super parameters can make the training accuracy of this model 99.8% and the verification accuracy 99.66%(the similarity is set to 0.97).3.Authentication platform.Using the experimental model of the first two items,the application platform of signature recognition is realized.The functions of this application platform include customer information collenction,customer reserved signature collection,business running signature collection,signature authentication,query function analysis,temporary signature authentication and so on.