Research And Implementation Of Multi-view-based Handwritten Signature Verification System

As a biological behavior characteristic,handwritten signature has been widely used in financial and judicial fields,and has been one of the most recognized identity identification methods.Depending on the information contained in the handwritten signature can be divided into offline handwritten signature and online handwritten signature.Offline handwritten signatures are less affected by the signature device and are more widely used.Research into techniques for automatically verifying offline handwritten signatures using computers has high engineering applications.However,the automatic verification of offline handwritten signatures is highly challenging because forgers are able to practice repeatedly to write signatures with only minor differences based on the signature of the user to be forged,and the dynamic information of the timing during the writing process is not captured.Therefore,this thesis combines the multi-view representation learning method and focuses on the research and application of offline handwritten signature identification technology with the following work.(1)Handwritten signature verification based on multi-view deep feature representation.Many researches have been conducted based on deep learning techniques for offline handwritten signature verification,and valuable results have been achieved on language-specific datasets.However,limited by the learning capability of deep learning models,such verification models usually suffer from a certain degree of overfitting and usually have diffculty in showing stable results when dealing with different signature datasets.In this thesis,a multi-view representation learning method named Deep Canonically Correlated Denoising Autoencoders is proposed to address this problem to improve the generalization performance of deep features on verification tasks.This method consists of two Denoising Autoencoders with the optimization objectives of maximizing the canonical correlation between the two encoder outputs and minimizing the error between the decoder output and the original unnoised input,and adjusting the impact of the two part objectives on the network optimization by setting hyperparameter thresholds.A series of experiments demonstrate that the proposed method significantly improves the generalization performance of deep features,and reduces the EER to 1.75%,2.40%,and 0.85% on three non-trained datasets of CEDAR,MCYT,and PUC-PR,respectively.Meanwhile,the reasons for the improved verification results are analyzed by feature visualization,and it is demonstrated that the multi-view-based approach can improve the generalization performance of deep features on signature verification tasks.(2)Handwritten signature verification based on multi-feature fusion.Handcrafted features and deep features abstractly represent offline handwritten signatures from different perspectives and contain information from different angles.In this thesis,we address this problem by proposing a multi-view representation learning method called Deep Canonically Correlated Contractive Autoencoder to fuse deep features and handcrafted features to improve the performance of fused features on offline handwritten signature verification tasks.This method is an improvement of the Deep Canonically Correlated Autoencoder that enhances the feature fusion by introducing contraction terms to improve the abstraction of the encoding layer.Experiments show that the proposed method achieves equal error rates of 1.04%,7.64%,2.65%,2.98%,and 6.37% on several different language datasets: CEDAR,UTSig,BHSig-B,BHSig-H,and Sig Comp2011,respectively,and verifies the stability of the model under many different case training sets.(3)Multi-view signature verification system.For the problems of high cost and low efficiency of manual verification of handwritten signatures,which are greatly impacted by the professional and technical level of the authenticator,this thesis develops a multi-view-based offline handwritten signature verification system.Using B/S architecture,the application system is developed using Django,My SQL,Bootstrap and other technologies.Tensorflow,Python and other technologies are used to integrate the model of this thesis and design the verification interface.Using B/S architecture,the application system is developed using Django,My SQL,Bootstrap and other technologies.Tensorflow,Python and other technologies are used to integrate the model of this thesis and design the verification interface.From the process of signature verification,combining the characteristics of the model proposed in this thesis,two functional modules are developed for user model construction and user signature verification.Providing signature verification service for users.