Research On Offline Handwritten Signature Verification And Privacy Protection Based On Deep Learning

Offline handwriting signature identification is an important application field in computer vision technology,and it has a wide range of application scenarios in the fields of contract signing,bill signing and government office.As a biological feature,handwritten signature is a common method of identity authentication,which has full legal effect.Therefore,it is of great practical value to study the automatic identification method of handwritten signature.With the development of the Internet,paperless office is becoming more and more popular.Paperless office can effectively save resources and improve office efficiency.In this process,identity confirmation is essential,and a large number of user signature images will be generated in the process of identity confirmation.These user data appearing in the Internet are easy to be used by web crawlers and used for neural network model training.In view of these situations,this thesis proposes an offline signature identification algorithm based on deep learning,and studies the privacy protection of handwritten signature images.The main work of this thesis includes:(1)The dilated convolution and attention mechanism are added to the dual channel neural network.In this thesis,the specific parameters and positions of the dilated convolution are determined by using the 5-fold cross validation experiment.Experimental results show that hole convolution can significantly improve the performance of the model.In this thesis,channel attention mechanism is used to give different weights to the handwriting features extracted from the network model.The experimental results of 5-fold cross validation show that the addition of attention mechanism can significantly improve the effect of the model.In the final experiment,the dilated convolution and attention mechanism are added to the network model at the same time,and the experimental results beyond the baseline model are obtained.(2)Three channel neural network is introduced into the field of offline signature identification.The first two channels are two signature images,and the data of the third channel is the difference graph of the two signature images.5-fold cross validation experiments show the effectiveness of three channel neural network.In the final experiment,the dilated convolution and attention mechanism are added to three channel neural network.It is found that the dilated convolution will reduce the accuracy of the model,while the attention mechanism will further improve the accuracy of the model.(3)The image scaling attack is introduced into the field of signature privacy protection.Image scaling is an essential part of deep learning.The image scaling attack algorithm is used to modify the signature image,so that the modified signature image will completely become another designated signature after scaling,which can protect the user's signature privacy and ensure that the signature image will not be abused by deep learning.By setting the maximum range of modified pixels,we can ensure that the signature image changes less in the human perspective.