Research On Handwritten Signature Verification Technology Based On Deep Learning

Today,with the rapid development of digitalization,biometric identification technology has replaced the traditional encryption authentication method.Offline handwritten signature verification technology is one of the most widely studied technologies in the field of biometrics,and its application areas cover many industries such as finance,justice,government and medical.Based on the existing deep learning technology and theory,this thesis studies offline handwritten signature verification based on the Siamese network,and proposes an improved multi-task feature measurement signature verification model and a graph neural network signature verification model,and designs and implements an offline handwritten signature verification system.The main research work is as follows:1.In the multi-task feature measurement signature verification model: First,when the signature image is preprocessed,in addition to the traditional grayscale,size normalization,and denoising methods,this thesis also introduces an additional operation of grayscale image inversion.In this way,the number of signature images can be doubled to complete the expansion of the data set to a certain extent.Secondly,in the feature extraction stage,in view of the problem that the background information in the signature image is relatively large but the stroke information is relatively sparse,the multi-layer feature fusion technology is used to fuse the shallow and deep features,so that the model can not only focus on the entire signature image,the shallow feature information can also model the signature itself on the deep features to extract the detailed features of the signature strokes.Finally,when performing feature measurement and outputting verification results,the model adopts a multi-task feature measurement method.Use multiple tasks to compare the differences between input signatures from different angles at the same time,and use two strategies to fuse the results of multiple tasks to get the final output.Experiments show that compared to a single decision task,the multi-task method has a more accurate verification effect.2.In the graph neural network signature verification model,the overall architecture of the model is still improved based on the Siamese network.In the same preprocessing stage,an additional gray-scale inversion operation is introduced,and then the gray-level co-occurrence matrix(GLCM)features are extracted from the pre-processed signature image.Then,a graph is generated by taking the input multiple signature images as nodes,and the graph neural network Graph SAGE model is used to aggregate them to obtain the embedding vectors of all nodes,and finally the classification judgment is performed to obtain the output.The experiment also explored the verification effect of multiple tasks,and compared the performance with the multi-task feature measurement signature verification model.3.According to the improved multi-task feature measurement signature verification model,this thesis also designs and implements an offline handwritten signature verification system based on the web side,and the performance of this algorithm is verified through testing.This thesis is based on self-built and several public handwritten signature data sets for experiments.The experimental results and system tests show that the algorithm and the system designed and implemented in this thesis have good verification results,which proves the effectiveness of the model in this thesis.