Offline Handwritten Signature Verification Based On Deep Learning

Offline handwritten signature verification is a technology that uses personal handwritten signature images for identity authentication.It has the advantages of low cost and easy to accept and it has very important applications in security,finance,justice,criminal investigation,and other fields.In recent years,with the rise of deep learning and other methods,the performance of offline handwritten signature verification systems has been continuously improved.However,in practice,high precision offline handwritten signature verification remains challenging due to the small distinction between skilled forgeries and genuine signatures,and the great difference between signatures of the same person at different moments.This thesis studies offline signature verification in depth and proposes a signature verification algorithm based on adaptive region segmentation.The algorithm first segments signature images adaptively.Then the features of cropped regions are extracted by a convolutional neural network and finally the average distance between the features is calculated to verify signatures.To get better performance,this thesis further improves the algorithm in the three aspects of distance measurement,distance fusion,and loss function.Specifically,the main work of this thesis is as follows:1.Signature images are properly preprocessed.In order to avoid background noise,the background of the signature image is removed using a binarization algorithm.To eliminate the inconsistency of the gray distribution and the position of different signatures,grayscale normalization and moment normalization are employed.2.A signature verification algorithm based on adaptive region segmentation is proposed.Aiming at the difficulty that the difference between a skilled forged signature and a genuine signature often lies in the stroke details,this thesis designs a segmentation network based on the spatial transformer network to segment the signature image adaptively,so that subsequent networks can focus on the details of the signature to improve verification performance.Experimental results demonstrate the superiority of the signature verification algorithm based on adaptive region segmentation.3.An improved signature verification algorithm based on adaptive region segmentation is proposed.In order to make the most of the details contained in signatures,this thesis uses the attentive recurrent comparator to compare the features of cropped regions alternately.To better fuse the distances between the signature regions,this thesis designs an adaptive distance fusion module to output the final comparison result.Moreover,according to the characteristics of the signature verification problem,this thesis designs a smooth double margin loss function,which effectively dampens the overfitting problem in the training process and further improves the performance.Experimental results verify the effectiveness and complementarity of the attentive recurrent comparator,the adaptive distance fusion module,and the smooth double margin loss function.In this thesis,a number of experiments are conducted on CEDAR(English),GPDS Synthetic(English),BHSig-H(Hindi),BHSig-B(Bengali),and Chn Sig(Chinese)handwritten signature datasets,which are publicly available internationally.Experimental results show that the proposed algorithm can achieve high precision and multilingual handwritten signature verification,and has high generalization performance.