Research On Offline Chinese Signature Verification Based On Convolution Neural Network

Offline Chinese signatures are mainly used in the signature verification of legal documents and financial documents.The traditional offline Chinese signature verification research method extracts the features of signature images and then uses classification methods to classify them.Feature engineering workload is large,and the extracted features do not necessarily meet scientific research requirements.The feature extraction method has not developed much in recent years,and the extracted signature feature verification results have not improved much.Offline Chinese signature verification research has not achieved relatively large development and practical application at present.The research of this article coincides with the boom of deep learning development,in order to break the limitations of traditional feature engineering in offline Chinese signature verification research,this thesis proposes a method of using deep learning in offline Chinese signature verification.Convolutional neural network excel in image verification.In this thesis,a large number of comparative experiments are used,and finally an effective research method is obtained in offline Chinese signature verification.First,an offline Chinese signature data set was established by collecting volunteer signatures and integrating the signature data set SigComp2011.Then,smoothing denoising,binarization,and normalization preprocessing are performed on off-line Chinese signature data sets.Preprocessing not only remove noise but also ensure that the initial condition of all signature images are consistent.Secondly,three excellent convolutional neural network models Alex Net,GoogleNet and VGGNet are respectively tested on Tensorflow1.0.0.The experimental results show that the Alex Net model performs best and the classification accuracy reaches 99.37%.Super other two models.And attempts to fine-tune the AlexNet model to achieve better results.The problem of small datasets for this thesis reduces a layer of convolutional layers of the network model,model code Alex Net-c1.The problem of occupying most of the parameters of the network for the fully connected layer reduces the network model AlexNet-f with a fully connected layer.From the past scholars believe that the network is deeper and more capable of increasing network recognition capabilities,a layer of convolutional layers has been added,and the network model is recorded as Alex Net-c2.The experimental results show that AlexNet-f performs better than the other two improved network models AlexNet-c1 and Alex Net-c2 both in the training set and the test set.And in the training set,AlexNet-f is also better than the original network model Alex Net.This thesis proposes offline Chinese signature verification based on Alex Net-f model.The test results show that the accuracy rate is 93.5%,the error recognition rate is 2%,and the false rejection rate is 4.5%.Compared with the traditional offline Chinese signature verification method,the proposed method is superior to other methods to some extent and avoids complicated feature engineering.