Image Recognition Of Counterfeit Currency Based On Deep Learning

The authentication of RMB based on mobile phone image is the key technology to realize the authentication of RMB by mobile phone photography,and it is also a crucial step to realize the authentication application of RMB.This paper focuses on image denoising,image correction,authentication based on traditional texture features and authentication based on deep convolution features.The details are as follows:(1)A data set of genuine and fake RMB has been established which contains 1000 genuine banknotes and 137 fake banknotes.With the assistance of various parties,137 original samples of counterfeits were collected,including 50 total fake banknotes,41 counterfeits whose left part is genuine while the right is fake,45 counterfeits whose left is fake while the right is genuine,and one counterfeit which left and right parts both are genuine.(2)A combined median filter based on RMB texture analysis is proposed and implemented.In view of the rich texture characteristics of RMB,this paper divides RMB into small blocks,carries out texture analysis on each block,and designs a texture complexity factor for quantifying texture.When the texture complexity factor in the image block is small,large square median filter is used,and vice versa,small linear median filter is used.The experimental results show that it has better visual effect than the traditional single median filter,and has better performance under PSNR and SSM standards.The weighted function of error penalty based on ideal line is designed,and the weighted least squares linear fitting method is improved,which effectively reduces the influence of outliers on the results of edge line fitting.(3)A RMB authentication method based on D-S fusion of GLCM texture features and Gabor texture features is proposed and implemented.The anti-counterfeiting area of RMB image is selected through comparative study.The feature extraction method based on GLCM texture feature extraction and the feature extraction method based on Gabor texture feature extraction are used to extract the feature of the selected area respectively.Then the SVM with RBF core in LIBSVM is used to classify the texture feature.The experimental results show that the recognition effect based on single feature is not good.Using D-S evidence theory to fuse these two features,the authentication of RMB is identified.The comparative experiments show that the identification method based on D-S multi-feature fusion can effectively improve the recognition rate based on single feature identification method,and reduce its false recognition rate and rejection rate.(4)The authentication of RMB based on VGG-19 convolution neural network is proposed and implemented.VGG-19 convolution layer is used to extract features of RMB.Two 1×1 convolution layers are designed to replace the full connection layer to fuse and reduce the dimension of extracted features.Finally,SVM with RBF core in LIBSVM is used to classify the features,and the authentication of RMB is identified.Compared with the authentication method based on traditional feature,the authentication method based on the convolutional neural network not only has higher accuracy,loweComputer Graphics and Image ProcessingAcademic Press,1979:394-407 r false recognition rate and rejection rate,but also has better fault tolerance for region translation and generalization ability.