| With the rapid development of information and digitalization of economy and society,the collection and application of personal information are becoming more and more extensive.At the same time,people pay more and more attention to their own data security and personal privacy protection needs.Accordingly,various technologies and applications for collecting personal information emerge at the historic moment and become a hot topic in this field.The use of human biometrics for identity authentication is widely respected by various industries because of its unique and stable representation ability,which is difficult to be tampered with and forged.Multimodal identification technology uses the same or multiple feature extraction network to achieve the purpose of identity authentication by fusing multiple biological information of users.Finger vein and fingerprint samples have the characteristics of low collection cost,strong correlation of physical location,high reliability of features and similar texture structure.Therefore,this thesis focuses on the collection of finger vein and fingerprint,and takes them as the main fusion object of dual-mode,and uses convolutional neural network(CNN)as the main method to carry out research.The main research contents of this thesis are as follows:(1)Due to the unique imaging mode of finger vein,the detected image suffers the problem of low image contrast and losing local shape.To address this problem,a multichannel residual neural network(MRNN)is proposed,which enhances the flow of information via multiple channels,builds the learning ability of global information,and excavates the characteristics of deep vein pattern to the maximum extent.Experiments has been carried out on the open data sets MMCBNU,FV-USM and Poly U-FV with the Equal Error Rate(EER)performance of 0.105%,0.101% and 0.155%,respectively,indicating the effectiveness of the model.(2)Due to the rich details of fingerprints and the low complexity of texture features,it is easy to cause large feature approximation of samples between classes.Therefore,the lightweight bilinear convolution neural network(LBCNN)is adopted to abstract the characteristic matrix with high discrimination,so as to improve the efficiency of model verification.Through theoretical analysis and experimental verification has been performed to show that the EER performances are 0.099% and 0.097% on the FVC2004 and SDUMLA-HMT-FP data sets respectively,indicating that the modified network has good classification accuracy.(3)In order to reasonably realize the complementation of bimodal feature information and explore the contribution rate of different feature output layers to performance improvement,on the basis of the experiment of single-level feature fusion structure,the bimodal multi-level fusion network(BMFN)is adopted to efficiently fuse the bimodal features of different levels.In addition,in order to reduce the interference of invalid information on model indicators after feature fusion,enable the network to capture key features quickly and enhance the sensitivity of the model to effective information.Based on BMFN and combined with attention mechanism,the Att-BMFN model is designed to improve the fitting efficiency of the model.On the SDUMLA and MMCBNU-SDUMLA datasets,the EER reaches 0.057% and 0.044% respectively,enabling the model to achieve a better balance between recognition performance and verification efficiency. |
PDF Full Text Request