A Study On Multi-finger Vein Recognition Based On Deep Learning

In the field of identity verification,biometric identification has become the focus of researchers’ study because of its advantages of quickness,convenience and security.Finger vein recognition,as a new biometric identification technology,has the advantages of high accuracy,high stability as well as difficult to steal,and has broad application prospects.In the past,the research on finger vein recognition mainly focused on the field of single-finger vein recognition.Considering that multi-finger vein images contain more features than single-finger vein images,this thesis proposes that multi-finger vein images can be used for identity recognition based on deep learning in a similar process as singlefinger vein recognition.The process can be divided into several stages:data acquisition,data preprocess,ROI region extraction,feature extraction,verification as well as identification,and corresponding solutions are put forward for some difficulties in the process.The specific contents of this thesis are as follows:1.A multi-finger vein image data set is collected and constructed by ourselves,which included 4,100 multi-finger vein images from 41 subjects’left and right hands.It contains the low-quality pictures that may be collected by the equipment in practical application,and marks them.2.A multi-finger vein ROI detection model GVMVN based on sliding vertex is proposed.The model uses multi-target detection algorithm for multi-finger vein ROI detection.The sliding vertex method is used to determine the four vertices,and the ROI region consistent with the biological shape characteristics of the finger vein is extracted.By adjusting the sliding range of the vertices,the model is better applied to finger ROI detection at various angles.The experiment demonstrates that,compared with horizontal detection frame and rotating rectangular detection frame,the IOU and accuracy of GVMVN are obviously improved.3.The method of multi-finger vein recognition based on voting and fusion were put forward,and the network of multi-finger vein feature extraction and recognition is built.The voting-based identification method can also meet the application requirements of high pass rate and high security by changing the threshold setting of the number of fingers.The fusion-based method synthesizes multi-finger vein ROI images into threechannel images,so that the three-channel images can be directly used for feature extraction and recognition by using the single-finger vein recognition model with excellent performance.Experiment finds out that compared with single-finger vein recognition,the multi-finger vein recognition is more flexible and the accuracy is greatly improved.4.This thesis proposes a feature extraction network based on dual attention mechanism,which enhances the model’s ability to select vein features and spatial regions by weight distribution to channels and spaces.In addition,it also can effectively improve the model’s recognition effect on low-quality images.The experimental result displays that for the problem that some or all of the vein features of a finger are missing in the image,the network can improve the model recognition performance well because of the dual attention mechanism.5.A relatively large virtual data set BGMVD is generated by using the biodynamic algorithm of finger vein,thus improving the current situation that the number of public data sets of finger vein is insufficient.Experiment illustrates that the BGMVD has a larger inter-class distance and a smaller intra-class distance,and is similar to the real finger vein data set.Eventually,a large-scale 1:N identification experiment of multi-finger vein model is carried out using BGMVD,which verifies the prospect of multifinger vein identification in practical application.