Deep Learning Based Palm Vein Recognition Method And Application

With the continuous development of society and the progress of Internet technology,people have higher and higher requirements for information security,and how to solve the problem of personal identification becomes very important.Biometric identification has the characteristics of accuracy,security and convenience,and has great advantages compared with traditional identification methods.But the existing biometric identification technology can not meet people’s requirements,palm vein recognition technology is an emerging biometric identification technology,its principle is to use near infrared camera to irradiate the palm to get the vein characteristics for identity verification,compared with other biometric identification its more difficult to be forged,and also more easily accepted by users.However,there are some technical difficulties in palm vein recognition,mainly including the localization of the region of interest and the performance and lightweight of the vein feature extraction network model,which will be studied in this paper.In order to reduce the invalid information and interference factors of palm vein image to improve the recognition accuracy and efficiency,firstly,the region of interest localization of palm vein image is needed,however,due to the interference of external environment such as light,position and angle during palm vein image acquisition,the traditional region of interest localization method may be wrong,in order to improve the robustness of palm vein region of interest localization method,this study proposes a target detection based network model.In order to improve the robustness of the palm vein region of interest localization method,this study proposes a target detection-based region of interest localization method to solve the existing problem of palm vein region of interest localization failure.Since the target detection algorithm YOLOv3 network model and the large number of parameters are not suitable for deployment to embedded devices,this paper uses the lightweight network model Mobile Net V2 to replace the YOLOv3 backbone network Darknet-53.In addition,this paper uses the GIo U loss function to optimize the region of interest localization problem to further improve the detection effect,and finally,the detection effect on the Poly U palm vein dataset was tested,and the accuracy of palm vein region of interest localization reached99.5%.In order to reduce the loss of important vein features due to downsampling,this paper proposes a new feature fusion module FFD to replace the Bottleneck module of Mobile Net V2,and at the same time adds the SENet module to Bottleneck.The SENet module is added to increase the weight of important features.In addition,this paper uses Focal Loss instead of the traditional cross-entropy loss function,and assigns different weights to each sample according to the sample difficulty to further improve the model performance.In this paper,we choose the method of dataset augmentation to overcome the problem of small samples of palm vein data,and finally,we test on Poly U palm vein dataset,and the iso-error rate is 0.58%.In addition,this paper deploys the network model into an embedded computing device to implement the user registration and authentication functions,and tests the recognition time of the network model,which is about 200 ms,proving the feasibility of the proposed method in this paper.