Design Of Finger Vein Image Acquisition And Recognition System

Compared with traditional personal identity verification,biometric identification technology has higher security,and has attracted wide attention in recent years.Among the existing biometric identification technologies,finger vein image identification technology is particularly outstanding in the aspects of security and stability,so it has been rapidly developed.Because of its strong learning ability and recognition and generalization ability,convolutional neural network has gradually replaced the artificial design operator to express the texture information in finger vein images.However,convolutional neural network has a lot of parameter redundancy and a large amount of computation.Considering the computational capacity of finger vein recognition system,the number of parameters and computation amount of algorithm model should be reduced as far as possible while maintaining its recognition accuracy.Below,the specific research content and experimental results of this paper are explained.Firstly,in view of the problem of forged vein information with special materials,a method of in vivo assisted determination based on physiological signals of pulse wave was proposed.Since veins distributed in the fingertip are excluded in venous image acquisition,we collect pulse wave signals from the fingertip,calculate pulse frequency,and compare it with the heart rate range of normal human body to complete the in vivo auxiliary judgment.Secondly,a lightweight structure(named PRCNet)based on structured Pruning and Restoring Channels(PRC)was proposed to solve the problem that convolutional neural network model has many redundant parameters in finger vein image classification.The PRC structure uses the norm of convolution kernel to judge the importance of each part of the convolution kernel,and then tailoring the unimportant convolution kernel structure,which can effectively eliminate redundant calculation.Considering that the convolution kernel will affect the forward propagation of the jump-connected structure through structural clipping,the channel recovery algorithm is added on the basis of pruning,which will not increase the calculation amount significantly and enrich the utilization rate of high-weight features.Experiments show that compared with the base network,the parameter number is reduced by4 times and the computation amount is reduced by 1.3 times by using the PRC optimization structure.Moreover,the equivalent error rates obtained from the finger vein database of Shandong University and Hong Kong Polytechnic University are 0.025 and 0.085 respectively,which are much lower than the equivalent error rates obtained from the Res Net18 basic model(1.17 and 2.13).Thirdly,aiming at the problem of slow feature matching in the recognition phase under the condition of large data volume,this paper proposes a progressive feature vector matching scheme based on hierarchical feature matching.The features in the venous image registration database were first screened by preset intermediate vectors and then presented to the next level for fine matching.Moreover,we also optimized the Loss function according to the distancebased matching scheme to regularize the cross-entropy function with the simplified Center Loss,which made the classification result more compact,strengthened the intra-class distance and weakened the class distance,so as to make the matching more fast and accurate.Experimental results show that the number of features retrieved can be reduced by more than95.57%,and the time of feature matching can be reduced by more than 84.27%,which greatly reduces the calculation pressure of the system.Fourthly,for the design of finger vein image acquisition and recognition system,our system takes Raspberry Pi 4 Model B and Intel Movidius as the main control core and computational power support and uses near-infrared light sets and cameras to collect finger vein images of people to be recognized.The recognition function of the system is realized by using the designed recognition method.