Study Of Finger Vein Recognition Based On Deep Learning

With the continuous integration of information technology and big data technology into our lives,biometric technology is also developing vigorously.In order to protect citizens' personal privacy and information security,more and more biometric technologies appear in our society.Finger vein recognition technology,a relatively novel biometric technology,stands out because of its low repetition rate,difficult to forge,stable features and the need for live detection,and has made great progress.Finger vein recognition technology generally uses an infrared camera to form a finger vein image according to the principle of finger blood absorbing infrared light,which is used as a feature for identity verification.In order to solve the problem that the traditional finger vein recognition method is usually easily affected by imaging quality,change of finger gesture or lighting conditions,the finger vein recognition method which is based on deep learning is studied in this paper.The main work and contributions are as follows:Firstly,this paper proposes a finger vein ROI extraction network which is fine-tuned by pre-training network to solve the problem of poor robustness and mobility of traditional ROI extraction methods.At the same time,by analyzing the factors that have the greatest impact on the accuracy of finger vein image recognition,this paper aims to expand the size of the training sample data set and improve the robustness of the model through data amplification.Secondly,the finer vein ROI extraction network proposed in this paper is used as a region proposal network,so as to construct a finger vein feature extraction convolution neural network(FV R-CNN).In addition,in order to adapt to the zero-shot learning task of finger vein recognition,this paper uses the metric learning method to learn the semantics of finger vein images,and proposes an adaptive region of interest loss function(Adaptive ROI Loss,ARL)to train the network at the same time.Through metric learning,the network can extract semantic expressions with stronger represent ability and stronger distinguishing ability.Because the ROI of the finger vein which is suitable for the traditional method may not be the best region for the convolution neural network,and there is even the possibility of mislabeling,this paper proposes an adaptive ROI loss function,which enables the network to learn the optimal ROI automatically.This paper uses an end-to-end multi-stage training method combined with online hard sample mining technology.For each finger vein image,the network finally outputs 128-dimensional feature vectors and matches them using cosine distance.Thirdly,because the model of the original feature extraction network of finger vein is heavy and the computational cost is high,so it is difficult to deploy on the mobile terminal or embedded devices,this paper uses a variety of model compression method to optimize and prune the network.In the training stage,the pruned network structure is generated and fine-tuned.The pruning method can repeatedly obtain the best model.After the completion of network pruning,the method of network quantization is used to further reduce the size of the model.The network of finger vein feature extraction proposed in this paper is 10.2MB and lightweight is 3.35MB.The author carried out several groups of comparative experiments on Shandong University data set,Tsinghua University data set and Chonbuk National University data set respectively.The original model achieved equal error rates of 0.48%,0.39%and 0.21%on the three data sets,while the equal error rates of the lightweight model on the three data sets were 0.62%,0.51%and 0.34%respectively.The experimental results verified the effectiveness of the proposed method.