Research And Application Of Finger Vein Recognition Based On Deep Metric Learning

Finger vein is a stable and safe biological feature,which has a wide range of application prospects in life.At present,there are many challenges in the application of finger vein recognition technology,such as: 1.The traditional finger vein feature extraction algorithm is based on manually set feature parameters（geometry and topology of the vein pattern,local binary code,etc.）,and the generalization ability of the algorithm is not good.the deep learning-based method does not make full use of the intra-class and inter-class relationships of the samples when learning the feature representation,and there is a certain room for improvement in the recognition performance of the algorithm.2.Finger vein images collected based on infrared sensors are prone to quality problems such as low contrast,overexposure,blur,etc.Therefore,it is necessary to evaluate the quality of finger vein images.The finger vein image quality assessment is generally based on human visual perception,and there is no connection between image quality and computer recognition.In response to the above problems,this paper has done the following work:（1）Finger vein feature extraction based on deep metric learning.First,in order to ensure the consistency of the data distribution and remove the background noise of the image,the region of interest of the finger image is extracted based on the Sobel operator；secondly,in order to fully express the deep features of the finger vein,a finger image is constructed based on the Res Net network structure and attention mechanism.Vein Feature Extraction Network（FV-FE-Net）； Then,in order to learn more discriminative features,a joint loss is proposed,which exploits the joint constraints of Euclidean space and cosine space to mine the intra-and inter-class relationships of sample features,And the improvement is proposed from the update direction of the feature.The method was tested on the finger vein datasets FV-USM,HKPU-FV,SDUMLA-HMT,MMCBNU＿6000,THUFVFDT3 and FV＿1030,and achieved 1.53%,1.06%,0.86%,0.78 respectively.%,1.37%and 1.83% equal error rates,and compared with the existing methods,the results show that the features extracted by this method have better performance improvement for finger vein recognition.（2）Finger vein image quality assessment based on metric learning.First,in order to avoid the interference of human subjective factors on the image quality score,the average similarity is used to connect the image quality with the computer recognition results.Secondly,in order to solve the problem of the small number of low-quality images,a distortion algorithm is proposed to synthesize low-quality images.Then,in order to make full use of the relative relationship between samples,a Siamese network that learns the contrastive relationship is proposed.Experiments are carried out on FV-USM,SDUMLA-HMT and MMCBNU＿6000,and the equal error rate on high-quality images is about 0.2% lower than the original data set.The results show that the proposed algorithm can effectively improve the performance of finger vein recognition.（3）Finger vein authentication system.Aiming at the increasing demand of finger categories and the problem of insufficient performance of terminal equipment,based on the feature extraction and quality assessment method proposed in this paper,a finger vein identity authentication system is designed and implemented.The system integrates the functions of image quality assessment,feature extraction,1:1 verification and 1:N recognition.