| Existing finger-vein recognition technology has already achieved great achievements in theoretical research,but there are still many technical challenges in practical application.For example,large-scale deep neural networks are difficult to directly deploy on embedded devices,and low-quality images seriously affect the system’s recognition accuracy.In response to these two major technical issues,this thesis carries out the following research:(1)Proposed a parameter-dependent lightweight feature extraction network algorithm.The technical approach mainly includes two aspects: firstly,selecting lightweight classic network structures-Mobile Net and Squeeze Net as the network backbone to reduce the network size from the overall structure.Secondly,using parameter-dependent convolutional kernels(PDKs)to further reduce the parameters of the backbone network,thereby achieving the ultimate compression of the model.The experimental results show that the proposed PDKs can compress the parameters of Mobile Net and Squeeze Net by 36.47% and 96.17%,respectively,while improving the performance of them: on the FV＿USM,MMCBNU＿6000 and SDUMLA datasets,the Equal Error Rate(EER)of Mobile Net were reduced by 0.617 %,0.05 % and 0.15 %respectively,and the EER of Squeeze Net were reduced by 2.34 %,4.42 % and 2.21 %respectively.(2)Proposed a finger-vein image quality assessment algorithm based on similarity distribution.The implementation of the algorithm is divided into two steps: firstly,automatically label the quality of the image based on the intra-class and inter-class similarity distribution of the feature vectors of the finger-vein images.Secondly,to explore the common properties of low-quality images,a lightweight image quality assessment model is constructed.To address the problem of imbalanced quantities of high and low-quality images,a low-quality image generation algorithm is proposed.The experimental results show that after using this algorithm to remove low-quality images from the FV＿USM,MMCBNU＿6000 and SDUMLA datasets,the performance of various finger-vein recognition systems in the comparative experiments has been improved to varying degrees.Specifically,on top of the EER results mentioned earlier,Squeeze Net further reduced the EER by 0.51%,0.8%,and 1.07%.(3)Developed a prototype of finger-vein recognition system that is suitable for practical application scenarios.The system integrates the lightweight feature extraction model and image quality assessment model proposed in this thesis,providing services such as finger-vein registration,recognition,and verification for ordinary users,and also providing related services such as model management and system performance statistical analysis for system administrators.The system is lightweight,occupying only 108 M of memory space.The test results show that the average response time of this system is298 ms,and its recognition accuracy is as high as 99.17%.It effectively simulates the vein recognition process in practical application scenarios and has certain practical value. |
