| With the rapid development of the information industry and the wide application of the Internet,information security has become more and more important.As a new identification technology,vein recognition is not easy to copy,forge,or destroy compared with fingerprints and faces.It is irreplaceable in high-security applications and has broad application prospects.However,the imaging quality of vein features is affected by various factors,such as ambient light,ambient temperature,user behavior,etc.Finger vein identification technology is facing severe challenges.For example,during the recognition process,the flipping of the user’s finger will change the projection angle of the three-dimensional blood vessels on the two-dimensional plane,resulting in ambiguity of vein features,that is,the same finger veins may have different two-dimensional plane projections,which reduces the system recognition accuracy.At present,deep learning models such as Transformer have shown good recognition performance in computer vision tasks.Therefore,this paper studies the Transformer model and applies it to finger vein recognition,and carries out the following work:(1)A researchable multi-view finger vein database is established.First,a camera is used to rotate around the finger to collect projection images of finger veins from all viewing angles.Then the collected finger vein images are preprocessed,and the U-net network is used to extract the region of interest(ROI)of the image.Finally,the ROI image is normalized to establish a multi-view finger vein database.(2)A single-view finger vein recognition algorithm based on an improved attention mechanism is proposed.According to the characteristics of finger vein images,the selfattention mechanism is improved,and a pixel attention module that can pay attention to the details of vein images is established.By stacking pixel attention modules,a Transformer model based on an improved attention mechanism is constructed.The model simultaneously extracts inter-block and intra-block correlation features of finger vein images,realizes simultaneous attention to coarse-grained and fine-grained single-view finger vein images,and improves the feature expression ability of the model.The experimental results on the public finger vein datasets FV-USM,Ploy U-FV and the self-built dataset CTBU-SFV show that compared with the existing Transformer model,the Transformer model based on the improved attention module effectively improves Recognition accuracy,while reducing the computational complexity of the model.(3)A multi-view finger vein recognition algorithm based on local attention Transformer is proposed.The method consists of two Transformer modules,one module is responsible for extracting the dependent features between patches in the vein image,and the other module is responsible for learning the dependent features between images under different viewing angles.In addition,a graph structure is introduced into the model using a masked self-attention mechanism to extract local attention features between adjacent nodes.To further improve the performance of model recognition,conditional position encodings are used to dynamically generate position encodings.The experimental results on the multi-view finger vein database in this paper show that the algorithm in this paper is superior to the existing multi-view vein recognition methods,and the recognition accuracy reaches 97.81% in the case of finger flipping. |
