Research On Finger Dual-modal Recognition Method Based On Convolutional Neural Network

Multimodal recognition technology has the advantages of high anti-counterfeiting and stability by fusing a variety of feature information for identity verification.Finger veins and knuckles are easy to obtain,stable and difficult to forge,which has become an important research object of multi-modal recognition technology.However,there are some problems that need to be further discussed in the process of multimodal biometric research.The shift of finger posture,illumination and wavelength of acquisition spectrum can easily lead to high intra-class difference and low inter-class difference,which leads to the difficulty of feature extraction,and the feature dimension and computational complexity are increased in the process of multimodal fusion.How to effectively extract image feature information and realize feature fusion has become a key problem in the field of multimodal feature recognition.Convolutional neural network(Convolutional neural Network,CNN)can abstract the high-level semantic information of images autonomously and has attracted wide attention in the field of multimodal biometric recognition.Therefore,this paper takes finger vein image and phalangeal pattern image as the object and studies a method of finger bimodal feature recognition based on CNN.The main contents are as follows:(1)The effect of image quality on feature extraction is analyzed.Aiming at the problem of the low contrast of finger vein image,a simplified NL module is introduced to model the global context information of vein feature map,to preserve the global information of vein image to the maximum extent,and to further design NL-Res Net extract finger vein feature.Aiming at the phenomenon of finger joint image migration,the multi-resolution image input strategy is adopted,and a two-stream network is built to realize the feature extraction of finger knuckles print.This method can make up for the problem of finger knuckle print image offset to a certain extent.Experimental results show that the two schemes can effectively improve the performance of a single-mode feature recognition system.(2)To take into account,the correlation between finger vein and phalangeal pattern features,and enhance the diversity of fusion features,a multi-scale feature fusion module is proposed MSIF,which is used to fuse the multi-scale features of each layer of the network to represent the fusion features comprehensively.To improve the discrimination ability of finger vein and phalanx fusion features,the multi-modal feature recognition based on CNN can obtain better results in finger biometric recognition.(3)In view of the difficulty of network fitting caused by the increasing complexity of feature information after multi feature fusion,an attention convolution binary tree network classification structure is proposed.The tree structure is combined with location attention mechanism and channel attention mechanism to emphasize the important spatial and channel information and improve the efficiency of network fitting.The experiments were carried out on the MMCBNU?6000?SDUMLA-HIT?Poly UFV three finger vein databases and the Poly U-FKP finger knuckle print database,which proved the effectiveness of the single-mode feature extraction method proposed in this paper.At the same time,based on the above four single-mode databases,the dual-mode database is recombined,and the comparative experiments show that the proposed feature fusion method is better than some common methods.