Research On Dual-modal Biometric Recognition Method Based On Deep Learnin

With the increasing demand for information security and the smart devices,biometric identification systems using single-modal data have limitations such as nonuniversality,intra-class variation,high false recognition rate,and spoofing attacks,etc.Multimodal biometric identification aims to make up for the shortcomings of singlemodal systems in security and accuracy,and has become a research hotspot in the field of personal identity authentication.However,the inconsistency in the distribution of different modal images leads to heterogeneity gaps,which makes it very challenging to represent and associate these heterogeneous biometric data.Therefore,in this paper,we propose a multimodal biometric recognition method based on deep learning using finger vein images and fingerprint images,and the main work is summarized as follows:(1)Due to the sensitivity of biometric data,real multimodal data with labels are scarce,and there are inconsistency in the number of samples between modalities.To construct a multimodal finger dataset with a balanced number of samples,this paper proposes a finger vein image generation method based on a category-based two-way diffusion model.First,based on the diffusion model,Gaussian noise is added to the given image;then,the deep semantic information in the low-contrast image is effectively learned by a two-way deep feature extraction block;finally,a category conditional branch is constructed to improve the identity representation of the generated image.From the image quality evaluation and recognition performance analysis,the method can accurately replicate the intra-class common features of finger veins and effectively solve the problem of sample scarcity and imbalance in multimodal biometric recognition.(2)To address the problem that finger biometric images contain features with different scales and different distribution characteristics of biometric images of different modalities,this paper proposes a finger feature extraction method based on axiallyenhanced neighborhood attention.On the one hand,the axially-enhanced neighborhood attention infer the query domain range of the attention module based on the distribution information of a given image and dynamically aggregate long and short distance information;on the other hand,the feature expansion is performed by using group convolution,which improves the similarity and accuracy of features.From the results of ablation and comparison experiments,the method can capture the representation more accurately and flexibly,which is a good trade-off between accuracy and parameters.(3)In response to the existing multimodal biometric recognition research dedicated to establishing correlations between modalities,which ignores the performance degradation caused by redundancy,this paper proposes a feature fusion method based on similarity-aware coding networks.First,a cross-modal excitation module is designed to learn the interdependencies;then,the encoder is constructed for nonlinear joint modeling;finally,the reconstructed representation generated by the decoder and the input are discriminated using the cosine dissimilarity loss,aiming to make the generated representation closer to the original input.From the validity analysis and ablation experiments,the method has higher accuracy compared with unimodal recognition,and effectively mitigates the performance degradation caused by information redundancy.(4)To address the problem that fusion methods based on individual feature spaces often need to select different optimization targets in practical applications,and there are bottlenecks in the performance improvement of recognition systems when the data types and morphologies are complex,this paper proposes a feature fusion method based on a multicore coding network.Firstly,multiple encoders with the same architecture are constructed for dimensionality reduction and fusion of multimodal features;secondly,the cross-modal components of heterogeneous features are learned separately according to a more appropriate individual loss function for representation;finally,learnable weighting coefficients are set for aggregation of mapping subspaces.From the results of ablation and comparison experiments,the method further improves the accuracy and security of multimodal biometric recognition,and provides the possibility of combining more features with different subspaces of mapping ability.