Research On Finger Vein Image Denoising And Inpainting Algorithm Based On Texture Constraint

The performance of finger vein recognition system largely depends on the quality of the collected finger vein images.Various types of noise generated by the collection equipment or environmental influences,as well as mirror oil stains in actual applications,and molting of the user's fingers,etc,will cause damage to the quality of the finger vein images and affect subsequent recognition performance.In this paper,the classic image denoising algorithms and inpainting algorithms are firstly studied,and it is found that these algorithms do not accurately use the texture information of the images,which easily leads to problems such as blurred veins,loss of details,and structural distortion for finger vein images with weak texture edge information.Therefore,this paper proposes a method for denoising and inpainting finger vein images based on texture constraints.In the process of reducing noise through traditional algorithms and inpainting damaged areas based on deep networks,texture information is used to constrain.At the same time,the detailed information of the vein is maintained.The specific research content is as follows:1.A Gabor Local Binary Pattern(GLBP)texture feature with strong anti-interference and good adaptability to light changes is proposed,which combines the advantages of Gabor and Local Binary Pattern(LBP).On the basis of Gabor extracting 8 directions of finger vein image response values,LBP is used for encoding,and a better characterization of vein image texture direction is obtained,named as GLBP.Gabor filters,with good directivity,which can better match the characteristics of the different extension directions of the finger veins,are used to obtain the 8directions response information of the finger vein images.Noise has a certain impact on the extraction of response values,and LBP has good anti-noise ability,so on the basis of Gabor extraction,LBP is used for encoding to obtain more stable GLBP features.Theoretical analysis and experimental results show that GLBP feature still has a good discrimination ability under the influence of noise and image damage.2.A Non-Local Mean filter denoising algorithm for finger vein images based on GLBP texture constraints is proposed,which adopts a weighted denoising kernel function based on GLBP texture feature constraints and gray distance information correction.The proposed algorithm can optimize the weight distribution of vein background and foreground,and improves the denoising effect of unclear finger vein images.This chapter firstly uses the new cosine function to obtain the GLBP texture similarity between image blocks.Considering that the gray information is affected by noise,but it also contains some information.Therefore,based on the gray information,the Gaussian weighted Euclidean distance is used to modify the kernel function,so that image blocks with high similarity can be assigned high weights,and image blocks with low similarity can be assigned low weights.Reasonable weight distribution has significantly improved the denoising performance of the algorithm.Theoretical analysis and simulation results show that,compared with the traditional Non-Local Mean denoising algorithm without texture constraints,the proposed algorithm can effectively protect the details and edge information of vein while denoising.3.A finger vein image inpainting algorithm based on Neighbor Binary-Wasserstein Generative Adversarial Networks(NB-WGAN)is proposed.The proposed model consists of a cascaded generator network for coarse inpainting and precise inpainting,and two Wasserstein Generative Adversarial Networks with Gradient Penalty(WGAN-GP)that focus on the repair effects of local damaged regions and the repair effects of the entire image respectively.The cascaded generator structure can extract richer and more comprehensive information during the inpainting process to achieve better recovery of the damaged area.The parallel discriminator networks combine local loss and global loss to constrain the network,which ensure that the entire image is consistent with the damaged area after repair.In order to better protect the texture information of finger vein images,a Neighbor Binary Pattern(NBP)texture loss is proposed,which combines reconstruction loss and adversarial loss to jointly constrain the update of network parameters.Finally,through a detailed analysis of the characteristics of the damaged finger vein images,a training set containing multiple damage conditions is proposed,which enhances the generalization ability of the network model.Theoretical analysis and experimental simulation show that,compared with the traditional Curvature Driven Diffusion(CDD)algorithm without texture constraints,an inpainting algorithm based on Gabor texture constraints,and an depth inpainting algorithm based on attention mechanism without texture constraints,the NB-WGAN algorithm proposed in this chapter has better inpainting performance.