Collaborative Blind Denoising of Similar Patch Group Based on Fully Convolutional Siamese Network

Recently,patch group collaborative denoising algorithm based on image self-similarity has made rapid development,but how to find similar patches quickly and accurately in a noisy environment is a difficult problem.The common block matching algorithm defines the similarity between image patches through Euclidean distance,which could not measure the internal structure information of image patches.A search method for similar patches is proposed based on fully convolutional siamese network.Firstly,the potential relationship between clean reference patch and noisy image patch is learned through siamese network.Then,Mahalanobis distance is used to measure the similarity between image patches by using structural information of image patches.Finally,the image is restored by using similar patches collaborative denoising.Experiments show that the PSNR value of the proposed algorithm is higher than GID by 0.51 dB,1.02 dB and 0.20 dB on Nam-CC15,Nam-CC60 and PolyU real image datasets respectively.By visual comparisons,the proposed algorithm can preserve more structural features than the competing methods.