基于全卷积孪生网络的相似块组协同盲去噪算法

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

近年来,基于图像自相似性的块组协同去噪算法取得了快速发展,但是如何在噪声环境下快速精确地寻找结构相似图像块是一个难点.常用的块匹配算法是通过欧几里得距离定义图像块之间的相似程度,无法度量图像块内部的结构信息.针对这类问题,提出一种基于全卷积孪生网络的相似块组搜索算法.首先通过孪生网络学习干净参考块与噪声图像块的潜在联系;然后利用马氏距离结合图像块的结构信息度量其相似性;进而对相似块组进行协同去噪恢复图像.实验表明,相比于GID算法,所提算法的峰值信噪比值在Nam-CC15,Nam-CC60和PolyU真实图像数据集上分别提高了0.51 dB,1.02 dB和0.20 dB;视觉效果上,所提算法能够使去噪图像保留更多的结构特征.

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.