Research On Non-local Self-similarity Based Denoising Algorithm

Image denoising,as a basic image restoration problem,has always been a research hotspot.In today's information age,images as an important method of information transmission are closely related to people's life.All applications of images can only meet the requirements of production and living under the premise of high-quality images.However,in the process of images capturing,storing and transmitting,noise can be generated due to interference from external factors thus leading to low-quality images.Therefore,research of image denoising is particularly important.Since the 1980 s,researchers have proposed many algorithms for image denoising.The proposed of non-local self-similarity theory makes research of image denoising into a new era.Afterwards,there are many image denoising algorithms based on non-local self-similarity priori proposed.The weighted nuclear norm minimization method is one of them,as with other low-rank matrix recovery methods,this method use the low-rank property of the matrix which composed of non-local self similar image blocks to acquire denoised images.However,there is a pretty obvious problem in this method,when apply non-local self-similarity to find the similar block matrix of a given image block,the Euclidean distance between noise image blocks is directly used,but the noise will affect the accuracy of similarity calculation thus influence the denoising effect.Based on the weighted nuclear norm minimization method,this dissertation focuses on the problem that noise influences similar image blocks' calculation.The main contents and innovations of this article include the following points:(1)It proposes to use the preprocessed image instead of the original noise image to find the nonlocal self-similarity matrix of the image block.As the signal-to-noise ratio rises,the preprocessed image solves the problem that this method is not robust to noise when finding similar blocks of the image.In addition,when reconstructing a denoised image block,instead of simply averaging the pixels to obtain each pixel value of the image block,a weighted average method is used.Experiment results show that this method can improve the denoising effect of weighted nuclear norm minimization method.(2)Study and analyze the properties of multi-scale image blocks.It proposed to combine multi-scale images blocks which has the property of scale invariance and noise level declining in a coarser scale to calculate the similarity between image blocks.Based on the texture of the image block,the method uses different directions' pyramid to process the image block.Besides,considering the different amounts of information contained in the image block at different scales,when the similarity between image blocks is calculated,the proportion of image blocks at different scales is different.The experiment results verify the feasibility and effectiveness of the method.