Image Denoising Algorithm Based On Sparse Representation

Images in the process of acquisition and transmission can produce noises inevitably,due to the impact of equipments and the surrounding environment,which will seriously affect the visual effect and the subsequent processing of the image.Therefore,in recent decades,image denoising has been one of the most important research topics in the image processing.Denoising algorithm based on non local similarity and sparse representation is a very important denoising method in recent years.The non-local mean algorithm considers the self-similarity of the image's internal structure and uses the redundant information to estimate the true value pixel of the image.The core idea of sparse representation model is to transform the input signal sparse representation in the transform domain,and set the threshold to sparse filtering in order to achieve the needs of denoising.In this paper,we improve the nonlocally centralized sparse representation denoising algorithm(NCSR)and weighted encoding with sparse non-local regularization algorithm(WESNR)on non local self-similarity and sparse representation denoising model.The main improvements include the following aspects:Firstly,we analyze the non local mean algorithm and sparse representation model which are widely used in image denoising in recent years.It is found that although the nonlocally centralized sparse representation model(NCSR)has good performance in non sparse noise(such as Gauss noise)processing,the denoising effect of sparse noise such as salt&pepper noise,periodic noise and the mixed noise composition of Gaussian,salt&pepper,periodic is not effective.Therefore,in this paper we propose a modified nonlocally centralized sparse representation algorithm(MNCSR).The new algorithm consists of two stages:first,the adaptive median filter is used to remove the sparsity noise,and then use the NCSR algorithm to remove the residual noise.The experimental results show that the proposed MNCSR is more effective than the original NCSR algorithm,especially for the mixed noise.Secondly,we have do lots of experiments around the mixed noise removal problem.We found that the weighted encoding with sparse nonlocal regularization algorithm(WESNR),which combines image sparsity and nonlocal similarity.The algorithm has a good performance in removing low-intensity mixed noise,but it is not ideal for high-intensity mixed noise.In this paper,based on the analysis of the theoretical basis of WESNR algorithm,using the idea of elastic net by adding l2 regular term,we propose the elastic network and weighted encoding with sparse non-local regularization algorithm(CN-WESNR).The experimental results show that the improved algorithm can make full use of the prior knowledge of sparsity and non-local similarity,so it has a better denoising ability for the high-intensity mixed noise.