Patch-based Image Denoising

Image denoising is a basic topic in the field of image processing and computer vision.The aim of image denoising is to remove noise while preserving image details as much as possible.As a specific class of image restoration,image denoising is an ill-posed inverse problem.Model-based optimization methods and discriminative learning methods have been the two dominant strategies for solving various inverse problems.In numerous image denoising methods adopting model-based optimization,the non-local mean method is a novel denoising strategy.For the existing drawbacks of original NLM algorithm,This article proposes three improved methods:the first one utilizes the superpixel segmentation algorithm to cluster pixels with similar local features and assigns higher weights to pixels with the same label in the process of estimating center pixel;the second one devises a parameterless edge detection method to generate the image edge distribution map,and then dynamically adjust the value of the filter parameter h according to the ratio of edge points within the search window so that the parameter h is better adapted to the local characteristics of the image;the last one is aimed at reducing the effect of noise on measuring similarity between image patches.using the orthogonal unit basis of pca,it generates the main patterns of the image patches and replaces original patches with their main patterns to compute the similarity.In addition,a new weight function is introduced to eliminate the contribution of dissimilar patches to the result.The experiments show the three improved methods have better denoising performance than the original NLM algorithm.In recent years,image denoising based on discriminant model learning has attracted much attention due to its good performance in removing certain levels of noise.In this article,a convolutional neural network is designed to learn noise from a noisy image and implicitly achieve the goal of denoising.The network model uses many advanced techniques and methods,including relu,batch normalization,and residual learning,to speed up the training process and improve the denoising performance.Through the training of a large number of sample data,the network model haves the more favorable denoising performance comparaed to the state-of-the-art EPLL and BM3 D algorithms in the aspect of removing specific levels of noise.