Research On Image Denoising Algorithm Based On Non-Local Mean

Image denoising is one of the most important and basic problems in the field of image processing.The purpose of image denoising is to eliminate the noise generated in the process of image transmission and recording.The results of image denoising directly affect the subsequent processing operations,such as image segmentation,edge detection,target recognition and so on.Image denoising algorithms usually assume that the image satisfies certain properties.Non-Local Means(NLM)algorithm is based on the self-similarity of the image.This algorithm makes full use of redundant information in the image,searches similar image blocks in the whole image and weighted average to restore the central pixel.This algorithm achieves good denoising effect when it is proposed,and is superior to most denoising algorithms.Block-matching and 3D Filtering(BM3D)takes NLM algorithm as the basic framework and introduces the concept of collaborative filtering.BM3 D is considered to be the best denoising algorithm in the field of image denoising.In this paper,the NLM and BM3 D algorithms are deeply studied.The NLM algorithm based on local Fourier transform and Laplacian and the BM3 D algorithm based on adaptive threshold are proposed.The research work of this paper mainly includes the following aspects:For the problem that the Gauss function in the NLM algorithm does not take into account the influence of noise on the Euclidean distance weighting,this paper analyzes the influence of noise on Gauss weighting,and proposes the Laplacian function combined with gradient information to improve the original Gauss function.The weight is used to measure the noise level in the image after denoising,and measure the influence of the neighborhood of pixels on the weight before the denoising,and gradient information is used to distinguish edges from noise.For NLM algorithm which only uses Euclidean distance to measure the similarity of image blocks can not consider the structural information in image blocks,this paper proposes a method to measure the similarity of image blocks based on local Fourier transform features of principal component analysis.Eight local Fourier transform templates are used to extract the local Fourier transform features of image blocks,and principal component analysis is used to reduce the dimension of features.The features obtained by this method can take into account the neglected structural information.To solve the problem that BM3 D can't get a proper number of image blocks because of using fixed threshold in block matching stage,this paper introduces a gradient-based structural similarity measurement method to analyze the number of blocks and matching threshold needed in block matching process,and the mathematical model of adaptive block matching threshold is constructed,which can calculate the optimal matching threshold for each image block.For the problem that the hard threshold processing of BM3 D in three-dimensional filtering will neglect the middle and low frequency parts of real signals,this paper proposes a soft threshold processing method in transform domain to preserve the middle and low frequency parts of original signals.A large number of experiments show that the two improved algorithms proposed in this paper are superior to the original ones.They are improved in peak signal-to-noise ratio and visual effect,and are superior to other excellent denoising algorithms in the current stage.The experimental results show that the improved NLM algorithm has a strong edge recovery ability,but a weak ability to restore texture areas,which is suitable for images with more edges.The improved BM3 D algorithm has a strong ability to restore texture regions,but it has a general ability to restore edges,which is suitable for images with more texture regions.