Research On Image Denoising Algorithm Based On Non-local Mean

In the field of image processing,the block matching method is a common method in image denoising and motion estimation.It finds similar blocks by querying a specified area,compares information and performs joint denoising,so as to weaken the noise that does not conform to the law of the picture.Among them,the three-dimensional block matching algorithm(BM3D)algorithm combines several algorithms in the space domain and the frequency domain,and is one of the best algorithms for image denoising.This paper studies the image denoising algorithm based on the non-local mean algorithm.The main research work is as follows:First of all,it takes a long time to calculate the three-dimensional block matching algorithm,and solves the problems that the extraction of similar blocks is not accurate enough,and similar features with different distributions in the divided blocks cannot be extracted.It is proposed to reduce the dimensionality of image information based on the Radon transform,and to determine the similar block information by solving the variance value of the dimensionality reduction function and the coordinate positioning of the maximum variance.In addition,the pixel average ratio of each similar block relative to the reference block is calculated,and then used for denoising after scaling.The experimental results show that the optimized algorithm denoises the image compared with the original algorithm to reduce the running time of the denoised image,and the image quality is improved.Then,for the features contained in the image with less regularity and strong randomness,there is no way to take advantage of denoising.This paper takes advantage of the non-local structural self-similarity in natural images and extends the BM3 D model,allowing the detection of self-similarity at different rotation angles to obtain a sparser representation and better signal and noise separation.Since the matching of different rotations requires interpolation,some blur will appear in the block matching process,and the rotation matching may cause unfavorable effects in some image regions.To this end,this paper also introduces a low-rank regularization process and a "hybrid" method to adaptively determine the weights to combine different denoising estimates.The resulting method improves the performance of BM3 D.Finally,this paper combines the idea of nonlocal sparse modeling with deep learning in the emerging field,and proposes a denoising method based on deep learning(DND).This method uses the densenet structure and a new nonlocal layer to achieve the best effect of image denoising.Through training and testing on MSCOCO data set,the experimental results show that this method is superior to other methods in terms of visual quality and peak signal-to-noise ratio.This paper also proposes a new non local connection structure,which enables the network to perform non local collaborative filtering effectively when denoising the image.By adding a non local layer,the performance of this method is further improved,and the availability of the non local algorithm in deep learning is proved.