Research On Image Denoising Based On Gaussian Mixture Model And Low Rank Model

Images are an important source of information and knowledge.However,due to imaging equipment and human factors,image degradation is inevitable.It often brings a lot of trouble to the subsequent image processing.Therefore,image restoration technology has been one of the hot research objects.Image restoration technology development up to now,from the spatial domain average filtering and median filtering method,to transform domain Fourier transform,wavelet transform and multi-scale transform method,as well as the rapid development in recent years the regularized image restoration method and based on sparse representation and image restoration model of low rank,image restoration is a very important content in the computer image processing.Based on the study of the classical image restoration algorithm and the analysis and summary of the main existing image denoising technologies,this paper proposes the improvement of two kinds of image denoising models aiming at the shortcomings of existing image denoising models: one is the EPLL image denoising algorithm based on the asymmetric gaussian mixture model;The other is a weighted kernel norm minimization image denoising algorithm that fuses relative total variation.(1)In nature,most images are complex.The gaussian distribution is a symmetric distribution,which is usually not fit into an image.Aiming at some asymmetric data in images,which are often not well represented by the gaussian mixture model,we propose an EPLL image denoising algorithm based on asymmetric gaussian mixture model,which improves the finite mixture model's assumption that all components of the natural image conform to the characteristics of gaussian mixture.The asymmetric gaussian mixture model can simulate the asymmetric distribution and is more consistent with the natural image data,so the recovery effect is improved.(2)In the existing image denoising model based on low-rank representation,the weighted kernel norm minimization algorithm has the characteristics of easy solution and good performance.It uses different weights to distinguish the singular value,uses the non-local similarity of graph structure,and obtains the similar block matrix by block matching.However,the texture part and edge structure of the irregular image are still not well expressed,and it is easy to produce smoothing phenomenon in the detail texture part of the image.So in order to solve this problem,we proposed a weighted nuclear fusion of relative total variation norm minimized image denoising algorithm.This method improves the method of weighted kernel norm minimization by combining the advantages of relative total variation in removing noise and keeping edge structure and texture features,and obtains good results.