Image Denoising Algorithm Based On Non-local Self-similarity

In today's Internet era,images always convey important information and play an extremely important role in the scientific research,education,and military fields of human society.Especially with the arrival of 5G technology,this effect is even more obvious.However,in the process of image acquisition and transmission,noise will be introduced due to external or human factors,which will seriously affect the quality of the image,cause the lack of useful information,and affect all aspects of society.Therefore,the removal of noise from images is a very meaningful and enduring topic.The research on image denoising technology has begun since the last century.Experts and scholars have proposed many valuable image denoising algorithms.With the introduction of the non-local self-similarity theory in 2005,the research on image denoising has entered a new era.Based on non-local self-similarity and combining low-rank and sparse representation theory,this paper studies and improves the weighted nuclear norm minimization(WNNM)method and nonlocally centralized sparse representation(NCSR)algorithm.The main contents and innovations are as follows:(1)The low-rank theory and the popular denoising algorithm based on low-rank in recent years are introduced,and the WNNM algorithm is mainly analyzed.On the basis of clarifying that the WNNM algorithm has a strong removal effect on non-sparse noise such as Gaussian noise,and it is not strong or even poor in removing sparse noise such as salt and pepper noise and periodic noise,an improved algorithm combining the advantages of WNNM algorithm and adaptive median filtering algorithm is proposed,that is,first use WNNM algorithm to effectively remove non-sparse noise such as Gaussian noise,and then use adaptive median filter algorithm to remove residual sparse noise such as salt and pepper noise and periodic noise.Experiments show that the improved algorithm has good effects both on subjective vision and image quality evaluation indicators.(2)The theory of sparse representation and the popular denoising algorithms based on sparse representation in recent years are introduced,and the NCSR algorithm is analyzed emphatically.Aiming at the fact that k-means clustering using Euclidean distance of noise image as the measure in the learning dictionary stage of NCSR algorithm will affect the similarity of image patches and even the quality of the dictionary,the gaussian mixture model based on priori knowledge of external clean image is proposed to guide the inner clustering of image,which effectively avoid the impact of noise;Aiming at the fact that the Euclidean distance is used in the sparse coefficient estimation stage of NCSR algorithm,but the brightness,contrast and structural similarity information is ignored,the structural similarity index measurement is introduced into the Euclidean distance formula,and the improved Euclidean distance formula is introduced into the calculation of the weighting coefficient.Experiments show that the improved algorithm has higher peak signal-to-noise ratio and better visual effect.