Research On Image Restoration Based On Structure Prior Information

In the context of today's information age,image is an important information transmission medium,which promotes the continuous development of image restoration technology.From the perspective of image internal similarity,the most representative one is the non-local mean method.This type of method considers that many pixels in the image are similar,so that multiple sets of similar patches can be extracted from one image,and then each pixel is weighted and summed to obtain the final result.Starting from the external similarity of images,algorithms such as expectated patch log-likelihood(EPLL),sparse representation,and low-rank approximation extract the statistical or structural features of images through a large number of similar patches,and make use of the redundancy of image information to achieve the purpose of denoising.Based on the prior information of image structure,new image denoising algorithms are proposed by combining different algorithm theories,and satisfactory results are obtained.The main works are as follows:(1)The EPLL algorithm is slightly inadequate in processing complex texture parts of the image.Coupling the low-rank prior can effectively solve this problem.According to the self-similarity of image,we obtain multiple sets of similar patches by the image patch extraction operator.Each group of similar patches is then vectorized into a low-rank matrix with high correlation between rows or columns.Adding this structural prior information to the prior learning processing step can better save the complex texture information of the image.The experimental results show that our method has not only better numerical results,but also better visual effects than the original EPLL and some other classical algorithms.(2)The traditional low-rank methods only consider the internal similarity of image when performing low-rank approximation on the similar patch matrix,which often limits its performance in denoising tasks.To this end,we first perform an initial sparse coding process on the image.This not only removes part of noise,but also makes the coefficient matrix of similar patch matrix under the sparse transformation have a better quality of low-rank,which can be seen as structured sparsity.A low-rank approximation of such a coefficient matrix canbe combined with the updated sparse transform to restore the image.Experiments show that the new method has a higher peak signal-to-noise ratio(PSNR),while retaining more image texture information.