Research On Image Denoising Algorithm Based On Nonlocal Self-similarity And Global Structure Sparsity

Due to the physical limitations of various image acquisition devices,the interference of various random noises inevitably appears in the process of image acquisition and transmission.Noise can be regarded as a kind of signal distortion,which hinders the observation and extraction of useful information.In terms of hardware,purchasing more high-quality equipment can improve the quality of image acquisition from the source,but this method will consume a lot of money.From the aspect of software technology,the image quality can also be improved through the post-processing method of image processing.Image denoising is the basic technology in the field of image analysis and processing,which can effectively remove the clutter and invalid information in the image and extract high-quality images.As the first step of image processing,image denoising helps to improve the accuracy of image subsequent processing,such as visual enhancement,feature detection,target recognition and so on.A variety of image priors play a very important role in the process of image denoising.By combining different image priors to establish an appropriate regularization denoising model,the quality of the denoised image can be effectively restored.In recent years,the methods based on various image priors have been widely studied,such as using the total variation algorithm of image gradient prior,using the non-local self-similarity prior and sparsity prior based on image blocks,and so on,which have achieved excellent denoising results.The nonlocal self-similarity of image is due to the existence of a large number of repeated textures and structure,which leads to the similarity of gray characteristics of many pixels although they are not adjacent in space.Low rank modeling is a classical method.In the process of denoising,the low rank matrix is constructed by using the non-local characteristics of the image,which can effectively restore most of the information of the image.Because the human visual perception system is more sensitive to the high-frequency information such as the texture and edge of the image,enhancing the high-frequency information of the target can improve the visual effect of the restored image.In this paper,we focus on the non-local self-similarity and the global structure regularization of the image to study the image denoising algorithm.The similarity group matrix is constructed by using the nonlocal self-similarity of image.Singular value decomposition(SVD)can achieve the best energy contraction in the sense of least squares,and effectively obtain the low rank approximate estimation of similar block matrix.Because different regions of the image have different redundancy of similar features,this paper first decomposes the image into smooth,texture and edge regions by nonsubsampled contourlet transform(NSCT)and an adaptive block matching method is proposed to improve the accuracy of low rank approximation.Then,adaptive block matching method is performed on different regions to obtain similar block matrix.A new noise standard deviation estimation formula is proposed to approximate the rank of low rank matrix by analyzing the mathematical relationship between image and noise.The low rank estimation of similarity group matrix is carried out by singular value truncation method.After image reconstruction,the underlying denoised image which retains most of the image information is obtained.In order to enhance the global structure of the image,the image is further decomposed into low-frequency components and high-frequency residuals.Four direction gradient operator constraints are added to the high-frequency components to filter the noise signal and reconstruct the high-frequency structure information.Finally,the two parts of prior are combined to optimize iteratively by the augmented Lagrange multiplier method and alternating direction multiplier method,so as to improve the accuracy of denoising effect.In this paper,an image denoising algorithm based on image self-similarity and global structure sparsity is proposed,which can make full use of the redundant information of the image,guide the regional decomposition of the image,and enhance the accuracy of similar patch classification of the image.The new method based on global structure regularization is conducive to enhance the edge and texture information of the image.In the experimental stage,this paper tests the standard gray image to verify the effectiveness of the denoising algorithm.The analysis of experimental results shows that this algorithm can reduce the noise in the image effectively.Compared with the denoising results of other representative algorithms,this algorithm is not only higher in the quantitative index,but also maintains the detail information in the visual perception effect.