Single image denoising via multi-scale weighted group sparse coding

• Published in: Signal processing Vol. 200; p. 108650
• Main Authors: Ou, Yang, Swamy, M.N.S., Luo, Jianqiao, Li, Bailin
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.11.2022
• Subjects: image denoising; Model-driven method; multi-scale self-similarity; weighted group sparse coding; weighted group sparse coding; Model-driven method; image denoising; multi-scale self-similarity
• ISSN: 0165-1684; 1872-7557
• DOI: 10.1016/j.sigpro.2022.108650

•We propose a new MS-WGSC model, in which the multi-scale NSS is exploited. Besides, the dictionary in each group is adaptively set via singular value decomposition rather than learning from external corpus.•We develop an alternating minimization method to solve our MS-WGSC model, wherein each of the steps has closed-form solution.•We show through extensive denoising experiments the competitiveness of our MS-WGSC model compared with many of the state-of-the-art methods, especially in perceptual quality. Nonlocal self-similarity (NSS) property of natural images, which means that the structure of the image sub-patches will appear repeatedly within a certain area, has been widely exploited as an effective prior to establishing various models in image denoising task. However, most of the existing NSS-based denoising models exploit the NSS prior in single scale only, and for some of the image patches that do not appear repeatedly, undesirable ringing artifacts will occur in the restored image, and even the image content may be lost. Considering the fact that NSS exists both within the same scale and across different scales, in order to better restore the structure and the edges of images contaminated by noise, we propose, in this paper, a novel multi-scale weighted group sparse coding model (MS-WGSC) for image denoising, wherein the patch groups are constructed using multi-scale NSS priors. Furthermore, an alternating minimization method is proposed to obtain the solution for our model. Extensive experiments are conducted that demonstrate the competitiveness of the proposed model compared with that of state-of-the-art methods not only in terms of the quantitative metrics such as PSNR and SSIM, but also in perceptual quality.