An edge map-guided acceleration strategy for multi-scale weighted nuclear norm minimization-based image denoising

By exploring the nonlocal self-similarity of images, weighted nuclear norm minimization (WNNM) based image denoising algorithm has achieved competitive performance. Multi-scale patches-based image denoising using weighted nuclear norm minimization (MS-WNNM) is one of the follow-up researches, which further improves WNNM by considering cross-scale self-similarity in block-matching. However, the better performance is achieved at the price of spending more time in a larger search space. In this paper, we propose an edge map-guided strategy to accelerate MS-WNNM based image denoising. Inspired by the fact that patches belonging to the structure part and patches belonging to the texture part follow different repetitive patterns, we propose a prior that the patches containing and surrounding the edges carry most of the structural and textural information of the image. The proposed prior is explained from an information-theoretic perspective. Based on this prior, an edge map-guided acceleration strategy is designed, where redundant patches are trimmed off and the informative patches are retained. Block-matching in the pruned search space, which still has patch diversity, can significantly reduce time without much loss of quality. As a result, comparable denoising can be achieved in less runtime. The experimental results demonstrate the quality and efficiency of the proposed method.