SC-BSN: Shifted Convolutions Based Blind-Spot Network for self-supervised image denoising

Abstract

Self-supervised image denoising methods have garnered significant attention recently due to their ability to train solely on noisy images without requiring paired clean-noisy data. However, real-world noise is often spatially correlated, leading to poor performance in self-supervised algorithms that assume pixel-wise independent noise. To address this limitation, we design multi-branch directional shifted operations to create blind spots in different regions, which effectively disrupt noise correlation. Further, the Shifted Convolutions Blind-Spot Network (SC-BSN) is proposed for self-supervised denoising. This network leverages three distinct blind-spot branches with varying shifted distances to effectively balance noise correlation suppression and the preservation of local spatial structures. Finally, we develop the Complementary Random-Replacing Refinement (CR3) to complement denoising results instead of relying on the iterative averaging of R3. The new post-processing technique efficiently retains the details of denoised images. Experimental results demonstrate that SC-BSN outperforms existing self-supervised denoising methods across multiple datasets, achieving superior performance in both visual quality and quantitative metrics.