WA-BSN: Self-Supervised Real-World Image Denoising Based on Wavelet-Adaptive Blind Spot Network

Abstract

Blind spot network (BSN) has gained increasing attention with its state-of-the-art performance in self-supervised image denoising. However, most existing BSN models are based on an unrealistic assumption of noise independence and use isotropic mask convolutions, which can lead to the loss of structural details in the denoised image. To address these limitations, we consider the spatially correlated noise and introduce directional adaptive downsampling and mask convolutions to the wavelet domain, resulting in a novel self-supervised denoising method called wavelet-adaptive BSN (WA-BSN). Specifically, we design the direction-adaptive pixel-shuffle downsamplings (PDs) and apply them to the wavelet decomposition subbands, where the spatial-correlated noise is eliminated and the inherent structure is well preserved in the wavelet domain. Then, based on the geometric direction of the wavelet subimages, we propose four shape-adaptive mask convolutions of a smaller size for each wavelet subband in WA-BSN. This enables adaptive pixel prediction within a structural neighborhood for each subband with reduced training time. Finally, total variation (TV) is added to the loss function to further preserve the edges. The results on public real-world datasets demonstrate that our method significantly outperforms existing self-supervised denoising methods and achieves great efficiency.