Asymmetric Mask Scheme for Self-Supervised Real Image Denoising

TL;DR

This paper tackles real-world image denoising in a self-supervised setting by removing blind-spot constraints with a mask-based denoising strategy inspired by MAE. It introduces an asymmetric framework (AMSNet) that trains with a single mask yet performs full-image denoising at inference using multiple masks (MMDB), coupled with pixel downsampling to handle correlated real noise. The approach is validated on real noisy datasets (SIDD, DND, PolyU) and self-captured images, achieving state-of-the-art results and demonstrating robustness across diverse denoisers, with ablations showing the importance of mask ratios and the proposed smoothness refinement. The method offers a flexible, scalable solution that improves denoising quality while relaxing network-design constraints, potentially benefiting a wide range of real-world imaging applications.

Abstract

In recent years, self-supervised denoising methods have gained significant success and become critically important in the field of image restoration. Among them, the blind spot network based methods are the most typical type and have attracted the attentions of a large number of researchers. Although the introduction of blind spot operations can prevent identity mapping from noise to noise, it imposes stringent requirements on the receptive fields in the network design, thereby limiting overall performance. To address this challenge, we propose a single mask scheme for self-supervised denoising training, which eliminates the need for blind spot operation and thereby removes constraints on the network structure design. Furthermore, to achieve denoising across entire image during inference, we propose a multi-mask scheme. Our method, featuring the asymmetric mask scheme in training and inference, achieves state-of-the-art performance on existing real noisy image datasets. All the source code will be made available to the public.

Figures (11)

• Figure 1: Visual comparison of denoising results on the SIDD validation dataset abdelhamed2018high with various methods and our AMSNet is able to preserve more details and achieve better visual effects.
• Figure 2: The effect of different receptive fields after blind spot convolution on the final denoising result.
• Figure 3: Denoising via mask matrix $M$ and the noisy image $I_N$.
• Figure 4: Pixel downsamplinga and mask. Here we take the $P_2$ as an example.
• Figure 5: Overview of the proposed asymmetric mask scheme. The $D$ is depicted in \ref{['fig:mask_denoise']}, which takes sub-samples set $I_s$ of the noisy image and the corresponding binary mask matrices set $M_{s}$ to denoise the specified regions. Here, we present a configuration with 2 denoising branches as an example. During the training phase, a single branch is employed for optimization of denoiser. During the inference phase, we utilize all branches to derive restoration results for entire noisy image.