Around2center: Self-Supervised Denoising From the Around to the Center Using Masks

For image denoising tasks, it is often fraught with many challenges to obtain paired noisy images and clean images. While self-supervised learning promises to mitigate reliance on manual labels, generating meaningful supervisory signals remains a challenge. Existing masking strategies, while augmenting training data diversity, can alter noise structure and lead to edge effects. To address these challenges, we proposed an framework named Around2center for denoising using only a single noisy image. Firstly, to enhance local feature extraction and accentuate noise-signal boundaries, we divided the input image into multiple scales. Then, we employed a novel edge corner masker to strategically mask the image, enabling targeted processing of specific regions while mitigating identity mapping and preserving relevant information. A mask integrator innovatively fused the denoised image and masked plates, organically combining local and global information to effectively address denoising tasks in diverse scenarios. Notably, the generated integrated image served as a robust self-supervised label, eliminating the need for paired clean-noisy datasets. Finally, we provided a theoretical analysis of the loss function’s convergence properties, ensuring the framework’s soundness and efficacy. Extensive quantitative and qualitative evaluation results fully confirmed the superior denoising performance of our proposed method when dealing with both synthetic noise and real noise datasets. The code is available at https://github.com/eternity-city/Around2center.