A multi-scale dual-branch spatial attention network with dynamic training for zero-shot image denoising

Abstract

Zero-shot image denoising aims to restore a clean image from a single noisy observation without relying on external training data, which remains a challenging problem due to extremely limited supervision and complex noise characteristics. To address this issue, we propose a zero-shot denoising framework that combines a robust and efficient network architecture with a Dynamic Multi-Scale Pixel Bank (DMPB) training strategy. On the network side, we design a Multi-Scale Dual-Branch Spatial Attention Network (MSDB-SANet), which integrates a Multi-scale Dual-branch Feature Extractor (MDFE) to jointly model fine-grained local details and enlarged receptive-field context. A lightweight Multiscale Spatial Attention (MSA) module is further introduced to emphasize informative spatial regions and suppress spatially correlated noise, followed by a Multi-scale Convolutional Refinement Block (MCRB) to improve reconstruction fidelity. Moreover, to enhance the effectiveness of pixel-bank-based zero-shot training, we propose a Dynamic Multi-Scale Pixel Bank (DMPB) training strategy. By dynamically rescaling the pixel bank during image-specific optimization, the proposed strategy exposes the network to pseudo training pairs at multiple spatial resolutions derived from a single noisy image, which enhances robustness to scale variations and improves optimization stability under the zero-shot setting. Extensive experiments on two publicly available benchmark datasets demonstrate that the proposed method achieves highly competitive performance compared with existing denoising approaches across different noise types and noise levels, while producing visually more consistent and faithful results.