Learning Multi-scale Spatial-frequency Features for Image Denoising

2025-06-19

AI Summary

Existing image denoising methods predominantly employ fixed single-scale U-Net architectures, limiting their capacity to model pixel-level multi-scale representations and failing to explicitly account for the distinct spectral characteristics of high-frequency (texture distortion) and low-frequency (blurring) noise components. To address these limitations, we propose the Multi-scale Adaptive Dual-domain Network (MADNet). Its core innovations include: (1) a learnable mask-driven Spatial-Frequency Adaptive Unit (ASFU) that enables band-aware noise modeling; and (2) an integrated design incorporating image pyramid inputs, frequency-domain decomposition, and cross-scale global feature skip connections. Extensive experiments on synthetic (SIDD, DND) and real-world noisy datasets demonstrate that MADNet consistently outperforms state-of-the-art methods—particularly in texture detail recovery and high-frequency noise suppression—with average PSNR and SSIM improvements of 0.82 dB and 0.013, respectively.

Technology Category

• Computer Vision: Multi-modal Vision
• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Multiagent Systems: Adversarial Agents

Application Category

• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Recent advancements in multi-scale architectures have demonstrated exceptional performance in image denoising tasks. However, existing architectures mainly depends on a fixed single-input single-output Unet architecture, ignoring the multi-scale representations of pixel level. In addition, previous methods treat the frequency domain uniformly, ignoring the different characteristics of high-frequency and low-frequency noise. In this paper, we propose a novel multi-scale adaptive dual-domain network (MADNet) for image denoising. We use image pyramid inputs to restore noise-free results from low-resolution images. In order to realize the interaction of high-frequency and low-frequency information, we design an adaptive spatial-frequency learning unit (ASFU), where a learnable mask is used to separate the information into high-frequency and low-frequency components. In the skip connections, we design a global feature fusion block to enhance the features at different scales. Extensive experiments on both synthetic and real noisy image datasets verify the effectiveness of MADNet compared with current state-of-the-art denoising approaches.

Problem

• Fixed single-input single-output Unet architecture limitations
• Uniform treatment of frequency domain ignoring noise differences
• Lack of multi-scale representations in pixel level

Innovation

• Multi-scale adaptive dual-domain network (MADNet)
• Adaptive spatial-frequency learning unit (ASFU)
• Global feature fusion block in skip connections