Scholarlink

<p>Traditional image restoration methods treat degradation types independently, neglecting intrinsic connections between degradation phenomena. While All-in-one models handle multiple degradations within single networks, they fail to fully exploit correlations between different degradation types, limiting performance in complex scenarios. This paper proposes FSDRNet (Frequency-Spatial Collaborative Dynamic Recovery Network), achieving efficient restoration through multi-domain collaborative enhancement and dynamic routing. The spatial module employs balanced feature mixing attention (EFMA) for precise local texture characterization, while the frequency domain module extracts global structural patterns via Fourier transforms, transmitting key information through degradation guidance signals to establish unidirectional enhancement. The network adaptively adjusts parameters according to degradation types and achieves precise restoration through physically-guided specialized filters. Extensive experiments demonstrate FSDRNet’s excellent performance across denoising, deblurring, dehazing, deraining, and low-light enhancement tasks, achieving 28.52dB average PSNR while exhibiting significant generalization capability for unknown compound degradations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FSDRNet: A dynamic image restoration network based on frequency-spatial collaborative processing

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Xiwen Zhang,
  • Junxian Li,
  • Fangyan Dong,
  • Kewei Chen

摘要

Traditional image restoration methods treat degradation types independently, neglecting intrinsic connections between degradation phenomena. While All-in-one models handle multiple degradations within single networks, they fail to fully exploit correlations between different degradation types, limiting performance in complex scenarios. This paper proposes FSDRNet (Frequency-Spatial Collaborative Dynamic Recovery Network), achieving efficient restoration through multi-domain collaborative enhancement and dynamic routing. The spatial module employs balanced feature mixing attention (EFMA) for precise local texture characterization, while the frequency domain module extracts global structural patterns via Fourier transforms, transmitting key information through degradation guidance signals to establish unidirectional enhancement. The network adaptively adjusts parameters according to degradation types and achieves precise restoration through physically-guided specialized filters. Extensive experiments demonstrate FSDRNet’s excellent performance across denoising, deblurring, dehazing, deraining, and low-light enhancement tasks, achieving 28.52dB average PSNR while exhibiting significant generalization capability for unknown compound degradations.