Scholarlink

<p>Underwater images often suffer from severe quality degradation due to selective absorption and scattering of light, leading to color distortion, contrast reduction, haze, and blurring. While various enhancement techniques exist, traditional methods often fall short in handling complex underwater degradation, and existing deep learning approaches typically overlook critical multiscale characteristics across spatial and frequency domains. To address these limitations, we propose DMF-Net, a novel underwater image enhancement network based on Dynamic Multiscale Feature Fusion. The network first decomposes the input image into red, green, and blue channels, each optimized by a Multiscale Spatial-Frequency Hybrid (MSFH) block and a Dual Attention with Parallel Aggregation (DAPA) module. A Dynamic Fusion Gate (DFGate) then adaptively fuses the channels. The fused features are further refined via lightweight dynamic convolution and DAPA before final reconstruction. Extensive experiments on public datasets, including UIEB (R90, C60), EUVP, SQUID, and U45, demonstrate that DMF-Net achieves state-of-the-art performance, with top scores in key metrics such as PSNR (24.86 dB on R90) and SSIM (0.9365 on R90). It also shows significant improvements in downstream tasks like keypoint detection and edge extraction. The source code is available at: <a href="https://github.com/PanChao777/DMF/tree/master">https://github.com/PanChao777/DMF/tree/master</a>.</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

DMF-Net: Dynamic Multiscale Feature Fusion for Underwater Image Enhancement

  • 期刊论文

  • 发表日期：

    2026年1月28日

查看全文
  • Chao Pan,
  • Yuxin Wu,
  • Xun Wang,
  • Xin Shu

摘要

Underwater images often suffer from severe quality degradation due to selective absorption and scattering of light, leading to color distortion, contrast reduction, haze, and blurring. While various enhancement techniques exist, traditional methods often fall short in handling complex underwater degradation, and existing deep learning approaches typically overlook critical multiscale characteristics across spatial and frequency domains. To address these limitations, we propose DMF-Net, a novel underwater image enhancement network based on Dynamic Multiscale Feature Fusion. The network first decomposes the input image into red, green, and blue channels, each optimized by a Multiscale Spatial-Frequency Hybrid (MSFH) block and a Dual Attention with Parallel Aggregation (DAPA) module. A Dynamic Fusion Gate (DFGate) then adaptively fuses the channels. The fused features are further refined via lightweight dynamic convolution and DAPA before final reconstruction. Extensive experiments on public datasets, including UIEB (R90, C60), EUVP, SQUID, and U45, demonstrate that DMF-Net achieves state-of-the-art performance, with top scores in key metrics such as PSNR (24.86 dB on R90) and SSIM (0.9365 on R90). It also shows significant improvements in downstream tasks like keypoint detection and edge extraction. The source code is available at: https://github.com/PanChao777/DMF/tree/master.