Scholarlink

<p>This paper proposes a novel denoising diffusion probabilistic model for underwater image enhancement (UIE). It effectively addresses common issues in underwater images&#xa0;such as color distortion, low contrast, and detail blurring. We design an innovative Transformer denoising network that integrates&#xa0;a multidimensional collaborative attention (MCA) mechanism, which employs&#xa0;a three branch parallel structure to model feature interactions&#xa0;across&#xa0;channel, height, and width dimensions,&#xa0;further improving enhancement performance. Additionally, we propose&#xa0;the&#xa0;channel-enhanced local attention (CELA) module,&#xa0;which first performs global interaction at the channel level and then applies local neighborhood attention. This design ensures&#xa0;that&#xa0;local details are guided by global context,&#xa0;enabling effective integration of&#xa0;long-range dependencies and local semantics. To address the limitations of conventional feed-forward networks (FFNs) in spatial and channel interaction, we further design the channel-enhanced feed-forward network (CEFN). This module combines grouped convolution with a lightweight&#xa0;yet efficient&#xa0;channel attention mechanism,&#xa0;facilitating&#xa0;synergistic enhancement of spatial patterns and channel features. Extensive experimental results on diverse datasets (UIEB, LSUI, UFO, C60, U45) demonstrate that our method outperforms state-of-the-art approaches in both qualitative and quantitative evaluations.</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

CELAT-DiffNet: channel-enhanced local attention transformer for underwater image enhancement based on diffusion models

  • 期刊论文

  • 发表日期：

    2025年8月29日

查看全文
  • Jiale Wang,
  • Yitao Liang,
  • Xuxu Yang,
  • Mengjuan Zhao,
  • Juan Xia

摘要

This paper proposes a novel denoising diffusion probabilistic model for underwater image enhancement (UIE). It effectively addresses common issues in underwater images such as color distortion, low contrast, and detail blurring. We design an innovative Transformer denoising network that integrates a multidimensional collaborative attention (MCA) mechanism, which employs a three branch parallel structure to model feature interactions across channel, height, and width dimensions, further improving enhancement performance. Additionally, we propose the channel-enhanced local attention (CELA) module, which first performs global interaction at the channel level and then applies local neighborhood attention. This design ensures that local details are guided by global context, enabling effective integration of long-range dependencies and local semantics. To address the limitations of conventional feed-forward networks (FFNs) in spatial and channel interaction, we further design the channel-enhanced feed-forward network (CEFN). This module combines grouped convolution with a lightweight yet efficient channel attention mechanism, facilitating synergistic enhancement of spatial patterns and channel features. Extensive experimental results on diverse datasets (UIEB, LSUI, UFO, C60, U45) demonstrate that our method outperforms state-of-the-art approaches in both qualitative and quantitative evaluations.