Scholarlink

<p>Single image dehazing remains a challenging task, particularly for dense and non-homogeneous haze distributions. Existing methods often struggle to accurately capture the complex distribution of haze features and lack effective detail recovery mechanisms, leading to suboptimal dehazing performance or the introduction of artifacts. To address this, we propose D3-Dehaze, a novel divide-and-conquer framework that comprises haze, channel, and pixel divide-and-conquer stages. The proposed method leverages a hierarchical feature extraction mechanism and adaptive attention fusion, effectively distilling haze-relevant information, enhancing multi-scale channel perception, and refining pixel-wise reconstruction, enabling robust dehazing with superior detail recovery and color accuracy. Extensive experiments on synthetic and real-world datasets, including RESIDE, NTIRE2020-2024, and non-homogeneous hazy images, demonstrate the superiority of our method. Notably, D3-Dehaze achieves state-of-the-art performance, with PSNR values of 41.10&#xa0;dB and 35.25&#xa0;dB on the SOTS-indoor and outdoor datasets, respectively. Furthermore, we extend D3-Dehaze to single image deraining, underlining its versatility and effectiveness. The code of this work is publicly available at: <a href="https://github.com/zhanxiyan/D3-Dehaze">https://github.com/zhanxiyan/D3-Dehaze</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

D3-Dehaze: a divide-and-conquer framework for enhanced single image dehazing

  • 期刊论文

  • 发表日期：

    2025年1月9日

查看全文
  • Zhanqiang Huo,
  • Xiyan Zhan,
  • Yingxu Qiao,
  • Shan Zhao

摘要

Single image dehazing remains a challenging task, particularly for dense and non-homogeneous haze distributions. Existing methods often struggle to accurately capture the complex distribution of haze features and lack effective detail recovery mechanisms, leading to suboptimal dehazing performance or the introduction of artifacts. To address this, we propose D3-Dehaze, a novel divide-and-conquer framework that comprises haze, channel, and pixel divide-and-conquer stages. The proposed method leverages a hierarchical feature extraction mechanism and adaptive attention fusion, effectively distilling haze-relevant information, enhancing multi-scale channel perception, and refining pixel-wise reconstruction, enabling robust dehazing with superior detail recovery and color accuracy. Extensive experiments on synthetic and real-world datasets, including RESIDE, NTIRE2020-2024, and non-homogeneous hazy images, demonstrate the superiority of our method. Notably, D3-Dehaze achieves state-of-the-art performance, with PSNR values of 41.10 dB and 35.25 dB on the SOTS-indoor and outdoor datasets, respectively. Furthermore, we extend D3-Dehaze to single image deraining, underlining its versatility and effectiveness. The code of this work is publicly available at: https://github.com/zhanxiyan/D3-Dehaze.