Scholarlink

<p>Image dehazing is crucial for enhancing environmental perception robustness in vision-based systems like autonomous driving and remote sensing. However, real-world haze presents complex multi-scale variations and spatial heterogeneity, posing significant challenges to existing dehazing algorithms. We propose Diff-HazeNet, a diffusion-based generative framework designed to enhance generalization under diverse real-world haze conditions. Unlike conventional end-to-end convolutional approaches, our method leverages a dual-branch diffusion architecture that jointly utilizes spatial and frequency-domain information. The diffusion restoration branch incorporates multi-frequency feature decoupling module and dual-prior restoration module to capture textures and structural cues while fusing multiple spatial priors. The deep feature prediction branch employs the wavelet enhancement module to reconstruct intrinsic inter-channel dependencies. Extensive experiments on real-world datasets demonstrate that Diff-HazeNet consistently outperforms existing methods across urban, forest, nighttime, and cross-climate scenarios, highlighting its robustness, adaptability, and practical potential. The code of this work is publicly available at: <a href="https://github.com/lijang2000/Diff-HazeNet">https://github.com/lijang2000/Diff-HazeNet</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

Diff-HazeNet: enhancing real-world image dehazing via diffusion models in complex conditions

  • 期刊论文

  • 发表日期：

    2025年8月22日

查看全文
  • Jiaang Li,
  • Jiaming Niu,
  • Yuting Pang,
  • Jianlei Liu

摘要

Image dehazing is crucial for enhancing environmental perception robustness in vision-based systems like autonomous driving and remote sensing. However, real-world haze presents complex multi-scale variations and spatial heterogeneity, posing significant challenges to existing dehazing algorithms. We propose Diff-HazeNet, a diffusion-based generative framework designed to enhance generalization under diverse real-world haze conditions. Unlike conventional end-to-end convolutional approaches, our method leverages a dual-branch diffusion architecture that jointly utilizes spatial and frequency-domain information. The diffusion restoration branch incorporates multi-frequency feature decoupling module and dual-prior restoration module to capture textures and structural cues while fusing multiple spatial priors. The deep feature prediction branch employs the wavelet enhancement module to reconstruct intrinsic inter-channel dependencies. Extensive experiments on real-world datasets demonstrate that Diff-HazeNet consistently outperforms existing methods across urban, forest, nighttime, and cross-climate scenarios, highlighting its robustness, adaptability, and practical potential. The code of this work is publicly available at: https://github.com/lijang2000/Diff-HazeNet.