Scholarlink

<p>Real-world image dehazing remains challenging under spatially varying haze, where most existing methods still treat haze as a static degradation to be directly removed. Such formulations often struggle with non-uniform haze and may produce residual haze, over-enhancement, color inconsistency, and structural instability. In this work, we shift the perspective from haze degradation removal to haze evolution modeling and propose a diffusion-inspired generative framework for image dehazing. Specifically, we represent haze as a structure-aware haze field and progressively evolve it toward a more region-consistent state in latent space, providing an explicit structural prior for modeling region-level haze inconsistency in non-uniform haze scenes. Based on this idea, we develop <b>LAD</b>, a latent anisotropic diffusion dehazing framework in which haze evolution guides haze-aware representation learning, structure-constrained latent generation, and luminance–color-consistent reconstruction. By coupling explicit haze-field modeling with generative restoration, our method better aligns the dehazing process with the spatially varying nature of real-world haze while reducing hallucinations and perceptual artifacts. Extensive experiments on multiple benchmarks demonstrate competitive overall performance, with clear advantages on several paired benchmarks and key metrics. Specifically, LAD attains 26.33 dB PSNR on Haze4K, outperforming the strongest competing method by 2.82 dB. It also achieves 0.77 SSIM on O-Haze, exceeding the previous best result by 0.09. The project code is publicly available at <a href="https://github.com/miraclebai/LAD">https://github.com/miraclebai/LAD</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

From haze degradation to haze evolution: an anisotropic-diffusion-inspired generative framework for image dehazing

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Jingyuan Bai,
  • Jingang Wang,
  • Peng Liu,
  • Zhenrong Zhang

摘要

Real-world image dehazing remains challenging under spatially varying haze, where most existing methods still treat haze as a static degradation to be directly removed. Such formulations often struggle with non-uniform haze and may produce residual haze, over-enhancement, color inconsistency, and structural instability. In this work, we shift the perspective from haze degradation removal to haze evolution modeling and propose a diffusion-inspired generative framework for image dehazing. Specifically, we represent haze as a structure-aware haze field and progressively evolve it toward a more region-consistent state in latent space, providing an explicit structural prior for modeling region-level haze inconsistency in non-uniform haze scenes. Based on this idea, we develop LAD, a latent anisotropic diffusion dehazing framework in which haze evolution guides haze-aware representation learning, structure-constrained latent generation, and luminance–color-consistent reconstruction. By coupling explicit haze-field modeling with generative restoration, our method better aligns the dehazing process with the spatially varying nature of real-world haze while reducing hallucinations and perceptual artifacts. Extensive experiments on multiple benchmarks demonstrate competitive overall performance, with clear advantages on several paired benchmarks and key metrics. Specifically, LAD attains 26.33 dB PSNR on Haze4K, outperforming the strongest competing method by 2.82 dB. It also achieves 0.77 SSIM on O-Haze, exceeding the previous best result by 0.09. The project code is publicly available at https://github.com/miraclebai/LAD.