Scholarlink

<p>Haze severely reduces image visibility and structural clarity, posing major challenges for high-level vision tasks. Single-image dehazing methods relying only on visible images often fail to restore fine structures and maintain color fidelity under dense haze. To address this issue, we propose a Haze-Aware Prompt Fusion Network for Visible-Infrared Image Dehazing (VI-HAPFNet). The network fully leverages the strong penetration capability of infrared imaging under hazy conditions, overcoming the limitations of single-image dehazing methods in structural restoration and color fidelity. We introduce haze-aware prompts to guide feature modulation, enabling the network to adaptively focus on heavily degraded regions while preserving fine-grained structures. By combining haze density estimation with structural inconsistency modeling, the network achieves dynamic fusion of visible and infrared features, effectively suppressing irrelevant infrared noise and enhancing the robustness of feature integration. Experimental results on the AirSim-VID, Dense-Haze, and NH-Haze datasets demonstrate that VI-HAPFNet outperforms the existing single-image and visible-infrared multimodal fusion methods in terms of PSNR and SSIM, while preserving more image details and structural information visually. The code is available at <a href="https://github.com/zhang-J06/VI-HAPFNet">https://github.com/zhang-J06/VI-HAPFNet</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

VI-HAPFNet: Haze-aware prompt fusion network for visible-infrared image dehazing

  • 期刊论文

  • 发表日期：

    2026年5月9日

查看全文
  • Wei Xiao,
  • Jiajia Zhang,
  • Man Liu,
  • Jingzhi Bi,
  • Yanji Guo,
  • Chao Pan

摘要

Haze severely reduces image visibility and structural clarity, posing major challenges for high-level vision tasks. Single-image dehazing methods relying only on visible images often fail to restore fine structures and maintain color fidelity under dense haze. To address this issue, we propose a Haze-Aware Prompt Fusion Network for Visible-Infrared Image Dehazing (VI-HAPFNet). The network fully leverages the strong penetration capability of infrared imaging under hazy conditions, overcoming the limitations of single-image dehazing methods in structural restoration and color fidelity. We introduce haze-aware prompts to guide feature modulation, enabling the network to adaptively focus on heavily degraded regions while preserving fine-grained structures. By combining haze density estimation with structural inconsistency modeling, the network achieves dynamic fusion of visible and infrared features, effectively suppressing irrelevant infrared noise and enhancing the robustness of feature integration. Experimental results on the AirSim-VID, Dense-Haze, and NH-Haze datasets demonstrate that VI-HAPFNet outperforms the existing single-image and visible-infrared multimodal fusion methods in terms of PSNR and SSIM, while preserving more image details and structural information visually. The code is available at https://github.com/zhang-J06/VI-HAPFNet.