Scholarlink

<p>Existing dehazing methods often overlook the heterogeneous haze distribution across RGB channels, rely heavily on synthetic datasets with limited real-world generalization, and underutilize the benefits of skip connections and semantic priors. To address these challenges, we propose ColorBranchNet, a novel dehazing framework that processes R, G, and B channels independently to capture channel-specific haze characteristics better. To fuse multi-branch features and improve adaptability, we introduce Residual Dense Meta Blocks (RDMB), which leverage meta-learning for dynamic feature weighting. We further design Attention-Guided Skip Connections (AGSC) to preserve fine-grained details and propose a dedicated color loss to mitigate chromatic distortions. A semantic segmentation module provides high-level contextual guidance, refined via attention for adaptive enhancement. Extensive experiments demonstrate that ColorBranchNet achieves superior performance over state-of-the-art methods in both quantitative and qualitative evaluations. Notably, it generalizes well to real-world hazy images despite training exclusively on synthetic data. Code is available at: <a href="https://github.com/71717171fan/ColorBranchNet">https://github.com/71717171fan/ColorBranchNet</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

ColorBranchNet: A Triple-Channel Branch Network for Single Image Dehazing

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Bingqing Yang,
  • Maoli Wang,
  • Jianlei Liu

摘要

Existing dehazing methods often overlook the heterogeneous haze distribution across RGB channels, rely heavily on synthetic datasets with limited real-world generalization, and underutilize the benefits of skip connections and semantic priors. To address these challenges, we propose ColorBranchNet, a novel dehazing framework that processes R, G, and B channels independently to capture channel-specific haze characteristics better. To fuse multi-branch features and improve adaptability, we introduce Residual Dense Meta Blocks (RDMB), which leverage meta-learning for dynamic feature weighting. We further design Attention-Guided Skip Connections (AGSC) to preserve fine-grained details and propose a dedicated color loss to mitigate chromatic distortions. A semantic segmentation module provides high-level contextual guidance, refined via attention for adaptive enhancement. Extensive experiments demonstrate that ColorBranchNet achieves superior performance over state-of-the-art methods in both quantitative and qualitative evaluations. Notably, it generalizes well to real-world hazy images despite training exclusively on synthetic data. Code is available at: https://github.com/71717171fan/ColorBranchNet.