Scholarlink

<p>Image dehazing is a key task in the field of computer vision, aiming to estimate the underlying haze-free image from the observed hazy images. Image dehazing methods based on the RGB color space often suffer from color distortion and insufficient detail restoration. To address these issues, this paper proposes HSVC-Net, an image dehazing network based on the HSV color space and the Color Attenuation Prior (CAP). The proposed network adopts a dual-branch architecture: one branch is dedicated to processing the H component, while the other branch focuses on the S and V components. This design is well aligned with the inherent characteristics of the HSV color space, facilitating targeted and component-specific optimization. The proposed network consists of three key modules: the Hue Process Module (HPM), the Top-K Enhanced Sparse Attention Module (TKESA), and the Dual Domain Refinement Module (DDRM). To maintain color consistency, HPM leverages hue frequency-domain information encoded by Fourier features to further enhance the perception and correction of local color deviations. Furthermore, to improve detail restoration, TKESA captures multi-scale spatial features and dynamically sparsifies the multi-head attention to adaptively focus on key local regions, thereby significantly enhancing the representation and recovery of textures and fine details. In addition, DDRM adopts a joint optimization strategy across HSV and RGB color spaces to effectively mitigate detail loss caused by the HSV to RGB conversion, further improving the dehazing performance. Extensive experiments demonstrate that the proposed HSVC-Net outperforms most of the existing methods.</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

HSVC-net: hierarchical spatial-visual collaborative network for color-preserving image dehazing

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Mengmeng Zhang,
  • Bowen Shao,
  • Hongyuan Jing,
  • Hongyun Lu

摘要

Image dehazing is a key task in the field of computer vision, aiming to estimate the underlying haze-free image from the observed hazy images. Image dehazing methods based on the RGB color space often suffer from color distortion and insufficient detail restoration. To address these issues, this paper proposes HSVC-Net, an image dehazing network based on the HSV color space and the Color Attenuation Prior (CAP). The proposed network adopts a dual-branch architecture: one branch is dedicated to processing the H component, while the other branch focuses on the S and V components. This design is well aligned with the inherent characteristics of the HSV color space, facilitating targeted and component-specific optimization. The proposed network consists of three key modules: the Hue Process Module (HPM), the Top-K Enhanced Sparse Attention Module (TKESA), and the Dual Domain Refinement Module (DDRM). To maintain color consistency, HPM leverages hue frequency-domain information encoded by Fourier features to further enhance the perception and correction of local color deviations. Furthermore, to improve detail restoration, TKESA captures multi-scale spatial features and dynamically sparsifies the multi-head attention to adaptively focus on key local regions, thereby significantly enhancing the representation and recovery of textures and fine details. In addition, DDRM adopts a joint optimization strategy across HSV and RGB color spaces to effectively mitigate detail loss caused by the HSV to RGB conversion, further improving the dehazing performance. Extensive experiments demonstrate that the proposed HSVC-Net outperforms most of the existing methods.