Scholarlink

<p>In hazy weather, the quality of aerial images suffers severe degradation, which affects the imaging capabilities of advanced remote sensing applications. Most existing learning-based dehazing algorithms utilize manually designed deep structures to enhance model performance. However, these deep networks often contain redundant branches, leading to a significant decrease in computational efficiency. To address this issue, we first conduct experiments to explore the distribution of haze and introduce an improved dehazing strategy in the YUV space. Subsequently, we propose a Runge-Kutta (RK) method-inspired aerial image dehazing network called RKNet, which consists of two parts: luminance (Y) domain haze removal and chrominance (U,V) enhancement. From the perspective of dynamical systems and based on the 3rd-order RK method, we design an RK3 block and incorporate it into RKNet to improve computational accuracy. Experimental evaluations on both synthetic and real-world benchmarks demonstrate that RKNet outperforms current haze removal algorithms and achieves superior performance. In addition, RKNet can improve the detection accuracy and capability of high-level vision algorithms and is also applicable to the processing of sandstorm images and underwater images.</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

Runge-kutta method inspired aerial image dehazing network in YUV space

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月3日

查看全文
  • Yazhong Si,
  • Jingyu Chen,
  • Wenqing Rao,
  • Chen Li

摘要

In hazy weather, the quality of aerial images suffers severe degradation, which affects the imaging capabilities of advanced remote sensing applications. Most existing learning-based dehazing algorithms utilize manually designed deep structures to enhance model performance. However, these deep networks often contain redundant branches, leading to a significant decrease in computational efficiency. To address this issue, we first conduct experiments to explore the distribution of haze and introduce an improved dehazing strategy in the YUV space. Subsequently, we propose a Runge-Kutta (RK) method-inspired aerial image dehazing network called RKNet, which consists of two parts: luminance (Y) domain haze removal and chrominance (U,V) enhancement. From the perspective of dynamical systems and based on the 3rd-order RK method, we design an RK3 block and incorporate it into RKNet to improve computational accuracy. Experimental evaluations on both synthetic and real-world benchmarks demonstrate that RKNet outperforms current haze removal algorithms and achieves superior performance. In addition, RKNet can improve the detection accuracy and capability of high-level vision algorithms and is also applicable to the processing of sandstorm images and underwater images.