Scholarlink

<p>Haze, a common environmental phenomenon, significantly reduces visibility and impairs tasks such as autonomous driving, traffic monitoring, and other computer vision applications. To improve the performance of downstream tasks, we propose a novel image dehazing algorithm, KACFormer, that integrates CNN and Transformer architectures. Our approach makes three key contributions. First, to effectively combine the strengths of CNN and Transformer architectures, we propose a novel knowledge transfer method. In this approach, a CNN-based teacher network guides a Transformer-based student network, enabling the model to leverage both local detail and global context, resulting in more efficient dehazing. Second, we enhance the teacher network’s dehazing performance by applying a multi-scale attention mechanism, which dynamically allocates resources based on haze density, prioritizing denser regions. Third, we optimize the skip connection module in the Transformer by incorporating depthwise separable convolutions and standard convolutions, enabling the model to better process both deep and shallow features. The processed features are then concatenated for subsequent operations, preserving critical details from both layers. Our method outperforms existing approaches, as demonstrated by experiments on both synthetic and real-world datasets. The code is available at <a href="https://github.com/71717171fan/KACFormer">https://github.com/71717171fan/KACFormer</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

Kacformer: a hybrid CNN–transformer framework for image dehazing via knowledge transfer

  • 期刊论文

  • 发表日期：

    2025年5月29日

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

摘要

Haze, a common environmental phenomenon, significantly reduces visibility and impairs tasks such as autonomous driving, traffic monitoring, and other computer vision applications. To improve the performance of downstream tasks, we propose a novel image dehazing algorithm, KACFormer, that integrates CNN and Transformer architectures. Our approach makes three key contributions. First, to effectively combine the strengths of CNN and Transformer architectures, we propose a novel knowledge transfer method. In this approach, a CNN-based teacher network guides a Transformer-based student network, enabling the model to leverage both local detail and global context, resulting in more efficient dehazing. Second, we enhance the teacher network’s dehazing performance by applying a multi-scale attention mechanism, which dynamically allocates resources based on haze density, prioritizing denser regions. Third, we optimize the skip connection module in the Transformer by incorporating depthwise separable convolutions and standard convolutions, enabling the model to better process both deep and shallow features. The processed features are then concatenated for subsequent operations, preserving critical details from both layers. Our method outperforms existing approaches, as demonstrated by experiments on both synthetic and real-world datasets. The code is available at https://github.com/71717171fan/KACFormer.