Scholarlink

<p>Image defogging is a challenging and hot-spot issue in the field of computer vision. Existing learning methods usually employ a single convolutional neural network (CNN) model to address it. However, such methods often overlook the restoration of edge details and exhibit poor defogging performance under non-uniform haze conditions. To solve the above two problems, this paper proposes a dual-branch defogging generative adversarial network that incorporates attention perception and contrastive learning. (1) The Residual Attention Branch (RAB) aims to generate attention feature maps, and the Scene Reconstruction Branch (SRB) is used to reconstruct haze-free images. (2) Considering that edge texture details may be lost in defogged images, an Attention-Aware Corrector (AAC) is proposed. (3) A Multi-Scale Fusion Discriminator (MFD) is used to supervise the restoration of haze-free images. (4) Contrastive learning is introduced as a loss function in network training, which further improves the quality of restored images. Experimental results show that the method proposed in this paper outperforms relevant classical methods on both synthetic and real-world datasets, providing new ideas and a technical baseline for image defogging.</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

A Dual-Branch Defogging Generative Adversarial Network Incorporating Attention Perception and Contrastive Learning

  • 期刊论文

  • 发表日期：

    2026年5月18日

查看全文
  • Yan Wang,
  • Xiaoyu Hou,
  • Jinyuan Hu,
  • Chengqing Cai

摘要

Image defogging is a challenging and hot-spot issue in the field of computer vision. Existing learning methods usually employ a single convolutional neural network (CNN) model to address it. However, such methods often overlook the restoration of edge details and exhibit poor defogging performance under non-uniform haze conditions. To solve the above two problems, this paper proposes a dual-branch defogging generative adversarial network that incorporates attention perception and contrastive learning. (1) The Residual Attention Branch (RAB) aims to generate attention feature maps, and the Scene Reconstruction Branch (SRB) is used to reconstruct haze-free images. (2) Considering that edge texture details may be lost in defogged images, an Attention-Aware Corrector (AAC) is proposed. (3) A Multi-Scale Fusion Discriminator (MFD) is used to supervise the restoration of haze-free images. (4) Contrastive learning is introduced as a loss function in network training, which further improves the quality of restored images. Experimental results show that the method proposed in this paper outperforms relevant classical methods on both synthetic and real-world datasets, providing new ideas and a technical baseline for image defogging.