Scholarlink

<p>Cracks are the most common form of disrepair in road maintenance projects. Accurate crack detection remains challenging due to diverse morphologies, complex background textures, illumination variations, interfering objects, and micro-crack characteristics. To address these challenges, we propose DHAF-Net, a dual-branch network with hybrid attention fusion for precise crack detection. The main branch extracts crack semantic information, while the auxiliary branch preserves detailed crack features. Furthermore, the network splits channel dimensions into multiple pathways, employing distinct convolution operations in each branch to establish global contextual relationships. Attention mechanisms are introduced to adaptively fuse high-level and low-level features. To verify the effectiveness and accuracy of the proposed method, experiments were conducted on three publicly available crack datasets: DeepCrack, Crack500 and CFD. Our method achieved <i>ODS</i>/ <i>MIoU</i> scores of 0.879/0.886 on DeepCrack, 0.739/0.777 on Crack500 and 0.626/0.721 on CFD, outperforming existing approaches.</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

DHAF-Net: A dual-branch hybrid attention fusion network for pavement crack detection

  • 期刊论文

  • 发表日期：

    2025年6月19日

查看全文
  • Zhong Qu,
  • Hang Liu,
  • Xuehui Yin

摘要

Cracks are the most common form of disrepair in road maintenance projects. Accurate crack detection remains challenging due to diverse morphologies, complex background textures, illumination variations, interfering objects, and micro-crack characteristics. To address these challenges, we propose DHAF-Net, a dual-branch network with hybrid attention fusion for precise crack detection. The main branch extracts crack semantic information, while the auxiliary branch preserves detailed crack features. Furthermore, the network splits channel dimensions into multiple pathways, employing distinct convolution operations in each branch to establish global contextual relationships. Attention mechanisms are introduced to adaptively fuse high-level and low-level features. To verify the effectiveness and accuracy of the proposed method, experiments were conducted on three publicly available crack datasets: DeepCrack, Crack500 and CFD. Our method achieved ODS/ MIoU scores of 0.879/0.886 on DeepCrack, 0.739/0.777 on Crack500 and 0.626/0.721 on CFD, outperforming existing approaches.