Scholarlink

<p>Road crack detection is an essential task for ensuring traffic and driving safety. However, due to the uneven strength of cracks, complex road environments, limited samples, and noise interference, neural network-based segmentation methods still face several challenges in practical applications. To address these issues, this paper presents a crack segmentation network called MFDA-Net based on DeepLabV3+. In MFDA-Net, a feature pyramid pooling module is incorporated which includes depthwise separable dilated convolutions along with a triple attention module to achieve dense multi-scale feature fusion, effectively preserving the detailed information. Furthermore, MFDA-Net introduces a hybrid attention mechanism that enhances the discriminative ability and generalization capacity of the feature representations. Experimental results demonstrate that the proposed model performs exceptionally well on multiple benchmark datasets for crack segmentation tasks, achieving significant improvement in detail capture capabilities along with enhanced noise resistance and robustness in complex scenes.</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

Multi-scale feature dense connection fusion with hybrid attention mechanism in crack segmentation

  • 期刊论文

  • 发表日期：

    2025年9月16日

查看全文
  • Hengyang Liu,
  • Yanjun Liu,
  • Yonglong Li,
  • Guifang Shao,
  • Cong Zhou,
  • Bolin Cao

摘要

Road crack detection is an essential task for ensuring traffic and driving safety. However, due to the uneven strength of cracks, complex road environments, limited samples, and noise interference, neural network-based segmentation methods still face several challenges in practical applications. To address these issues, this paper presents a crack segmentation network called MFDA-Net based on DeepLabV3+. In MFDA-Net, a feature pyramid pooling module is incorporated which includes depthwise separable dilated convolutions along with a triple attention module to achieve dense multi-scale feature fusion, effectively preserving the detailed information. Furthermore, MFDA-Net introduces a hybrid attention mechanism that enhances the discriminative ability and generalization capacity of the feature representations. Experimental results demonstrate that the proposed model performs exceptionally well on multiple benchmark datasets for crack segmentation tasks, achieving significant improvement in detail capture capabilities along with enhanced noise resistance and robustness in complex scenes.