Scholarlink

<p>Road surface crack detection plays a vital role in intelligent transportation and infrastructure maintenance. Existing lightweight detection models struggle with slender crack feature extraction, fine crack missed detection, and inaccurate localization in complex scenes. This work proposes a lightweight detection framework with directional self-attention multi-scale feature interaction, direction-guided edge gated convolution, and CIoU loss optimization. The method enhances crack feature representation, suppresses background noise, strengthens fine edge perception, and improves bounding box regression accuracy. Experiments on the RDD2022 dataset show that our model achieves 83.6% precision, 88.1% recall, and 80.8% mAP@0.5 with only 2.6M parameters and 127.6 FPS, balancing accuracy and efficiency for real-world deployment. The codes are available at <a href="https://github.com/miao185/DAM-D-FINE">https://github.com/miao185/DAM-D-FINE</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

Directional attention and edge enhancement for lightweight road surface crack detection

  • 期刊论文

  • 发表日期：

    2026年7月4日

查看全文
  • Yun Bai,
  • Shuangjie Miao,
  • Qian Xu

摘要

Road surface crack detection plays a vital role in intelligent transportation and infrastructure maintenance. Existing lightweight detection models struggle with slender crack feature extraction, fine crack missed detection, and inaccurate localization in complex scenes. This work proposes a lightweight detection framework with directional self-attention multi-scale feature interaction, direction-guided edge gated convolution, and CIoU loss optimization. The method enhances crack feature representation, suppresses background noise, strengthens fine edge perception, and improves bounding box regression accuracy. Experiments on the RDD2022 dataset show that our model achieves 83.6% precision, 88.1% recall, and 80.8% mAP@0.5 with only 2.6M parameters and 127.6 FPS, balancing accuracy and efficiency for real-world deployment. The codes are available at https://github.com/miao185/DAM-D-FINE.