Scholarlink

This paper presents an enhanced YOLOv8m-based model for the task of pothole detection in road damage scenarios. Our contributions involve architectural modifications to the original YOLOv8m-seg by integrating additional C2f modules at levels 3, 4, and 5, and incorporating Squeeze-and-Excitation (SE) blocks into each bottleneck to improve feature recalibration. These enhancements aim to improve the model’s precision and robustness when identifying pothole regions, which often present ambiguous edges, small sizes, or occlusions. Experiments conducted on the Pothole Detection Dataset demonstrate notable improvements in mAP50, mA P50 - 95 , and overall fitness. Specifically, the enhanced model achieves a 3.1% increase in fitness (from 0.892 to 0.923), with notable improvements in precision and mAP across both bounding box (B) and mask (M) predictions. While recall shows a slight reduction, this trade-off emphasizes higher precision, which is often desirable for real-world pothole detection where false positives are costly.

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

Robust Pothole Segmentation Using Multi-level C2f Integration and Squeeze-and-Excitation Bottlenecks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Thong Minh Phuc Nguyen,
  • Vinh Dinh Nguyen

摘要

This paper presents an enhanced YOLOv8m-based model for the task of pothole detection in road damage scenarios. Our contributions involve architectural modifications to the original YOLOv8m-seg by integrating additional C2f modules at levels 3, 4, and 5, and incorporating Squeeze-and-Excitation (SE) blocks into each bottleneck to improve feature recalibration. These enhancements aim to improve the model’s precision and robustness when identifying pothole regions, which often present ambiguous edges, small sizes, or occlusions. Experiments conducted on the Pothole Detection Dataset demonstrate notable improvements in mAP50, mA P50 - 95 , and overall fitness. Specifically, the enhanced model achieves a 3.1% increase in fitness (from 0.892 to 0.923), with notable improvements in precision and mAP across both bounding box (B) and mask (M) predictions. While recall shows a slight reduction, this trade-off emphasizes higher precision, which is often desirable for real-world pothole detection where false positives are costly.