Scholarlink

<p>Timely detection of road infrastructure defects is important for traffic safety, maintenance efficiency, and transportation system resilience. While existing methods can detect small targets such as road cracks, challenges remain in low-contrast crack recognition, small-target detection, multi-category road facility defect identification, and edge-device deployment. This study introduces a comprehensive road defect detection system deployed on an edge device with a Rockchip RK3576 chip, enabling near-real-time inference and automatic result uploading to cloud platforms. To address the difficulty of detecting low-contrast and crack-like defects, we design a task-oriented lightweight crack enhancement block, ECEM, within the YOLOv11 framework. Unlike generic attention-based YOLO modifications, ECEM integrates multi-scale depth-wise feature extraction, attention refinement, and background suppression to enhance elongated crack responses while maintaining lightweight computation. Evaluated on the RFDD test set, the S-scale variant ECEM-YOLO-S achieves 73.0% Precision, 67.3% mAP50, and 895.60 ± 1.97 FPS on an NVIDIA RTX 4090 under the unified Ultralytics validation pipeline. Compared with YOLOv11-S, it improves Precision and mAP50, with a slight decrease in Recall, reflecting a precision-oriented accuracy–efficiency trade-off. After INT8 conversion using the standard RKNN-Toolkit2 workflow, ECEM-YOLO-S reaches 10.13 FPS on the RK3576 NPU, approximately 21 × faster than CPU-only inference. Codes and dataset are available at <a href="https://github.com/zdpf122/ECEM-YOLO-RoadDefect">https://github.com/zdpf122/ECEM-YOLO-RoadDefect</a> and <a href="https://doi.org/10.5281/zenodo.17792244">https://doi.org/10.5281/zenodo.17792244</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

Enhanced road infrastructure defect detection via attention-based visual inspection

  • 期刊论文

  • 发表日期：

    2026年7月29日

查看全文
  • Ningbo Gao,
  • Dupengfei Zhai,
  • Guifang Shi,
  • Guanya Hao,
  • Yong Qi,
  • Zi Yang

摘要

Timely detection of road infrastructure defects is important for traffic safety, maintenance efficiency, and transportation system resilience. While existing methods can detect small targets such as road cracks, challenges remain in low-contrast crack recognition, small-target detection, multi-category road facility defect identification, and edge-device deployment. This study introduces a comprehensive road defect detection system deployed on an edge device with a Rockchip RK3576 chip, enabling near-real-time inference and automatic result uploading to cloud platforms. To address the difficulty of detecting low-contrast and crack-like defects, we design a task-oriented lightweight crack enhancement block, ECEM, within the YOLOv11 framework. Unlike generic attention-based YOLO modifications, ECEM integrates multi-scale depth-wise feature extraction, attention refinement, and background suppression to enhance elongated crack responses while maintaining lightweight computation. Evaluated on the RFDD test set, the S-scale variant ECEM-YOLO-S achieves 73.0% Precision, 67.3% mAP50, and 895.60 ± 1.97 FPS on an NVIDIA RTX 4090 under the unified Ultralytics validation pipeline. Compared with YOLOv11-S, it improves Precision and mAP50, with a slight decrease in Recall, reflecting a precision-oriented accuracy–efficiency trade-off. After INT8 conversion using the standard RKNN-Toolkit2 workflow, ECEM-YOLO-S reaches 10.13 FPS on the RK3576 NPU, approximately 21 × faster than CPU-only inference. Codes and dataset are available at https://github.com/zdpf122/ECEM-YOLO-RoadDefect and https://doi.org/10.5281/zenodo.17792244.