Lightweight pavement crack detection model for edge computing devices.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Infrastructure Maintenance

Background:

• Road surface crack detection faces challenges like false positives, missed detections, and large model sizes.
• Existing models struggle with complex backgrounds and varying crack shapes.

Purpose of the Study:

• To propose a lightweight and accurate road crack detection model, YOLO-DGVG.
• To improve detection performance while reducing computational complexity and parameter volume.

Main Methods:

• Introduced deformable convolution (DCNv2) into the backbone network (C2f-DCNv2) for adaptive shape adjustment.
• Integrated lightweight GSConv and VoVGSCSP modules in the neck network to enhance feature extraction and reduce complexity.
• Developed a grouped convolution detection head (GCH) to minimize parameter volume.

Main Results:

• Ablation studies on the PID dataset showed increased Recall (0.3-1.6%) and mAP, with a 22.28% reduction in parameters compared to YOLOv8.
• Generalization experiments on UAPD, RDD2022, and custom datasets confirmed effectiveness.
• YOLO-DGVG outperformed RT-DETR and YOLOv10 in detection tasks.
• Successful deployment on edge devices for static image crack detection.

Conclusions:

• YOLO-DGVG offers a significant improvement in road crack detection accuracy and efficiency.
• The model's lightweight design and enhanced modules make it suitable for real-world applications, including edge computing.