Research and optimization of YOLO-based method for automatic pavement defect detection

Abstract

<abstract> <p>According to the latest statistics at the end of 2022, the total length of highways in China has reached 5.3548 million kilometers, with a maintenance mileage of 5.3503 million kilometers, accounting for 99.9% of the total maintenance coverage. Relying on inefficient manual pavement detection methods is difficult to meet the needs of large-scale detection. To tackle this issue, experiments were conducted to explore deep learning-based intelligent identification models, leveraging pavement distress data as the fundamental basis. The dataset encompasses pavement micro-cracks, which hold particular significance for the purpose of pavement preventive maintenance. The two-stage model Faster R-CNN achieved a mean average precision (mAP) of 0.938, which surpassed the one-stage object detection algorithms YOLOv5 (mAP: 0.91) and YOLOv7 (mAP: 0.932). To balance model weight and detection performance, this study proposes a YOLO-based optimization method on the basis of YOLOv5. This method achieves comparable detection performance (mAP: 0.93) to that of two-stage detectors, while exhibiting only a minimal increase in the number of parameters. Overall, the two-stage model demonstrated excellent detection performance when using a residual network (ResNet) as the backbone, whereas the YOLO algorithm of the one-stage detection model proved to be more suitable for practical engineering applications.</p> </abstract>