Research on asphalt road cavity defects detection and data annotation methods based on GPR

Accurate detection of road cavity defects is a crucial part of infrastructure maintenance. It is of great significance to optimize the quality of data annotation to improve the detection performance of the model. This research has solved three key challenges in asphalt pavement cavity detection: selection of single-stage model architecture, optimization of pre-training strategies, and design of annotation methods. Based on the 3D GPR equipment, this research constructed a dataset of asphalt road cavity defects. Under the condition of controlling hyperparameters, by comparing YOLOv5 − v12 series models, it was found that AP of YOLOv5 was best. The experiment involving nine different training strategies revealed that models without using pre-trained weights achieved a 0.9% improvement in AP for cavity defect detection. Based on characteristic region partitioning of cavity defects, the research proposed a new annotation method named GPR-HITBZ. Experiments demonstrated that excluding the bottom characteristic regions of cavity defects from bounding boxes could improve the F1-score by approximately 30%. The GPR-HITBZ achieved the best results in both AP and F1 score among the eight annotation methods, providing a standardized annotation paradigm for detection of hidden defects on roads.