Addressing Data Scarcity in GPR Road Defect Detection: A Novel Framework Combining Stable Diffusion and Efficient GCP-YOLO

The application of deep learning to ground penetrating radar (GPR) detection of urban road subsurface defects is often limited by the scarcity of real-world data and high computational costs. To address this, we propose a novel framework that integrates Stable Diffusion for data augmentation, constructing mixed data sets of real and synthetically generated GPR B-scans. This approach effectively mitigated data scarcity and reduced reliance on time-consuming field measurements, resulting in significant gains in detection precision—with improvements in mean average precision ranging from 5.5% to 18.2%. Furthermore, this study conducts a thorough analysis of the time costs associated with data set construction and model training, offering practical guidelines for balancing precision and efficiency. Several enhanced YOLOv8 models were tailored for this task. Among them, GCP (2)-YOLO achieved the highest detection precision with a higher training time cost, while GC-YOLO and GCP (6)-YOLO attained a better trade-off.