Intelligent Identification of Hidden Defects in Asphalt Roads Using GPR Based on Loss Function and Anchor Box Optimization

Ground penetrating radar (GPR) is a critical tool for identifying hidden defects in asphalt pavements. This study focuses on the automated detection of asphalt road defects, including hollow defects and defects between layers, using optimized YOLOv5 models applied to GPR images. The influence of different feature scale detection layers on the identification of hidden defects in asphalt roads is systematically evaluated. Ten models, integrating the efficient intersection over union loss function and the K-means++ clustering algorithm, are compared for their performance. The results demonstrate that the YOLOv5m model outperforms others during training, achieving the highest mean at [email protected] and AP values of 0.841 and 0.748 for hollow defects and defects between layers, respectively. In contrast, the YOLOv5l6 model exhibits the lowest detection accuracy. Notably, the findings reveal that increased model complexity does not necessarily enhance detection accuracy. Furthermore, in the context of asphalt road-defect detection, three-scale feature detection layers show superior applicability compared to four-scale feature detection layers.