Road Underground Target Detection in Ground Penetrating Radar Images Based on YOLO Models

Ground penetrating radar (GPR) is extensively employed for subsurface road target detection, offering benefits such as convenience, nondestructive testing, rapid data acquisition, and superior resolution. Despite these advantages, interpreting GPR data often depends on the expertise of professionals, resulting in low detection efficiency and low accuracy. To address these challenges, this study introduces an intelligent detection technique for GPR images, utilizing an enhanced YOLOv5 framework. First, considering the problems of the small amount of GPR image datasets and the unclear characteristics caused by the complex underground media, a Dense-C3 module is built by utilizing the structure of DenseNet to enhance the network's capability for extracting features. Subsequently, a channel and spatial hybrid attention module is introduced into the backbone for feature refinement and improving the efficiency. Finally, the multi-class focal loss function is devised to enhance the precision in cases of imbalanced sample classes. Experimental results show that the proposed model surpasses the original YOLOv5 model and various contemporary advanced models.