Highways are prone to developing cracks, potholes, sinkholes, and other sur-face and structural layer hidden diseases due to the combined effects of traffic loads, weather conditions, and environmental influences during their ser-vice life. GPR is a widely employed NDT technique for detecting such concealed defects in pavement structures. However, in practical detection applications, GPR mapping relies heavily on experience for recognition and analysis, resulting in low recognition efficiency. This study collected disease map-ping data at a GPR standardized disease test road and utilized simulation analysis software to generate ortho maps. These maps were used to construct an automatic identification dataset containing 900 maps. Three target recognition models, SSD, Faster R-CNN, and YOLO v5, were employed to analyze the dataset and compare their effectiveness. The results revealed that YOLO v5 model had the best application effect, with an average recognition accuracy of 90.45%. The YOLO v5 model's backbone was improved based on Swin-Transformer and FE Module, leading to significant improvements in recognition. The improved model achieved an average recognition accuracy of 94.21%, demonstrating its superior performance compared to the other two models.