Currently, the manual method of annotating deep learning samples for Ground Penetrating Radar (GPR) with open-source tools such as LabelImg and Labelme is not only time-consuming and labor-intensive, but it also annotates images rather than radar data. This fails to satisfy the requirements of deep learning for large sample sizes and hinders the sharing and reuse of GPR data. It is essential to design a unified data storage format for the manual interpretation results of GPR data, establish a mapping relationship between hidden road defects and their GPR data, and enable autonomous retrieval, positioning, cropping, and automatic annotation of GPR data samples. Based on the YOLO network model, this study has developed an intelligent diagnostic software system for GPR images pertaining to hidden road defects. This system can automatically annotate GPR sample data pertaining to hidden road defects and implement methods such as adaptive gain adjustment, digital filtering, automatic zero drift removal, and background subtraction to enhance radar sample data, generating deep learning samples with different signal characteristics. Through a comparative analysis of the deep learning training performance of the YOLOv8n and YOLOv8x models on GPR samples pertaining to hidden road defects, a manual verification method for intelligent diagnostic results has been developed. The testing results of the algorithm and software reveal that automatic annotation and data enhancement of GPR data pertaining to hidden road defects can significantly expedite the generation speed of GPR deep learning samples and enrich the diversity of such samples. Compared with YOLOv8n, the YOLOv8x model achieves smaller training losses, higher training accuracy, and is more suited for intelligent diagnosis of GPR images pertaining to hidden road defects.