Lightweight multiscale deep learning for pavement defect detection using ground penetrating radar

Ground-penetrating radar (GPR) is a valuable method for assessing road infrastructure without causing damage, helping quickly identify and locate potential pavement defects. As GPR acquisition shifts toward vehicle-mounted platforms for large-scale surveys, automated interpretation methods are required for lightweight deployment and real-time inference, while remaining robust to complex background noise and multi-scale defects. This study develops MR-YOLOv5s, a lightweight model for interpreting GPR B-scan images. Leveraging 1,038 SAGAN-augmented images, the model integrates self-attention mechanisms and lightweight techniques to reduce computational load while maintaining multi-scale feature extraction. Results show that MR-YOLOv5s achieves an inference speed of 105.2 FPS and a compact size of 10.9 MB, satisfying vehicle-mounted edge deployment requirements. Additionally, the model reaches 96.1% mAP by effectively capturing defect waveforms and contextual features, ensuring reliable detection and robust generalisation across complex pavement defects. This approach offers a versatile and efficient solution for large-scale infrastructure management.