Enhancing GPR-based tunnel lining delamination detection: An unsupervised domain-adaptive YOLO network

The ground penetrating radar (GPR) is one of the most recommended methods for tunnel lining inspection, yet the interpreting the GPR data requires significant time and expertise. Recent attempts to automate this procedure using deep learning techniques have made significant progress, yet is still hindered by the need of extensive and accurately annotated data, which is challenging to gather in real scenarios. This paper introduces a domain-adaptive YOLO network, evolving from YOLOv7, which trains on FDTD data for source domain detection and real-world data for target domain adaption. The proposed network enhances the network’s ability to generate generalized features across domains and hence improves the detection performance on real GPR data.