Road sub-surface defect detection based on gprMax forward simulation-sample generation and Swin Transformer-YOLOX

Abstract

Training samples for deep learning networks are typically obtained through various field experiments, which require significant manpower, resource and time consumption. However, it is possible to utilize simulated data to augment the training samples. In this paper, by comparing the actual experimental model with the simulated model generated by the gprMax [1] forward simulation method, the feasibility of obtaining simulated samples through gprMax simulation is validated. Subsequently, the samples generated by gprMax forward simulation are used for training the network to detect objects in existing real samples. At the same time, aiming at the detection and intelligent recognition of road sub-surface defects, the Swin-YOLOX algorithm is introduced, and the excellence of the detection network, which is improved by augmenting the simulated samples with real samples, is further verified. By comparing the prediction performance of the object detection models, it is observed that the model trained with mixed samples achieved a recall of 94.74% and a mean average precision (mAP) of 97.71%, surpassing the model trained only on real samples by 12.95% and 15.64%, respectively. The feasibility and excellence of training the model with mixed samples are confirmed. The potential of using a fusion of simulated and existing real samples instead of repeatedly acquiring new real samples by field experiment is demonstrated by this study, thereby improving detection efficiency, saving resources, and providing a new approach to the problem of multiple interpretations in ground penetrating radar (GPR) data.