A CNN-Based Automatic Detection Method for Tunnel Lining Internal Defect Using Ground Penetrating Radar

Regular tunnel lining internal damage (TLID) inspection is very crucial for the railroad operation safety. However, traditional manual identification methods using ground penetrating radar (GPR) images suffer from limited accuracy and efficiency. Fortunately, CNN-based automatic inspection methods have the potential to replace manual inspections, offering improved precision and productivity. In this paper, we propose a CNN-based detection model that builds upon the improved YOLACT framework for identifying defects in tunnel linings. To enhance feature extraction, we employ the Res2Net architecture as the backbone of our model. The overall architecture includes the FPN feature pyramid, prediction heads, protonet branch, as well as assembly and crop processes. Through experimentation, we compare the performance of our model when utilizing different backbone networks, namely Res2Net50, Res2Net101, ResNet50, and ResNet101. Our experiments, conducted on a GPR dataset, demonstrate that the improved YOLACT with Res2Net50 achieves a high detection rate and exhibits good adaptation capabilities for identifying internal damage in complex tunnel environments.