DSLCD-based Tunnel Secondary Lining Cavities Detection in Ground Penetrating Radar Images

Because the depth is large, the detection area is wide, the interference is much, and the discrimination depends on expert experience, it is difficult to detect cavities inside the secondary lining of the tunnel efficiently. In order to solve this problem, we propose an end-to-end improved Faster R-CNN based network model named Deep Secondary Lining Cavity Detection (DSLCD) network. The model is divided into three parts: the feature extraction network, the region proposal network, and the classification and regression network. In the feature extraction network, a Balanced Feature Pyramid is used to strengthen the multi-layer features, while in the region suggestion network, the Guided Anchoring method is used to improve the model’s adaptability to the various size of the secondary lining voids target and to enhance the accuracy of the secondary lining voids target positioning by optimizing the ROI alignment process. The experimental results show that the DSLCD network has a decent detection accuracy on our GPR image dataset containing both reinforced concrete and plain concrete, which meets the engineering application requirements.