ABSTRACT
Maintenance of tunnel structure involves investigation of tunnel wall with Ground Penetration Radar (GPR). By analyzing GPR investigation images, engineers determine the locations of steel ribs, thickness of concrete lining and the location and size of void if any. The identification of defects, however, requires experience and also is a time-consuming work. YOLO (You Only Look Once) and Mask R-CNN are CNN (Convolutional Neural Network) algorithms with single-stage detection and two stage detection, respectively. In this study, a model based on YOLO was trained to detect the steel ribs in GPR images and Mask R-CNN was trained to detect the concrete lining, and tested. Overall, both of the CNN algorithms provided satisfactory results; however, the images with severe scattering resulted in lower accuracy.
