Arbitrarily-oriented tunnel lining defects detection from Ground Penetrating Radar images using deep Convolutional Neural networks

Abstract Tunnel lining internal defect detection is essential for the safe operation of tunnels. This paper presents an automatic scheme based on rotational region deformable convolutional neural network (R2DCNN) and Ground Penetrating Radar (GPR) images for the accurate detection of defects and rebars with arbitrary orientations. The R2DCNN comprises inter-related modules, specifically, deformable convolution, feature fusion, and rotated region detection modules. In this study, synthetic GPR images, including rebars and various structural defects with different permittivities, as well as real GPR images obtained by model experiments, were constructed for the R2DCNN. Improved results were obtained while testing the R2DCNN on GPR images in comparative experiments. The mean average precision of the R2DCNN was enhanced by 8.21% compared to the R2CNN on synthetic GPR images. The R2DCNN showed satisfactory results in on-site experiments, which demonstrated the applicability of the R2DCNN to practical tunnels.