GPR image denoising based on unpaired data: Enhancing defect detection inside tunnels by TID-CycleGAN

Abstract

Detection of defects in tunnel linings is essential for ensuring structural safety and long-term stability. However, noise in ground penetrating radar (GPR) images significantly reduces the accuracy of defect detection. In this study, we propose a migration denoising model, TID-CycleGAN, based on an unpaired dataset. By incorporating the enhanced skip connection module (ESCM), the model efficiently fuses spatial and frequency domain information while preserving defect details during the denoising process. Experimental results indicate that TID-CycleGAN outperforms other models in peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM), notably improving the visual quality of radar images. Furthermore, by optimizing the mixing ratio of denoised and real images, a high-quality defect detection dataset is developed, achieving an optimal balance between background noise and defect information. Detection comparison experiments demonstrate that cutting-edge models such as YOLOv11 achieve average accuracies exceeding 90 %, meeting the requirements for tunnel inspection tasks and enabling automated detection of defects in tunnel linings.