Identification and 3D reconstruction of semi-rigid base loose distress from GPR B-scan using Generative Adversarial Network

Abstract

Ground Penetrating Radar (GPR) is widely utilized in detecting subsurface distress. However, its identification and analysis still face challenges. We propose a method based on the Generative Adversarial Network (GAN) to process GPR B-scan data containing semi-rigid base loose distress. An end-to-end GAN invert GPR B-scan into cross-sectional images of the road. A U-net is used in the generator to transform from GPR B-scan to road loose models. A patchGAN is used in the discriminator to determine the correlation between the GPR B-scan and road loose models. We used Finite-Difference Time-Domain (FDTD) and random mediums to construct the road loose distress model. Based on this model, the simulated dataset of 14,000 sets of images was randomly generated. Post-processing of the simulated dataset generated the synthetic dataset of 14,000 sets of images. The identification results trained on the simulated dataset and synthetic dataset achieved 90 % and 97 % average Structural Similarity Index (SSIM) compared to the source images. Through threshold segmentation of the generated images, 3D models are reconstructed using Marching Cubes (MC). Validation with the actual project indicates that this method effectively recognizes loose distress, and the generated 3D distribution model accurately represents the road condition. This approach offers a promising solution to the challenge of radar image identification and introduces a new data inversion method for road nondestructive testing.