Haochuan Zhou, Fanwen Meng, Jiaqi Li, Weiguang Zhang, Zheng Tong
Structural cracks that develop within semi-rigid asphalt pavement structures may remain concealed beneath the pavement surface, making their opening widths, depths, and morphologies difficult to determine nondestructively. Although ground-penetrating radar (GPR) can localize subsurface reflective cracks, the hyperbolic anomalies in a B-scan, a two-dimensional cross-sectional radar profile obtained from sequential measurements along a survey line, do not directly represent the actual geometrical dimensions of the cracks. To solve this problem, this paper proposes a geometrical dimension detection and morphological inversion (GDMI) method, called the GPR-GDMI. The GPR-GDMI comprises two independently trained networks: GPR-GCDNet for geometrical dimensions detection and GPR-CMIGAN for morphology inversion. GPR-GCDNet utilizes trapezoidal boxes to represent crack dimensions in B-scan, enhanced by directional crack attention (DCA) for hyperbolic pattern extraction, a SetTrap Head for trapezoidal proposal generation, and multi-scale feature fusion for improved receptive field and coverage. GPR-CMIGAN is an unsupervised framework consisting of two independently designed U-Net generators and discriminators, with Gaussian noise injected to prevent mode collapse. The crack-dimension-constrained cycle-consistency loss incorporates GPR-GCDNet's detected dimensions as physical constraints in the inversion, while Wasserstein loss mitigates over-constraints and improves stability during training. On the test set, GPR-GCDNet achieved an AP of 0.84, an F1-score of 0.89, and an mIoU of 0.74 under an IoU threshold of 0.5. GPR-CMIGAN achieved a performance on MS-SSIM, LPIPS, and FID of 0.9984, 0.0566, and 2.3145, respectively. Ablation studies confirm the contribution of each component. By quantitative experiments, GPR-GDMI achieves accurate detection of crack geometrical dimensions and morphological inversion based on field-collected GPR data.