Direction-Aware and Expert-Inspired Learning for Internal Crack Size Detection Using On-Site GPR Data

Internal cracking in asphalt pavements develops beneath the surface and can rapidly propagate upward, threatening structural integrity and traffic safety. Crack size, especially top width, bottom width, and depth, is a key parameter for selecting maintenance strategies (e.g., grouting positioning and repaving decisions). Ground-penetrating radar (GPR) enables non-destructive subsurface inspection, yet practical crack size interpretation remains challenging due to (i) limited robustness when transferring signal–size relationships from simulations to heterogeneous field conditions, and (ii) the difficulty of directly characterizing crack size from raw GPR B-scan image features.

This study proposes an internal crack size detection network (ICSD-Net) trained on on-site GPR B-scans with interpreted crack size labels. The method targets the trapezoidal geometry of internal cracks (narrow top, wider bottom) and the fact that size-relevant information is concentrated near the hyperbolic apex of crack reflections, where confounding layer reflections often exist and conventional anchor-based/anchor-free detectors struggle with positive-sample matching.

ICSD-Net integrates three key designs. First, a deformable Cross Stage Partial (CSP) backbone improves geometric adaptability for irregular hyperbolic reflections. Second, a Directional Fusion Attention Module (DFAM) constructs direction-aware channel attention using 1D pooling along height/width and generates spatial interaction weights via directional feature broadcasting and multiplicative fusion, enhancing modeling of long-range dependencies across both sides of a hyperbola while suppressing background clutter. Third, an expert-inspired Bipartite Matching (BM) head adopts a DETR-like global set prediction strategy: the network outputs a fixed number of trapezoidal size candidates and uses Hungarian matching to select the optimal assignment between predictions and ground truth, emulating expert global reasoning on an entire B-scan.

A field dataset was built using a 3D GPR array system (24 channels, 800 MHz) from a highway rehabilitation project; signals were minimally processed (direct-wave removal and normalization). Crack size labels were derived by combining forward-model-informed relationships between reflection amplitude and crack top/bottom widths (high correlation reported) with travel-time-based depth estimation, then annotated as four-corner trapezoids on B-scans. The dataset contains 1968 labeled B-scans (small/medium/large targets) split into train/validation/test at 7:2:1.

Experiments show ICSD-Net outperforms multiple state-of-the-art baselines (including YOLO pose variants and DETR adaptations), achieving the highest mAP and mIoU with approximately 8-12% mAP improvement over the strongest baseline, while maintaining real-time feasibility. Ablation studies indicate that DFAM and the BM head contribute most to accuracy gains, improving attention focus toward the hyperbolic apex and reducing misdetections caused by layer reflections. Stability tests demonstrate consistent performance across antenna frequencies and pavement structures, supporting practical deployment. Field validation using coring measurements indicates predicted crack sizes generally meet engineering requirements, with remaining difficulty in accurately estimating bottom width for water-saturated and small cracks due to strong dielectric-contrast-induced multiple reflections and deeper-layer noise.