Structure-Aware Conditional GAN for High-Fidelity GPR B-Scan Image Synthesis in Airport Runway Defect Detection
Year: 2026
Type: article
Abstract: Airport runways are critical transportation infrastructures whose structural integrity is continuously threatened by subsurface defects induced by cyclic aircraft loading and environmental erosion. Ground-penetrating radar (GPR) B-scan imaging is widely employed for subsurface inspection, yet deep learning-based interpretation is severely constrained by the scarcity of high-quality labeled data. Synthetic data generation offers a promising remedy, but existing GPR generators typically operate at a single feature scale and lack defect-category awareness, so that weak reflections of subtle defects are easily suppressed while large-scale defects are rendered as fragmented or structurally implausible patterns. To address these limitations, we propose SA-CGAN, a structure-aware conditional generative adversarial network that synthesizes high-fidelity, defect-specific GPR B-scan images for multiple defect types within a unified conditional framework, jointly guided by a defect-class condition and a spatially aligned bounding-box mask. Two complementary attention modules are embedded into the generator to close the fidelity gap from opposite ends. A Class-aware Local Structure Enhancement Attention (CLSEA) amplifies weak directional reflections of subtle defects by coupling a parameter-free multi-directional energy prior with class-conditional modulation under mask guidance, whereas a Structure-Localized Deformable Enhancement Attention (SLDEA) preserves the spatial coherence of large-scale, irregularly shaped defects via mask-conditioned deformable sampling and gated residual refinement. Experiments on real airport runway GPR data collected from 11 runways show that SA-CGAN achieves favorable FID, KID, and mFSIM scores compared with competitive generative baselines, and that augmenting real data with SA-CGAN-synthesized samples consistently improves downstream YOLOv10s detection, raising the F1-score to 0.94, 0.86, 0.93, and 0.93 on cracks, voids, looseness, and subsidence, respectively.
Source: Electronics
Authors: Nansha Li, Xinyu Liu, Zhongcheng Gui
Institution: Civil Aviation University of China
Language: English
Subfield: Geophysics
Field: Earth and Planetary Sciences
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7225832003
FWCI:
Cites:
Cited by:
Related to: