Аннотация
Ground-penetrating radar (GPR) is widely used for detecting subsurface defects in airport runways. However, GPR data is often noisy, complex and inconsistent due to different subsurface structures and environmental conditions across airports. These factors pose serious challenges to existing detection models, as similar features across different defect types and diverse patterns within the same type make it hard to learn stable and discriminative representations. To address these issues, this study proposes a multimodal detection framework, Text-Image Guided Network (TIGNet), which integrates GPR image with subsurface layer information and textual semantics, enhancing both feature learning and target discrimination. Furthermore, a learnable text embedding mechanism is introduced, enabling the model to adaptively refine textual features during training, rather than relying on manually designed templates. Experiments on data collected from eleven airports demonstrate that TIGNet achieves superior performance over state-of-the-art methods in detection accuracy, false positive reduction, and cross-domain generalization. Specifically, our method achieves F1 score at 89%, 82%, 90%, and 92% for four types of subsurface features (i.e. gap, crack, subsidence and rebar), respectively, which indicate strong application potential in runway inspection.