Enhanced Subsurface Defect Detection via Context-Guided and Lightweight Convolutional Model

Efficient detection of subsurface road defects is crucial for ensuring roadway safety and facilitating intelligent maintenance. Traditional models often struggle with feature extraction accuracy, particularly under complex radar signal conditions. This study introduces LUCGD, a transfer learning-enhanced model that addresses these challenges through structural and algorithmic innovations. By initializing the backbone with YOLOv11n pretrained weights, we reduce training complexity. A hybrid dataset combining GPRMax simulations and GS8000 radar measurements is constructed to overcome limited GPR data. The model incorporates a Lightweight Asymmetric Detection Head (LADH) for precision and computational efficiency, and a Context-Guided Convolution (ContextGuidedConv) within the C3k2 block for enhanced feature discrimination. An Adaptive Downsampling (ADown) module preserves high-frequency details, improving sensitivity to small or blurred defects. LUCGD achieves a 0.0373 increase in mAP@50 over YOLOv11n while reducing computational cost by 3 GFLOPs. The source code will be publicly available at https://github.com/zwb-ong/LUCGD-The-Visual-Computer.