MARD-Net: Fusing Multi-Path Attention, Re-Parameterization, and Asymmetric Detection Head for Subsurface Road Defect Detection in GPR Imagery

Timely detection of subsurface road defects is critical for structural safety and pavement longevity. Here, we propose MARD-Net, an enhanced deep learning framework designed for the accurate identification of urban subsurface road defects. First, addressing the scarcity of defect samples, a hybrid dataset was constructed by integrating empirical data acquired via the GS8000 ground-penetrating radar (GPR) system and synthetic data generated by gprMax. Second, to address complex geological backgrounds and variations in defect waveform scales, the RepNCSPELAN4_CAA module was integrated into the architecture. By combining structural re-parameterization with context anchor attention (CAA), this module enhances feature reuse and inter-channel interaction, enabling the fine-grained discrimination of subtle defects. Third, a lightweight asymmetric detection head (LADH), incorporating depthwise separable convolution (DSConv) within its regression branch, was developed to significantly reduce computational costs while maintaining robust detection performance. Finally, to overcome weak and uneven defect signals across imaging depths, a multi-path coordinate attention (MPCA) mechanism adaptively fuses global and local contextual information for precise defect recognition. Empirical experiments show that MARD-Net achieves a 2.07% increase in mean average precision at an intersection over union threshold of 0.50 (mAP@50) over the baseline YOLOv11n while reducing floating point operations (FLOPs) by 2.0 giga floating point operations (GFLOPs).