Abstract
To address the challenges of low accuracy and limited real-time efficiency in detecting subsurface defects within concrete structures, this study proposes an enhanced YOLOv5 model integrated with an Efficient Channel Attention (ECA) mechanism for automated ground-penetrating radar (GPR) defect detection. A Deep Convolutional Generative Adversarial Network (DCGAN)-based augmentation strategy is introduced to mitigate class imbalance, synthesizing realistic minority-class defect samples while preserving wave scattering characteristics. A specialized dataset encompassing diverse defect types was constructed to reflect real-world concrete inspection scenarios. The proposed YOLOv5 + ECA model was rigorously evaluated against other attention-enhanced variants and the baseline YOLOv5. Experimental results demonstrate that ECA's channel-specific feature recalibration significantly improves detection accuracy, achieving the highest mean average precision, while maintaining real-time inference speeds suitable for unmanned aerial vehicle (UAV)-mounted deployment. This work advances the precision and efficiency of infrastructure health monitoring, offering a robust solution for subsurface defect diagnosis in concrete structures such as tunnel linings and bridge decks.