Ground-penetrating radar (GPR) imaging is widely used for detecting hidden defects in urban roads. However, the complex noise environment, large-scale variations in defect features, and the sensitivity of slender defects to annotation errors pose significant challenges to accurate detection. To address these issues, this study proposes an improved object detection framework, termed DFF-MoCA-YOLO, based on YOLOv11 for identifying void and loose defects in GPR images. First, a multi-strategy gated feature fusion module (MSGFF-C3k2) is designed to enhance feature robustness against complex noise and scale variations. Then, a Monte Carlo Attention (MoCAttention) module is introduced to improve defect-feature representation via stochastic sampling and channel recalibration. Subsequently, an adaptive aspect-ratio penalty CIoU loss (CIoU-ARP) is developed to improve bounding box regression accuracy for slender defects. A labeled dataset containing void and loose defects is constructed using multi-source GPR data collected from eight urban roads. Finally, a series of ablation experiments is conducted on the proposed modules. Experimental results demonstrate that the proposed method achieves consistent performance improvements over the baseline YOLOv11 and other mainstream YOLO variants, while maintaining relatively low computational complexity. The results indicate that the proposed framework offers an effective and practical solution for detecting hidden defects in urban roads using GPR images. Moreover, the model’s robustness to noise and ability to accurately detect defects at varying scales make it a promising tool for urban infrastructure maintenance. Its efficient performance with minimal computational overhead makes it suitable for real-time defect detection.
Chen et al. (Thu,) studied this question.