Ground Penetrating Radar Image Analysis for Underground Barrier Detection by Combining YOLOv12 with Channel-wise Attention and Denoising Auto-Encoder

Le résumé fourni par la source

Accurate detection of underground barriers such as pipelines is crucial for urban safety and infrastructure management. Ground Penetrating Radar (GPR) image provides a non-destructive means for subsurface exploration, but its B-scan images often contain strong noise and clutter that hinder reliable recognition. To address these challenges, we propose a YOLOv12-based detection framework enhanced with a denoising autoencoder (AE) and channel-wise attention (CBAM). The AE suppresses noise while preserving hyperbolic signatures, and CBAM adaptively highlights informative features, which improves robustness under complex soil conditions. Experiments on real GPR datasets of gas pipelines show that our method achieves higher precision, recall, and mAP@50 than baseline YOLOv12. Efficiency analysis further reveals that the CBAM-enhanced variant offers the best accuracy-training trade-off, while the combined AE+CBAM model provides the most balanced performance. These results demonstrate the effectiveness of integrating denoising and attention mechanisms into modern detectors for robust underground barrier detection.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.