Abstract:To address the challenges of low accuracy, difficulty in recognizing hyperbolic features of structural anomalies, and high computational complexity in Ground Penetrating Radar (GPR) image detection for urban underground exploration, this paper proposes a lightweight detection model, RBD-YOLOv11, based on an improved YOLOv11n. A high-quality dataset encompassing metal pipes, PVC pipes, and rebar meshes is constructed by utilizing GprMax for simulation and CycleGAN for domain transfer. The proposed method introduces three key improvements: First, RepViT is adopted as a lightweight backbone network to significantly reduce model parameters and computational load. Second, the BiFPN (Bidirectional Feature Pyramid Network) and SCDown (Spatial-Channel Decoupled Downsampling) modules are integrated into the Neck architecture to enhance multi-scale feature fusion capabilities while lowering computational costs. Finally, a lightweight dynamic upsampler, Dysample, is employed to achieve effective recognition of structural anomaly hyperbolas under complex backgrounds. Experimental results demonstrate that the proposed method achieves high recognition accuracy with a lightweight architecture, exhibiting significant advantages in the field of GPR image category detection.