Ground penetrating radar signal recognition and localization method based on improved Yolov8

This paper proposes an improved Yolov8 network (ADC-Yolov8) for ground penetrating radar (GPR) signal recognition and localization of buried pipelines. The method addresses the challenges in processing GPR images with multiple targets, small targets, and weak textures. The key innovations include: (1) constructing an efficient feature fusion structure using the ADown module; (2) enhancing multi-scale features through the DualConv_C2f module with channel attention; and (3) improving complex feature extraction capability by embedding the CA mechanism of HS-FPN. Experimental results demonstrate that the improved model achieves 3.5% and 1.5% higher Precision and mAP@0.5 respectively compared to the original Yolov8, while reducing FLOPs by 24% and parameter quantity by 50%. These advancements enable the proposed method to deliver superior performance in complex underground environments, particularly for detecting small and weakly textured targets, where it shows a remarkable 0.4% accuracy improvement (from 93.1% to 97.1%).