A Two-Stage Hyperbola Localization Method in GPR Images Based on Frequency-Domain Feature Aggregation

The hyperbola detection in the ground-penetrating radar (GPR) data is of real significance for subsurface object localization. However, current detection methods are difficult to achieve both accuracy and efficiency. While existing two-stage methods rely on computationally intensive region proposal networks or complex preprocessing pipelines, we propose a novel Two-stage GPR Signal Localization (TGSL) framework that achieves precise localization through a lightweight progressive refinement strategy. Unlike conventional two-stage approaches that rely on rectangular bounding boxes, TGSL employs a coarse detection–fine localization cascade architecture based on YOLO11. In the first stage, the improved YOLO11 identifies potential target regions in GPR images. In the second stage, instead of traditional classification refinement, TGSL performs direct keypoint regression on cropped regions using a specialized YOLO11 equipped with Frequency-Driven Feature Aggregation (FDFA) module. This unique approach leverages shared backbone weights between stages and focused processing on target-relevant regions, achieving both computational efficiency and high precision. The framework accurately detects five critical keypoints of GPR signal hyperbola. Experimental results on tunnel lining detection demonstrate that TGSL achieves 92.7% mAP50 outperforming both traditional two-stage methods in speed and single-stage methods in accuracy.