YOLO 및 회귀 모델을 활용한 GPR B-scan 영상 기반 공동 크기 및 위치 추정

Abstract

Ground Penetrating Radar (GPR) is a non-destructive technique used to detect subsurface voids. In this studypaper, synthetic B-scan data hiddenwithwith spherical voids were generated using gprMax, and a YOLOv8-s model was trained to detect hyperbolic patterns. Since YOLO provides bounding boxes in pixel-level coordinates, regression mapping to ground-truth (GT) labels was applied to enable real-world size estimation. The predicted bounding boxes were mapped to Ground Truth (GT) labels through regression to evaluate void size inference.Since YOLO provides bounding boxes in pixel-level coordinates, a regression-based mapping was developed to convert detection results into real-world metric (m) coordinates, enabling direct estimation of void size and position. Results show that the predicted center coordinates (X, Y) exhibited linearity with GT values (R²>0.99). The bounding box height correlated with GT (R²=0.98). Also, the bounding box width and depth (Y), the estimation of GT width shows (R²≈ 0.98). These results demonstrate that YOLOv8-s based detection can reliably support both void localization and quantitative size estimation in GPR B-scan images.These results demonstrate that the proposed method can reliably support both void localization and real-size estimation in GPR B-scan images. Thus, the proposed approach provides an automated pixel-to-meter translation framework for quantitative GPR interpretation.