Comparative Study of Deep Learning Models in GPR-Based Structural Health Monitoring

초록

Structural Health Monitoring (SHM) plays a critical role in ensuring the safety and longevity of civil infrastructure by supporting the interpretation of subsurface conditions such as embedded reinforcement, delamination, voids, and other anomalies. Among non-destructive testing techniques, Ground Penetrating Radar (GPR) is widely used for subsurface inspection, where embedded objects and anomaly-related responses often appear as hyperbolic patterns in B-scan images. However, accurate interpretation of GPR data remains challenging because it depends heavily on expert knowledge and labor intensive manual analysis. This study presents a comparative analysis of three object detection models YOLOv8, YOLO11, and Faster R-CNN for automatic hyperbola detection in GPR B-scan images. The experiments were conducted using a GPR B-scan object detection dataset consisting of 512 images and 6,138 labeled instances. The dataset was divided into 328 training images and 184 validation images, and the three models were independently trained and evaluated under the same class definition. The evaluated classes were noise, hyperbola, and hyperbola line. The results show that YOLOv8 achieved the best overall AP based detection performance, with an mAP50 of 0.824 and an mAP50–95 of 0.503. YOLO11 achieved the highest precision and recall, with values of 0.754 and 0.770, respectively, while also providing the fastest inference speed. Faster R-CNN trained for 15,000 iterations showed stable loss convergence but achieved lower validation performance, with an mAP50 of 0.468 and an mAP50–95 of 0.262. These results indicate that YOLO-based one stage detectors are more suitable than the evaluated Faster R-CNN baseline for noisy and geometrically complex GPR B-scan hyperbola detection. © 2026 International Association for Computer Information Systems. All rights reserved.

키워드