Dual-Channel CNN-Based Framework for Automated Rebar Detection in GPR Data of Concrete Bridge Decks

Ground Penetrating Radar (GPR) is widely used for assessing the deterioration of concrete bridge decks. GPR surveys generate large amounts of data in the form of B-scan images, which display rebar traces as hyperbolas. Accurate analysis of the GPR scans relies on the effective extraction of rebar locations and amplitudes. This paper presents two automated rebar detection algorithms based on Convolutional Neural Network (CNN) machine learning techniques. Two models are proposed: CNN-1 and CNN-2. CNN-1 was trained on raw GPR images to identify hyperbolas, while CNN-2 model used both raw and migrated GPR images for enhanced analysis. The models were evaluated using GPR data collected from three bridges with different overlay types. Performance was assessed through the visual comparison of the generated bridge amplitude maps against ground-truth data, as well as precision, recall, and F1-score metrics. The results demonstrate that CNN-2 outperforms CNN-1 in terms of accuracy and efficiency for rebar detection.