Single Shot Multibox Detector for Onsite Rebar Detection in Deteriorated Concrete Bridge Deck GPR B-Scan Images

In this paper, rebar detection in deteriorated bridge deck Ground Penetrating Radar (GPR) B-scan images using Convolutional Neural Network (CNN) for onsite application is addressed. A novel approach based on Single Shot Multi-Box meta-architecture (SSD) for real-time rebar detection in bridge decks is adapted for first time in this paper. Extensive experiments on accuracy and resource usage tradeoffs are presented. Two Depthwise Separable CNN backbone networks, MobileNet V1 and MobileNet V2 are used as feature extractor for SSD detector. Various strategies are implemented to trade accuracy for resource usage of the models while keeping Average Precession (mAP) in an acceptable range.