Parallel B-Scan YOLO Plus A-Scan Voting Method for Pavement Integrity Assessment Based on Three-Dimensional Ground-Penetrating Radar Data

Abstract

Pavement integrity assessment based on ground-penetrating radar (GPR) remains challenging because of the uncertainty of the relationship between pavement damage and GPR reflection patterns. Further, existing image-processing techniques for analyzing GPR data still need significant improvements. In this study, a novel framework termed “parallel B-scan You Only Look Once (YOLO) plus A-scan voting” (P-BYPAV) was proposed to improve the performance of YOLO classification in the presence of ambiguous B-scan image patterns. The framework was developed by utilizing GPR-aided coring and falling-weight deflectometer test data. GPR B-scan reflection patterns can be classified into four categories: (1) the valley-peak-valley pattern (V-P-V) representing intact and healthy interfaces where moisture damage is minimal; (2) the peak-valley-peak pattern (P-V-P) representing the presence of air voids at layer interfaces, with a strong correlation to severe debonding; (3) the amplified valley-peak-valley pattern (Amp. V-P-V) representing the presence of moisture at layer interfaces, corresponding to various level of pavement damages; and (4) the symmetric peak-valley pattern, which indicates ambiguity in distinguishing moisture content and air voids. A dataset of 300 labeled images was created using the knowledge of these patterns. The dataset was then used to train the proposed P-BYPAV framework, demonstrating that P-BYPAV successfully integrated the advantages of YOLO and A-scan data recognition. The results show that P-BYPAV resolved more than 97% of the ambiguous regions outlined by YOLO alone.