Comparative analysis of road subsurface defect detection models using multiple GPR display formats and multi-image feature fusion networks

Numerous algorithms have been proposed for detecting road subsurface defects using data from Ground Penetrating Radar (GPR). However, GPR operates across multiple display formats, and existing methods typically rely on a single display format for analysis. Our comprehensive tests have shown that different GPR display formats vary in their effectiveness at detecting specific types of defects; certain formats may be optimal for one type of defect but less effective for another. Therefore, this study has conducted a comparative analysis using the YOLOv7 framework as the benchmark model to assess the impact of different display formats and multi-image feature fusion modes (pixel-level fusion, feature-level fusion, and decision-level fusion) on the detection outcomes of subsurface defects. Using a dataset of actual road subsurface defects for validation, the results demonstrate that different GPR display formats excel at detecting specific types of road subsurface defects. The decision-level fusion method notably enhances detection accuracy and recall rates, with improvements of 2.78% for cavities and 10% for cracks compared to single-input modes.