Multimodal YOLO-based identification and 3D characterization of concrete surface defects via integrated LiDAR and photogrammetry

Traditional two-dimensional (2D) visual inspections lack the depth information required for accurate three-dimensional (3D) defect quantification in aging concrete infrastructure. This paper proposes a deep learning-based methodology that integrates LiDAR and photogrammetry for the identification and 3D characterization of concrete surface defects. A multimodal defect dataset with six-dimensional spatial-spectral features is first constructed, incorporating a cross-modal mapping mechanism that fuses visual texture information (r, g, b) with high-precision geometric data (x, y, z). A dual-stage cascaded network is developed to achieve pixel-level recognition of multiple defect types based on the You Only Look Once (YOLO) model. Furthermore, a 2D-3D mapping-based quantification framework is proposed to project pixel-level masks into 3D space. Validation demonstrates maximum relative errors of 3.22% for crack width and 5.58% for spalling volume. Notably, the proposed 3D centerline-based method reveals that 2D projections underestimate physical crack lengths, providing more reliable information for structural safety assessment.