Enhancing bridge defect semantic segmentation via synthetic point cloud data generation

The structural integrity of bridges is critical for modern transportation systems, yet current methods for defect detection and segmentation on curved concrete surfaces remain limited in precision and cost-effectiveness. Data scarcity is also a significant problem. This study addresses these challenges by proposing a framework leveraging synthetic data generation and a Surface Normal Enhanced PointConv (SNEPointConv) model for bridge defect semantic segmentation. The proposed approach includes a low-cost method for generating synthetic cracks and spalling defects that mimic real-world bridge defect geometries, enabling effective training of deep learning models. Additionally, the SNEPointConv model integrates normal vector enhancements to improve feature extraction from irregular point clouds. Experiments demonstrate the feasibility of using synthetic datasets for defect semantic segmentation, bridging the gap between real-world defect characteristics and digital twin applications for structural health monitoring. Key contributions include the development of scalable synthetic defect data generation techniques, improved defect segmentation accuracy through feature enhancement, and a comprehensive exploration of the effectiveness of synthetic dataset in model training.