Visual-semantic alignment for automatic structural defect detection and diagnosis of prestressed concrete bridges

Abstract

Defect detection and diagnosis are vital for ensuring safe operations of in-service bridges. However, detecting diverse defects from limited, low-quality image datasets remains challenging. Besides, interpreting identified bridge defects into knowledge for diagnosing bridge health conditions is also difficult. This paper develops a visual-semantic alignment tool for automatic bridge defect detection and diagnosis through computer vision and semantic analysis. The proposed visual-semantic alignment tool contains 1) an enhanced YOLOv10-based detection model for capturing bridge defects; 2) a semantic extraction model for extracting defect information from bridge inspection reports; and 3) a Graph Neural Network (GNN)-based diagnosis model for reasoning structural health conditions. Results show that the developed method achieves 91.7 % precision in defect detection and 81.3 % precision in defect diagnosis. Results indicate that aligning visual with semantic information could support effective bridge defect detection and diagnosis. Future research will focus on advancing computational efficiency to support in-situ bridge inspections.