Component-aware post-earthquake damage recognition for RC structures using instance segmentation and oriented bounding box detection

Post-earthquake damage recognition of reinforced concrete structures is vital for safety and recovery. However, most deep learning-based approaches primarily focus on damage detection without considering component information, limiting their effectiveness in structural assessment. This study proposes a component-aware damage identification framework integrating instance segmentation with oriented bounding box (OBB) detection to enhance precision and interpretability. It employs Hybrid Task Cascade to delineate components and YOLOv11 to localize damage. Three innovations support the proposed framework: a convex hull-based OBB labeling method that provides a standardized orientation-aware annotation process; a structural-knowledge-guided data augmentation strategy, damage-intact paste, to address data scarcity and imbalance; and a semantic-edge branch to refine boundary delineation through contextual supervision. Experimental results demonstrate promising performance in both component segmentation (segm_mAP 50 of 0.806) and damage detection (obb_mAP 50 of 0.522). The framework improves automation in structural health monitoring and post-earthquake decision-making, with practical value for safety-critical infrastructure. • A framework integrates component instance segmentation and OBB damage detection. • A convex hull-based OBB labeling standardizes annotation for complex-shaped damage. • A structural knowledge-based augmentation boosts detection under data scarcity. • A semantic-edge branch improves RC component segmentation via boundary supervision.