Development and Evaluation of a Task-Adaptive and Hierarchical Deep Learning Framework for Multitype Damage Detection in Reinforced Concrete Structures

With the rapid advancement of deep learning, automated inspection of reinforced concrete (RC) structures has become increasingly viable. However, existing models are typically task-specific, limiting their utility across diverse scenarios. This study proposes a unified hierarchical deep learning framework that integrates lightweight classification, near real-time object detection, and pixel-level segmentation to support comprehensive multidamage detection in RC structures. The proposed MobileNet-MHA model, augmented with multihead attention, offers an efficient balance between speed and accuracy for initial screening. YOLOv11 and DeepLab v3+ are further employed for localized damage detection and precise boundary segmentation. Experimental results demonstrate robust performance across five damage types—cracks, spalling, rebar exposure, rebar corrosion, and crushing—even in complex structural scenes. Importantly, this work also conducts a detailed performance bias analysis, revealing that semantic similarity and data imbalance significantly affect recognition reliability, particularly for rebar-related defects. These findings underscore the potential of this task-adaptive framework for deployment in emergency assessments, maintenance planning, and long-term monitoring of RC infrastructures.