Damage-driven detection and segmentation network: a deep learning framework for fine-grained post-earthquake damage segmentation in RC double-column piers

Reinforced concrete (RC) double-column piers are susceptible to seismic damage, such as concrete spalling and exposed rebar. Accurate pixel-level segmentation of these damage types is essential for automated post-earthquake assessments. However, fine-grained damage features, especially exposed rebar, remain difficult to identify because they occupy only a small proportion of pixels in high-resolution images. This paper proposes the Damage-Driven Detection and Segmentation Network (D3SeN) to address this challenge. The framework employs a “locate, crop, then segment” strategy, utilising YOLOv11 for preliminary localisation and an optimised MLSA-DeepLabv3+ model for fine-grained segmentation. The segmentation model integrates multi-level semantic features and a self-attention mechanism. Evaluations on the local validation set verify the optimisation stability of MLSA-DeepLabv3+, while a strictly independent and unseen global test set provides the primary evidence for framework-level performance. D3SeN improves rebar Intersection over Union (rIoU) from 18.24% under the Direct-Global baseline to 27.92%, while achieving an overall F1-score of 61.92% and an end-to-end damage recall of 62.15%. These results demonstrate the effectiveness of preserving information density for fine-grained post-earthquake damage segmentation.