Damage-driven detection and segmentation network: a deep learning framework for fine-grained post-earthquake damage segmentation in RC double-column piers
Computational study demonstrates improved post-earthquake damage segmentation in concrete bridge piers, indicating the efficacy of a locate-crop-segment deep learning framework.
Key Points
To develop and evaluate a fine-grained deep learning segmentation framework capable of accurately identifying small, sparse post-earthquake damage features like exposed rebar in reinforced concrete double-column piers.
Engineered the D3SeN framework using a 'locate, crop, then segment' pipeline combining YOLOv11 for initial localization with an optimized MLSA-DeepLabv3+ model for detailed segmentation.
Integrated multi-level semantic features and self-attention mechanisms within the segmentation architecture.
Evaluated performance using a local validation set and an independent, unseen global test set.
Improved rebar Intersection over Union (rIoU) from 18.24% with the Direct-Global baseline to 27.92% using D3SeN.
Achieved an overall F1-score of 61.92% and an end-to-end damage recall of 62.15% on the global test set.