Answers are grounded in this paper and cite their sources.
Post-earthquake manual inspection of reinforced concrete (RC) components suffers from low efficiency, strong subjectivity, difficult quantification, and lack of non-contact automation. Traditional semantic segmentation models perform poorly on fine cracks under small-sample and class-imbalanced conditions, yielding low accuracy, blurred boundaries, and insufficient lightweight design. To achieve fine-grained, pixel-level, lightweight segmentation of multi-scale seismic damage, this paper developed a customized DAM-DeepLab configuration based on DeepLabV3+, which systematically integrated lightweight backbone, dense multi-scale feature extraction, attention mechanism, and Dice loss to address the specific challenges of RC seismic damage segmentation. The model integrated four enhancements: a lightweight MobileNetV2 backbone, a DenseASPP module for dense multi-scale feature extraction, a CBAM attention mechanism, and the Dice loss objective. A seismic damage dataset covering cracks, spalling, and crushing was constructed, and training was subsequently completed using a combination of data augmentation and transfer learning. Experimental results show that the proposed model achieved a mean intersection over union (mIOU) of 64.65% and a mean F1 score (mF1) of 76.10%, significantly outperforming DeepLabV3+, U-Net, SegFormer-B0, and PSPNet, while requiring only 64 s per training epoch. The model preserved fine crack continuity and damage region integrity, remained robust under complex backgrounds, and enabled precise calculation of damage area ratio as well as crack complexity quantification via persistent entropy ratio and minimum spanning tree topology. This study effectively addresses high-precision multi-class seismic damage segmentation under small samples and class imbalance, offering an efficient tool for post-earthquake structural inspection and image-based quantitative damage characterization, with potential for future residual-performance assessment.
No takes yet. Share an insight, caveat, or question.
Zheng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: