Scholarlink

<p>To mitigate on-site safety risks for field investigators and provide quantitative support for the conservation decision-making of historic buildings, this paper presents a lightweight damage-detection and safety-assessment framework for historic building components named HBDSegformer. The framework employs SegFormer-B0 as its backbone and enhances segmentation accuracy by integrating an Object-Contextual Representation (OCR) module that models pixel-region relationships and a parameter-free attention mechanism—SimAM. To maintain a lightweight footprint comparable to SegFormer-B0, standard 3 × 3 convolutions within the OCR module are replaced with Partial Convolution. Evaluated on our self-constructed historic-building damage datasets, the proposed model achieves a mIoU of 65.61% and a mPA of 74.53%, outperforming existing baselines. Furthermore, we establish a damage-rate-based safety grading system that maps segmentation outputs into risk levels. On 165 test images, the safety-assessment task attains a prediction accuracy of 86.10%, offering a reliable technical foundation for on-site inspections and conservation decisions of historic buildings.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Damage Detection and Safety Assessment for Historic-District Buildings Using a Semantic Segmentation Model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月7日

查看全文
  • Guangyuan Wang,
  • Jianmin Zhou,
  • Huajian Lu,
  • Qifan Xu,
  • Zihan Li,
  • Weihao Huang,
  • Guosen Tan

摘要

To mitigate on-site safety risks for field investigators and provide quantitative support for the conservation decision-making of historic buildings, this paper presents a lightweight damage-detection and safety-assessment framework for historic building components named HBDSegformer. The framework employs SegFormer-B0 as its backbone and enhances segmentation accuracy by integrating an Object-Contextual Representation (OCR) module that models pixel-region relationships and a parameter-free attention mechanism—SimAM. To maintain a lightweight footprint comparable to SegFormer-B0, standard 3 × 3 convolutions within the OCR module are replaced with Partial Convolution. Evaluated on our self-constructed historic-building damage datasets, the proposed model achieves a mIoU of 65.61% and a mPA of 74.53%, outperforming existing baselines. Furthermore, we establish a damage-rate-based safety grading system that maps segmentation outputs into risk levels. On 165 test images, the safety-assessment task attains a prediction accuracy of 86.10%, offering a reliable technical foundation for on-site inspections and conservation decisions of historic buildings.