Scholarlink

<p>Point cloud semantic segmentation plays a crucial role in the information modeling of traditional Chinese architecture. The complex structures of traditional Chinese buildings, along with rare-class components and large size variations of their characteristic elements, pose considerable challenges for automated semantic segmentation. To address these challenges, we present UNet-PADRB, an enhanced residual U-Net architecture that integrates Dynamic Graph Convolutional Neural Networks (DGCNN) block within a novel Position-Aware Dilated Residual Block (PADRB) framework for simultaneous local and global feature extraction. Specifically, U-Net processes the hierarchical characteristics of rare-class components, while PADRB combines residual learning and position-sensitive graph convolution to handle spatial topology of large size variations features. Experiments on Zhangguying Village point clouds show state-of-the-art performance (OA: 91.89%, mIoU: 71.07%), with significant improvements for small rare-class elements (e.g., doors: +32.7%) and large composite structures (e.g., courtyards: +18.57%). This advances heritage documentation for virtual restoration and risk management.</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

Multi-scale geometric variations semantic segmentation in Chinese architecture point clouds

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月24日

查看全文
  • Yuan Liu,
  • Bo Wu,
  • Jinbiao Yan,
  • Yunyuan Deng

摘要

Point cloud semantic segmentation plays a crucial role in the information modeling of traditional Chinese architecture. The complex structures of traditional Chinese buildings, along with rare-class components and large size variations of their characteristic elements, pose considerable challenges for automated semantic segmentation. To address these challenges, we present UNet-PADRB, an enhanced residual U-Net architecture that integrates Dynamic Graph Convolutional Neural Networks (DGCNN) block within a novel Position-Aware Dilated Residual Block (PADRB) framework for simultaneous local and global feature extraction. Specifically, U-Net processes the hierarchical characteristics of rare-class components, while PADRB combines residual learning and position-sensitive graph convolution to handle spatial topology of large size variations features. Experiments on Zhangguying Village point clouds show state-of-the-art performance (OA: 91.89%, mIoU: 71.07%), with significant improvements for small rare-class elements (e.g., doors: +32.7%) and large composite structures (e.g., courtyards: +18.57%). This advances heritage documentation for virtual restoration and risk management.