Scholarlink

<p>The integration of modern technology in the protection and restoration of cultural heritage is of paramount importance, particularly for the Great Wall, a significant architectural landmark facing numerous challenges. In this study, we present an innovative deep learning model—the Multi-Scale Neighborhood Selection Network—designed to enhance the automatic segmentation capabilities of the Great Wall of China and complex architectural heritage point cloud data, thereby facilitating its automated modeling. This model employs adjustable multi-scale neighborhood sizes tailored to different categories of point clouds within the same scene. In addition, it incorporates a decoupling strategy that enables the direct learning of geometric features from the original coordinate space. This approach markedly enhances the model’s ability to comprehend the scene, yielding exceptional performance in the semantic segmentation of architectural heritage point clouds. To comprehensively validate the efficacy of our algorithm, we conducted further tests on publicly available datasets including ArCH and S3DIS, focusing on semantic segmentation tasks. The results across all tests were outstanding.</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 neighborhood selection network for the Great Wall point cloud segmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年6月6日

查看全文
  • Lei Wang,
  • Ming Huang,
  • Xingxing Xiao,
  • Rui Wu,
  • Dong Li,
  • Cai Chen,
  • Miaoxing Zhao

摘要

The integration of modern technology in the protection and restoration of cultural heritage is of paramount importance, particularly for the Great Wall, a significant architectural landmark facing numerous challenges. In this study, we present an innovative deep learning model—the Multi-Scale Neighborhood Selection Network—designed to enhance the automatic segmentation capabilities of the Great Wall of China and complex architectural heritage point cloud data, thereby facilitating its automated modeling. This model employs adjustable multi-scale neighborhood sizes tailored to different categories of point clouds within the same scene. In addition, it incorporates a decoupling strategy that enables the direct learning of geometric features from the original coordinate space. This approach markedly enhances the model’s ability to comprehend the scene, yielding exceptional performance in the semantic segmentation of architectural heritage point clouds. To comprehensively validate the efficacy of our algorithm, we conducted further tests on publicly available datasets including ArCH and S3DIS, focusing on semantic segmentation tasks. The results across all tests were outstanding.