Scholarlink

<p>The accurate and fast classification of 3D mesh models is crucial due to their wide application. In recent years, the Broad Learning System (BLS) has brought an essential change to traditional classification methods due to its simple structure and excellent computational efficiency. However, the performance of BLS is inadequate when processing 3D mesh models. Therefore, we propose MeshBLS, a mesh-based broad learning 3D object classification network. Firstly, we select the coordinates of the face’s center point, the curvature of the mesh surface, and other descriptors to input them into the network in parallel. Secondly, we design a feature extractor to replace the feature mapping layer of BLS and then generate the enhanced layer. Finally, the output from each branch is fused to get the result. MeshBLS tackles common problems in the current 3D mesh model classification network, such as lengthy training times, high computing resource demands, and complex structures. Experiments on Shrec-11 and Cube Engraving datasets demonstrate that MeshBLS can significantly reduce the training time and achieve accuracy close to the latest mesh classification network. The results indicate that MeshBLS is suitable for real-time classification scenarios. Our code and data are available at <a href="https://github.com/haozhix/MeshBLS">https://github.com/haozhix/MeshBLS</a>.</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

MeshBLS: mesh-based broad learning 3D object classification network

  • 期刊论文

  • 发表日期：

    2025年1月6日

查看全文
  • Guoyou Zhang,
  • Zhixiang Hao,
  • Lihu Pan,
  • Wei Guo,
  • Jiaxin Zuo,
  • Xuenan Zhang

摘要

The accurate and fast classification of 3D mesh models is crucial due to their wide application. In recent years, the Broad Learning System (BLS) has brought an essential change to traditional classification methods due to its simple structure and excellent computational efficiency. However, the performance of BLS is inadequate when processing 3D mesh models. Therefore, we propose MeshBLS, a mesh-based broad learning 3D object classification network. Firstly, we select the coordinates of the face’s center point, the curvature of the mesh surface, and other descriptors to input them into the network in parallel. Secondly, we design a feature extractor to replace the feature mapping layer of BLS and then generate the enhanced layer. Finally, the output from each branch is fused to get the result. MeshBLS tackles common problems in the current 3D mesh model classification network, such as lengthy training times, high computing resource demands, and complex structures. Experiments on Shrec-11 and Cube Engraving datasets demonstrate that MeshBLS can significantly reduce the training time and achieve accuracy close to the latest mesh classification network. The results indicate that MeshBLS is suitable for real-time classification scenarios. Our code and data are available at https://github.com/haozhix/MeshBLS.