Scholarlink

<p>In point cloud classification tasks, the inherent disorder and sparsity of 3D point clouds often make it challenging for existing models to effectively capture both local details and global dependencies, which limits the improvement of classification accuracy. Therefore, to simultaneously capture local and global information, this paper proposes a network model, PointLGFN, which effectively enhances the feature representation capability of point cloud data. We developed the Local and Global Feature Extraction (LGFE) module, utilizing diverse pooling strategies to isolate and extract both salient and smooth features. Next, we introduce the Local Context Fusion (LCF) module, which employs bilinear regularization to reinforce local feature representation in point clouds. Lastly, we incorporate the Global Feature Transformer (GFT) module, leveraging a multi-head self-attention mechanism to effectively capture long-range dependencies in point clouds and improve global feature extraction. We conduct experiments on the ModelNet40 and ScanObjectNN datasets, and the results demonstrate that PointLGFN achieves significant performance improvement in classification tasks, achieving accuracy rates of 93.4% and 88.2%, respectively.</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

Pointlgfn: local–global fusion network for point cloud classification

  • 期刊论文

  • 发表日期：

    2025年12月6日

查看全文
  • Tong Liu,
  • Shengwei Tian,
  • Long Yu,
  • Chaoyue Wu,
  • Jie Li,
  • Guoqi Wang,
  • Pusen Xia

摘要

In point cloud classification tasks, the inherent disorder and sparsity of 3D point clouds often make it challenging for existing models to effectively capture both local details and global dependencies, which limits the improvement of classification accuracy. Therefore, to simultaneously capture local and global information, this paper proposes a network model, PointLGFN, which effectively enhances the feature representation capability of point cloud data. We developed the Local and Global Feature Extraction (LGFE) module, utilizing diverse pooling strategies to isolate and extract both salient and smooth features. Next, we introduce the Local Context Fusion (LCF) module, which employs bilinear regularization to reinforce local feature representation in point clouds. Lastly, we incorporate the Global Feature Transformer (GFT) module, leveraging a multi-head self-attention mechanism to effectively capture long-range dependencies in point clouds and improve global feature extraction. We conduct experiments on the ModelNet40 and ScanObjectNN datasets, and the results demonstrate that PointLGFN achieves significant performance improvement in classification tasks, achieving accuracy rates of 93.4% and 88.2%, respectively.