东北大学学报(自然科学版) ›› 2024, Vol. 45 ›› Issue (6): 786-792.DOI: 10.12068/j.issn.1005-3026.2024.06.004

• 信息与控制 • 上一篇    

基于注意力机制的自适应图卷积三维点云识别算法

马原, 佘黎煌(), 李佳蔚, 鲍喜荣   

  1. 东北大学 计算机科学与工程学院,辽宁 沈阳 110169
  • 收稿日期:2023-10-03 出版日期:2024-06-15 发布日期:2024-09-18
  • 通讯作者: 佘黎煌
  • 作者简介:马 原(1999-),男,河南驻马店人,东北大学硕士研究生.
  • 基金资助:
    辽宁省教育厅高等学校基本科研项目(LJKZ0011);辽宁省科学技术计划项目(2021JH1/10400011)

Adaptive Graph Convolutional 3D Point Cloud Recognition Algorithm Based on Attention Mechanism

Yuan MA, Li-huang SHE(), Jia-wei LI, Xi-rong BAO   

  1. School of Computer Science & Engineering,Northeastern University,Shenyang 110169,China.
  • Received:2023-10-03 Online:2024-06-15 Published:2024-09-18
  • Contact: Li-huang SHE
  • About author:SHE Li-huang, E-mail: shelihuang@ise.neu.edu.cn

摘要:

为了更好地捕捉三维点云的局部几何结构信息,提出了一种基于注意力机制的自适应图卷积三维点云识别算法.为了解决固定卷积核忽略特征的缺点,首先通过图结构特征动态学习自适应卷积核;其次为了提高模型对局部几何结构的建模能力,通过向量注意力机制自适应地调整卷积核的权重分配;而后使用点云的位置特征构建图,并利用自适应卷积核来对新构建的图结构特征进行卷积操作;最后通过池化得到新的点云特征.实验结果表明,相较之前的点云卷积算法,所提算法在采样点较少时仍可以很好地提取局部几何结构信息并在分类任务上取得较高精度.所提算法在ModelNet40,ScanObjectNN和ShapeNetPart数据集上的效果对比目前的点云分类和分割方法具有一定的优势.

关键词: 三维点云, 注意力机制, 自适应, 图卷积, 动态学习

Abstract:

To better capture the local geometric structural information of 3D point clouds, an adaptive graph convolutional 3D point cloud recognition algorithm is proposed based on attention mechanism. To address the drawback of fixed convolutional kernels ignoring features, the algorithm first dynamically learns adaptive convolutional kernels based on graph structural features. Furthermore, to enhance the modeling capability of the model for local geometric structures, the weight distribution of the convolutional kernels using a vector attention mechanism is adjusted adaptively. Subsequently, a graph is constructed using the position features of the point cloud and perform convolution operations on the newly constructed graph structural features using the adaptive convolutional kernels. Finally, new point cloud features through pooling is obtained. Experimental results demonstrate that the proposed algorithm effectively extracts local geometric structural information and achieves higher accuracy in classification tasks even with a limited number of sampled points, outperforming previous point cloud convolutional algorithms. The proposed algorithm also exhibits certain advantages compared to existing methods for point cloud classification and segmentation, as evidenced by the performance evaluation on the ModelNet40, ScanObjectNN, and ShapeNetPart datasets.

Key words: 3D point clouds, attention mechanism, self?adaptation, graph convolution, dynamically learn

中图分类号: