Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (19): 267-275.DOI: 10.3778/j.issn.1002-8331.2103-0037

• Graphics and Image Processing • Previous Articles     Next Articles

Research on Feature Extraction of 3D Point Cloud Based on MANet

WANG Baole, HUO Zhanqiang   

  1. School of Computer Science and Technology, Henan Polytechnic University, Jiaozuo, Henan 454000, China
  • Online:2022-10-01 Published:2022-10-01

基于MANet的三维点云特征提取方法研究

王宝乐,霍占强   

  1. 河南理工大学 计算机科学与技术学院,河南 焦作 454000

Abstract: For feature extraction of 3D point cloud data, the sparsity and irregularity of the point cloud data can affect its global feature representation and the existing methods do not consider the importance difference of different feature channels, which is not conducive to the global optimization of point cloud features. A novel MANet network based on multi-group representation and attention mechanism is proposed for 3D point cloud feature representation. Firstly, to obtain complete point cloud feature information, the point cloud data are input into the multi-group representation module to obtain the initial representation of the point cloud. Then, to learn the importance of different channels in point cloud, channel attention mechanism is introduced to emphasize the important channels for feature representation and suppress the unimportant channels, which optimizes the feature representation further. Finally, the optimized features are input into point cloud classification network. Experimental results show that multi-group representation can perceive local information, attention mechanism can optimize global feature representation, and the proposed method can effectively learn point cloud data, which helps to improve the robustness and accuracy of point cloud classification and component segmentation. The overall accuracy of ModelNet10/40 is 95.1% and 92.5% respectively, the overall accuracy of ScanNet and SHREC15 datasets reaches 78.6% and 97.2% respectively, all outperforming the PointNet++.

Key words: 3D point cloud, feature extraction, multi-group representation, attention mechanism, deep learning

摘要: 在三维点云数据特征提取过程中,点云数据本身的稀疏性和不规则性会影响输入数据的全局特征表示,且现有方法未考虑不同特征通道的重要性差异,不利于点云特征的全局优化。提出一种基于多分组表征和注意力机制的MANet网络进行三维点云特征描述。为获得完整的点云特征信息,将点云数据输入多分组表征模块获得初始点云特征。为学习点云不同通道的重要性,引入新的通道注意力机制强调对特征表示重要的通道,抑制不重要的通道,进一步优化特征表示。将优化后的特征输入点云分类网络,实验结果表明,多分组表征可以感知局部信息,注意力机制能够优化全局特征表示,所提方法能够对点云数据进行有效学习,有助于提高点云分类的鲁棒性和准确率。在ModelNet10/40分类数据集上总体准确率(overall accuracy)分别达到95.1%和92.5%,在ScanNet和SHREC15数据集上总体准确率分别为78.6%和97.2%,上述结果均优于PointNet++网络。

关键词: 三维点云, 特征提取, 多分组表征, 注意力机制, 深度学习