计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (12): 170-180.DOI: 10.3778/j.issn.1002-8331.2303-0221

• 图形图像处理 • 上一篇    下一篇

级联边缘卷积与注意力机制的点云分类分割研究

王秋红,徐杨,蒋诗怡,熊举举   

  1. 1.贵州大学 大数据与信息工程学院,贵阳 550025
    2.贵阳铝镁设计研究院有限公司,贵阳 550009
  • 出版日期:2024-06-15 发布日期:2024-06-14

Research on Point Cloud Classification and Segmentation of Cascaded Edge Convolution and Attention Mechanism

WANG Qiuhong, XU Yang, JIANG Shiyi, XIONG Juju   

  1. 1.College of Big Data and Information Engineering, Guizhou University, Guiyang 550025, China
    2.Guiyang Aluminum Magnesium Design & Research Institute Co., Ltd., Guiyang 550009, China
  • Online:2024-06-15 Published:2024-06-14

摘要: 近几年点云的分类分割研究多采用多层次架构提取点云特征的方法,提取到较为稳定的高层语义特征,但是全局特征和邻域特征提取不足并且缺乏对上下文信息的特征融合。因此,提出一种新的LAM-EdgeCNN网络,采用边缘卷积与注意力机制级联的方式对点云进行多层级特征提取,获取高层次特征信息。为了加强对特定通道特征和关键空间点的捕捉,提出一种轻量型LAM注意力机制,使用CAM特征通道注意力获取各通道的关联,定位关键通道特征的捕获,使网络更加关注特定通道特征以减少信息弥散和特征冗余;引入SAM空间注意力机制获取点空间的位置信息的注意力权重,增加获得浅层信息的细粒度。采用注意力机制与边缘卷积EdgeConv相结合的方式,增强上下文感知能力,充分提取和融合点云的局部特征和上下文特征,获得面向下游任务的点云特征。将模型应用于公开数据集,实验表明,模型在点云分类、部件分割、语义分割任务中取得良好效果且具有较好鲁棒性。

关键词: 点云分类, 点云分割, 轻量型, 注意力机制, 边缘卷积

Abstract: In recent years, the classification and segmentation research of point cloud mostly adopts the method of extracting the features of point cloud with multi-level architecture, and obtains relatively stable high-level semantic features. However, the extraction of global features and neighborhood features are insufficient, and the feature fusion of context information is lacking. Therefore, a new LAM-EdgeCNN network is proposed in this paper, which adopts the cascade of edge convolution and attention mechanism to extract multi-level feature information from point clouds and obtain high-level feature information. Firstly, in order to enhance the capture of specific channel features and key spatial points, a lightweight LAM attention mechanism is proposed, which uses CAM feature channel attention to acquire the correlation of each channel and locate the capture of key channel features, so that the network pays more attention to specific channel features to reduce the information dispersion and feature redundancy. Secondly, SAM spatial attention mechanism is introduced to obtain the attention weight of the location information of the point space and increase the granularity of the shallow information. Finally, a combination of attention mechanism and edge convolution EdgeConv is used to enhance context awareness, fully extract and fuse the local features and context features of point cloud, and obtain the downstream task-oriented point cloud features. The model is applied to the public data set, and the experiment shows that the model has good effect and robustness in the tasks of point cloud classification, component segmentation and semantic segmentation.

Key words: point cloud classification, point cloud segmentation, lightweight type, attention mechanism, edge convolution