计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (13): 171-179.DOI: 10.3778/j.issn.1002-8331.2304-0141

• 模式识别与人工智能 • 上一篇    下一篇

结合区域结构关系和自注意力的边卷积点云分类分割网络

吕志玮,杨家志,周国清,沈露   

  1. 1.桂林理工大学 信息科学与工程学院,广西 桂林 541006
    2.广西嵌入式技术与智能系统重点实验室,广西 桂林 541006
  • 出版日期:2024-07-01 发布日期:2024-07-01

Point Cloud Classification Segmentation Combining Inter-Region Structure Relations and Self-Attention Edge Convolution Network

LYU Zhiwei, YANG Jiazhi, ZHOU Guoqing, SHEN Lu   

  1. 1.School of Information Science and Engineering, Guilin University of Technology, Guilin, Guangxi 541006, China
    2.Guangxi Key Laboratory of Embedded Technology and Intelligent Systems, Guilin, Guangxi 541006, China
  • Online:2024-07-01 Published:2024-07-01

摘要: 针对深度学习点云网络中区域内上下文和关系特征捕获不充分问题,提出了一种新的网络框架ISEC-Net(inter-region structure relations and self-attention edge convolution network)。该网络由IrConv(inter-region convolution)模块和SaConv(self-attention convolution)模块组成,SaConv模块可以提取到更细粒的边特征,而IrConv可以动态地将局部结构信息集成到点特征中,然后自适应地捕获区域间关系。在ModelNet40数据集和ShapeNet数据集上分别对点云分类和部件分割进行了大量实验,实验结果表明,在ModelNet40数据集上,ISEC-Net模型的总体精度(OA)达到了93.5%,平均精度(mAcc)达到了90.7%;而在ShapeNet数据集上,平均并交比(mIoU)达到了86.1%,并且在单类并交比(IoU)实验中吉他、耳机、杯子等部件分割精度表现优异,说明ISEC-Net与传统的动态图卷积相比,能够精确地捕捉点云的局部特征和精细结构并加强全局特征的聚合,具有出色的有效性和泛化能力。

关键词: 边卷积, 区域上下文, 区域关系, 自注意力, 深度学习

Abstract: A new network framework, ISEC-Net (inter-region structure relations and self-attention edge convolution network), is proposed to address the problem of insufficient capture of context and relational features within a region in point cloud networks for deep learning. The network consists of two modules:IrConv (inter-region convolution) and SaConv (self-attention convolution). The SaConv module can extract finer edge features, while the IrConv can dynamically integrate local structural information into point features and adaptively capture inter-regional relationships. Extensive experiments are conducted on the ModelNet40 and ShapeNet datasets for point cloud classification and part segmentation. The results show that on the ModelNet40 dataset, the overall accuracy (OA) of the ISEC-Net model reaches 93.5%, and the average accuracy (mAcc) reaches 90.7%. On the ShapeNet dataset, the average intersection-over-union (mIoU) reaches 86.1%, and the part segmentation accuracy of guitar, headphone, cup and other parts in the single-class intersection-over-union (IoU) experiment is excellent. This demonstrates that compared with traditional dynamic graph convolutional networks, ISEC-Net can accurately capture the local features and fine structure of point clouds and enhance the aggregation of global features, thus having excellent effectiveness and generalization ability.

Key words: edge convolution, regional context, regional relationship, self-attention, deep learning