基于可变形图卷积的点云表征学习  

Deformable Graph Convolutional Networks Based Point Cloud Representation Learning

作　　者：李宗民[1] 张玉鹏[1] 刘玉杰[1] 李华[2] LI Zong-min;ZHANG Yu-peng;LIU Yu-jie;LI Hua(College of Computer Science and Technology,China University of Petroleum,Qingdao,Shandong 266580,China;Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中国石油大学(华东)计算机科学与技术学院,山东青岛266580 [2]中国科学院计算技术研究所,北京100190

出　　处：《计算机科学》2022年第8期273-278,共6页Computer Science

基　　金：国家重点研发计划(2019YFF0301800);国家自然科学基金(61379106);山东省自然科学基金(ZR2013FM036,ZR2015FM011)。

摘　　要：虽然深层神经网络较为成功地解决了点云数据稀疏和不规则等问题,但是点云局部特征的学习仍然是一个非常具有挑战性的问题。现有的用于点云表征学习的网络存在点与点之间相互独立提取特征的问题,为此提出了一种全新的空域图卷积。首先,在构造图结构时提出了自适应空洞K近邻算法,以最大程度地捕获局部拓扑结构信息;其次,在卷积中加入了卷积核每条边与感受野图之间的角度特征,保证了更具鉴别力的特征提取;最后,为了充分利用局部特征,提出了一种全新的图金字塔池化,以更好地融合多尺度特征。在标准公开数据集ModelNet40和ShapeNet上测试该算法,分别取得了93.2%与86.5%的准确度。实验结果表明,该算法在点云表征学习中处于领先水平。Although the sparseness and irregularity of point cloud data have been successfully solved by deep neural networks.However,how to learn the local features of point clouds is still a challenging problem.Existing networks for point cloud representation learning have the problem of extracting features independently between points and points.To this end,a new spatial graph convolution is proposed.Firstly,an adaptive hole K-nearest neighbor algorithm is proposed when constructing the graph structure to maximize local topo-logical structure information.Secondly,the angle feature between each edge of the convolution kernel and the receptive field map is added to the convolution,which ensures more discriminative feature extraction.Finally,in order to make full use of local features,a novel graph pyramid pooling is proposed.This algorithm is tested on the standard public data sets ModelNet40 and ShapeNet,and the accuracy is 93.2%and 86.5%respectively.Experimental results show that the proposed algorithm is at a leading level in point cloud representation learning.

关 键 词：图卷积神经网络 表征学习 点云 局部特征 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]