集改进图卷积和多层池化的点云分类模型  

A point cloud classification model with improved graph convolution and multilayer pooling

作　　者：周锐闯 田瑾[1] 闫丰亭 朱天晓 ZHOU Rui-chuang;TIAN Jin;YAN Feng-ting;ZHU Tian-xiao(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620

出　　处：《激光与红外》2024年第2期193-201,共9页Laser & Infrared

基　　金：国家基金委民航联合基金重点项目(No.U2033218)资助。

摘　　要：针对基于图卷积的点云分类模型在提取点云不同语义区域的特征信息以及高效利用聚合的高维特征方面存在的问题,本文提出了一种新的点云分类模型,该模型采用了动态自适应图卷积和多层池化相结合的方法。具体而言,本文采用了残差结构来构建更深层的卷积,以学习不同语义区域点对特征中不同层次的特征信息,从而生成动态自适应调整卷积核,针对不同的点对动态更新边的特征关系,从而提取更为精确的局部特征。同时,本文将聚合的高维特征输入到多层最大池化模块中,回收利用第一次最大池化后丢弃的特征信息进行多层最大池化,从而获取更为丰富的高维特征,提高分类模型的精度。实验结果表明,在ModelNet40数据集上,本文提出的分类模型的总体精度达到93.3%,平均精度为90.7%,明显优于目前主流的点云分类模型,并具有较强的鲁棒性。Aiming at the problems of graph convolution-based point cloud classification models in extracting feature information from different semantic regions of the point cloud and efficiently utilizing aggregated high-dimensional features,a novel point cloud classification model is proposed,which combines dynamic adaptive graph convolution with multi-layer pooling.Specifically,residual structures is employed to construct deeper convolutions and learn feature information from different semantic regions of point pairs at different levels to generate dynamically adaptive adjusted convolution kernels that update the feature relationships of different point pairs,thus extracting more accurate local features.At the same time,the aggregated high-dimensional features are input into a multi-layer max pooling module to recover the discarded feature information from the first max pooling layer and obtain richer high-dimensional features to improve the accuracy of the classification model.The experimental results show that the proposed model achieves an overall accuracy of 93.3%and an average accuracy of 90.7%on the ModelNet40 dataset,which is significantly better than the current mainstream point cloud classification models,and has strong robustness.

关 键 词：深度学习 图卷积神经网络 多层池化 点云分类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN958.98[自动化与计算机技术—计算机科学与技术]