面向点云理解的双邻域图卷积方法  

Dual-neighborhood graph convolution method for point cloud understanding

作　　者：李宗民[1,2] 徐畅 白云[1] 鲜世洋[1] 戎光彩 LI Zongmin;XU Chang;BAI Yun;XIAN Shiyang;RONG Guangcai(College of Computer Science and Technology,Qingdao Institute of Software,China University of Petroleum(East China),Qingdao 266580,China;Information Engineering College,Qingdao Binhai University,Qingdao 266580,China)

机构地区：[1]中国石油大学(华东)青岛软件学院计算机科学与技术学院,山东青岛266580 [2]青岛滨海学院信息工程学院,山东青岛266580

出　　处：《浙江大学学报(工学版)》2025年第5期879-889,共11页Journal of Zhejiang University:Engineering Science

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

摘　　要：针对现有方法对局部点云结构建模时空间跨度有限以及传统特征聚合方法造成一定信息损失的问题,提出双邻域图卷积网络(DNGCN).在原始点云中增加角度先验,以增强对点云局部几何结构的理解,捕捉局部细节.对原始邻域进行扩展,在局域内设计双邻域图卷积,通过集成高斯自适应聚合,在提取较大感受野范围内显著特征的同时,充分保留原始邻域信息.通过局部-全局信息交互来增大局部点的空间跨度,捕获远距离依赖关系.本文方法在分类数据集ModelNet40和ScanObjectNN上分别取得了94.1%、89.6%的总体精度,与其他先进算法相比有显著提升,较DGCNN分别提升了1.2%、11.5%.在部件分割数据集ShapeNetPart和语义分割数据集ScanNetv2、S3DIS上均获得优秀的性能,平均交并比分别为86.7%、74.9%和69.8%.通过大量的实验,证明了该模型的有效性.A dual-neighborhood graph convolutional network(DNGCN)was proposed in order to address the limitations of existing methods in modeling local point cloud structures with restricted spatial spans and the information loss caused by conventional feature aggregation strategies.Angular priors were incorporated into raw point coordinates in order to enhance geometric awareness for capturing fine-grained local structures.A dual-neighborhood graph convolution operator that integrated Gaussian adaptive aggregation was designed by extending the original neighborhood,enabling simultaneous extraction of salient features from enlarged receptive fields and preservation of intricate local details.A local-global cross-scale interaction mechanism was introduced to expand spatial perception spans and model long-range dependencies.The proposed method achieved an overall classification accuracy of 94.1%on ModelNet40 and 89.6%on ScanObjectNN,significantly outperforming other advanced algorithms.The increases were 1.2%and 11.5%respectively compared with DGCNN.Excellent performance was obtained on the ShapeNetPart dataset for part segmentation,as well as the ScanNetv2 and S3DIS datasets for semantic segmentation,with mean IoU scores of 86.7%,74.9%and 69.8%,respectively.Experiments proved the effectiveness of the model.

关 键 词：点云特征 图卷积网络 几何增强 局部全局交互 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]