摘要
三维点云分类分割网络忽视了融合特征中的冗余信息,缺乏放大有效特征占比能力,不能充分挖掘特征的表达性。在CurveNet网络基础上,提出了一种能够筛选和丰富融合特征的方法,对点云的识别与分割效果达到了较先进水平。首先,提出了对融合特征具有筛选能力的特征选择子网络,利用结合了打分机制的Top K算子选出包含有效信息的融合特征,并且能够自适应地赋予被选特征权重。其次,在聚合曲线特征模块中增加了两个新分支,分别学习曲线内部点距离特征和曲线之间的线距离特征,通过快速通道相关性注意力机制提取各分支的内部相关性,增强了网络特征的信息描述能力。实验结果表明,分类任务在ModelNet40数据集上准确率达到了93.8%,分割任务在ShapeNet Part数据集上平均交并比达到了86.4%。与基准网络相比,分类效果与分割效果均有所提高,证明了算法的有效性。
The 3D point cloud classification and segmentation networks ignore the redundant information in the fusion features,lack the ability to amplify the proportion of effective features,and cannot fully explore the expressiveness of features.Based on the CurveNet network,this paper proposes a method that can filter and enrich the fusion features,and the recognition and segmentation effect of point cloud reaches a relatively advanced level.Firstly,a feature selection subnetwork with filtering ability for fusion features is proposed,which combines Top K operator and a scoring mechanism to select fusion features containing valid information and adaptively assign weight to the selected features.Secondly,two new branches are added to the aggregation curve feature module,so as to learn the curve internal point distance features and the curve line distance features,respectively,and extract the internal correlation of each branch through the quick channel affinity attention mechanism,which enhances the information description ability of the network features.The experimental results show that the accuracy of the classification task on the ModelNet40 dataset reaches 93.8%,and the average intersection over union of the segmentation task on the ShapeNet Part dataset reaches 86.4%.Compared with the benchmark network,the classification effect and segmentation effect are improved,which proves the effectiveness of the proposed algorithm.
作者
朱安迪
达飞鹏
盖绍彦
ZHU Andi;DA Feipeng;GAI Shaoyan(School of Automation,Southeast University,Nanjing 210096,China;Key Laboratory of Measurement and Control of Complex Systems of Engineering of Ministry of Education,Southeast University,Nanjing 210096,China)
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2024年第5期52-63,共12页
Journal of Xi'an Jiaotong University
基金
江苏省前沿引领技术基础研究专项项目(BK20192004C)。
关键词
三维点云
融合特征筛选
曲线特征
注意力机制
分类分割
3D point cloud
fusion feature screening
curve feature
attentive mechanism
classification and segmentation
