基于双支融合与结构采样的点云分类算法

A classification algorithm for point cloud based on the dual-branch fusion and structure sampling

目的:利用双支融合和结构采样,提取较充分的局部特征并保留不同尺度下的全局结构信息,使得在点云分类任务中获得高精度。方法:首先,提出双支融合模块,并用之提取不同角度的局部特征,再加入融合框架剔除重复的特征,突出显著特征,使得局部特征更具有辨别力。其次,设计了结构采样单元以关注不同尺度下的全局结构信息。该单元通过对每个采样点学习,以增强采样稳定性,并平滑结构信息,从而学习点云的全局结构信息。最后,利用包含局部-全局的特征进行分类,完成点云的分类任务。结果:在ModelNet40和ScanObjectNN基准数据集中分别获得93.6%和86.6%的分类精度。结论:基于双支融合与结构采样网络模型在点云分类任务中具有先进性。

Aims:This paper aims to improve the accuracy in the point cloud classification task by using the dual-branch fusion and structure sampling to extract adequate local features and retain global structure information at different scales.Methods:Firstly,a dual-branch fusion module was proposed to extract the local features from different angles.Then the fusion framework was added to eliminate the repeated features,highlight the salient features,and make the local features more discriminating.Secondly,a structure sampling unit was designed to focus on the global structure information at different scales to learn the global structure information of the point cloud by learning from each sampling point to enhance the sampling stability and smooth the structure information.Finally,the local-global features were used to classify and complete the classification task of the point cloud.Results:The classification accuracy was 93.6%and 86.6%in ModelNet40 and ScanObjectNN benchmark datasets,respectively.Conclusions:The network model based on the dual-branch fusion and structural sampling is advanced in the point cloud classification task.