Fine-grained point cloud classification based on hierarchical feature enhancement. Journal of Zhejiang University (Science Edition),2025,52(1):70⁃80(基于层次特征增强的细粒度点云分类)
Aiming at the problem of insufficient local feature extraction of general point cloud classification methods in fine-grained classification tasks, we propose a point cloud-oriented 3D model classification framework, HFE-Net. The Veronese mapping-based point feature enhancement module (V-PE) is used...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | zho |
| Published: |
Zhejiang University Press
2025-01-01
|
| Series: | Zhejiang Daxue xuebao. Lixue ban |
| Subjects: | |
| Online Access: | https://doi.org/10.3785/j.issn.1008-9497.2025.01.008 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Aiming at the problem of insufficient local feature extraction of general point cloud classification methods in fine-grained classification tasks, we propose a point cloud-oriented 3D model classification framework, HFE-Net. The Veronese mapping-based point feature enhancement module (V-PE) is used to enhance the point cloud data, so that the network learns higher-order information of the normal and the attitude; the multi-scale context-aware intra-cluster feature enhancement module (CA-IntraCE) utilizes different scales of K-nearest neighbor algorithms and cross-attention to achieve different scales of features and eliminate the loss of information caused by maximal pooling; the inter-cluster feature enhancement module (GSS-InterCE) based on grouped sparse sampling utilizes the furthest-point-sampling (FPS) algorithm to obtain sparse points and the cross-attention to achieve the enhancement of different clusters, so that the network has stronger fine-grained discriminative ability.In the experimental results on the three sub-datasets Airplane, Car, and Chair of FG3D, the overall accuracies of HFE-Net reach 97.40%, 80.53%, and 83.83%, respectively, which have ex-ceeded those of the existing SOTA methods, DC-Net and FGPNet, showcasing the superior classification performance of HFE-Net.(针对粗粒度点云分类方法在细粒度数据集中局部特征提取不足的问题，提出了一种基于层次特征增强的三维细粒度点云分类网络（HFE-Net）。基于Veronese映射的点特征增强模块（V-PE）对点云数据进行数据增强，辅助网络学习法线和姿态高阶信息；基于多尺度上下文感知的簇内特征增强模块（CA-IntraCE），利用不同尺度的K近邻（K-nearest neighbors，KNN）算法以及交叉注意力实现不同尺度特征的增强，以消除最大池化带来的信息丢失；基于分组稀疏采样的簇间特征增强模块（GSS-InterCE），利用最远点采样（FPS）算法获得稀疏点，并采用交叉注意力实验不同簇间的特征增强，从而提高网络的细粒度判别能力。在FG3D数据集Airplane、Car和Chair 3个类别上的实验结果显示，HFE-Net的总体准确率分别达97.40%，80.53%和83.83%，已超过现有最优方法DC-Net、FGPNet的分类框架，说明HFE-Net的分类性能具有一定的优越性。) |
|---|---|
| ISSN: | 1008-9497 |