Abstract: Abstract: Existing 3D point cloud classification methods struggle to effectively handle fine-grained features and intra-class classification issues. A point cloud fine-grained classification model based on high-frequency attention filtering is proposed. Firstly, a feature dimensionality enhancement module based on spatial cosine encoding is introduced to address the sparsity of point cloud data. This approach preserves more high-frequency fine-grained features without increasing the model's parameter count. Then, a multi-layer bilinear attention feature fusion module is employed to further capture the feature differences in different point cloud neighborhoods. Finally, a graph attention-based high-frequency downsampling module is used to extract feature information of high-frequency points at different scales, with residual connections introduced to fuse features from different levels. The final fused features are used for fine-grained classification. The model achieves an overall classification accuracy of 93.78% on the standard public dataset ModelNet40, and an accuracy of 96.43%, 82.7%, and 79.23% on the three sub-datasets, Airplane, Chair, and Car, of the fine-grained classification dataset FG3D.

Key words: deep learning, point cloud classification, fine-grained classification, bilinear pooling, graph neural networks, attention mechanism

摘要： 摘 要: 针对3D点云现有分类方法难以有效处理点云的细粒度特征和类内分类的问题，提出了一种基于高频注意力滤波的点云细粒度分类模型。首先，主要针对点云信息的稀疏性提出了基于空间余弦编码的特征升维模块，在不提升模型参数量的情况下保留更多的高频细粒度特征。然后通过多层的双线性注意力特征融合模块进一步提取不同点云邻域的特征差异。最后通过图注意力高频下采样模块，获取不同尺度下高频点的特征信息，引入残差连接融合不同层级的特征，最终融合特征进行细粒度的特征分类。在标准公开数据集ModelNet40的总体分类精度为93.78%，在细粒度分类数据集FG3D三个子数据集Airplane、Chair和Car上的精度为96.43%、82.7%、79.23%。

关键词: 关键词: 深度学习, 点云分类, 细粒度分类, 双线性池化, 图神经网络, 注意力机制