结合偏移自注意力机制和残差连接的点云分类

Point cloud classification combining offset self-attentionmechanism and residual connection

现有基于深度学习的点云分类网络通常无法有效利用点云特征间的相关信息,并且存在鲁棒较低的问题。为了提高点云分类网络对有效特征的提取能力,增强模型鲁棒性,本文提出了一种结合偏移自注意力机制和残差连接的点云分类网络。首先在PointCNN基础上引入偏移自注意力模块,更好地关注于有效特征;然后引入残差网络的思想,在注意力层增加残差连接,将残差连接和注意力层的输出特征进行融合形成点云特征;最后使用多层感知机对点云特征进行分类。将本文模型与PointNet、PointCNN、DGCNN等其他点云分类模型在ModelNet40数据集上进行对比实验,结果表明本文网络的分类效果更好,获得了最高的分类准确率92.9%,相比于PointCNN提升了0.7%。在鲁棒性实验中,本文网络相比于PointCNN,在稀疏点云上的总体分类准确率提升了2.4%,在噪声点云上提升了11.6%,表明本文网络具有更好的鲁棒性。

Existing deep learning-based point cloud classification networks usually cannot effectively utilize the correlation information among point cloud features and suffer from low robustness.In order to improve the ability of point cloud classification networks to extract effective features and enhance model robustness,a point cloud classification network combining an offset self-attention mechanism and the residual connection is proposed in this paper.Firstly,an offset self-attention module is introduced on top of PointCNN to better focus on the effective features.Then,the idea of the residual network is introduced to add residual connections to the attention layers and fuse the output features from residual connections and attention layers to form point cloud features.Finally,a multi-layer perceptron is used to classify the point cloud features.The proposed network is compared with other point cloud networks such as PointNet,PointCNN and DGCNN on the ModelNet40 dataset.The results show that the proposed network achieves better classification results,obtaining the highest classification accuracy of 92.9%,which is 0.7%better than that of PointCNN.In the robustness experiments,the overall classification accuracy of the proposed network is improved bY_(2).4%on sparse point clouds and 11.6%on noisy point clouds compared with PointCNN,indicating that the proposed network has better robustness.