Dynamic graph CNN point clouds classification network based on channel attention and SoftPool
Jiale Xu, H. Liu, Yichen Shen, Yichen Shen, Xiangyu Wang, Wenkai Pan · IET conference proceedings. · 2023
In recent years, related studies on point clouds have become the hot topic in the field of deep learning due to the rich 3D spatial information. Among classification networks, Dynamic Graph CNN (DGCNN) is efficient, which can deal with point clouds directly and has good performance. Squeeze-and-Excitation Networks (SENet) is a channel attention module that can highlight important information while reducing useless information. Compared with max pooling or average pooling, SoftPool is able to retain more information during the downsampling process. In this paper, we design a novel network called Channel Attention-based Dynamic Graph Convolutional Neural Network (CA-DGCNN). CA-DGCNN combines SENet and SoftPool with DGCNN to improve the accuracy of classification. The experimental results on public ModelNet40 show that CA-DGCNN increases classification accuracy from 92.2% to 93.44% and average class accuracy from 90.2% to 91.02% compared to DGCNN.