Abstract: In order to improve the classification accuracy of 3D point cloud objects, the attention mechanism is integrated with the improved Pointnet network, and the extracted local features and global features are weighted to obtain more effective features for the classification task. Firstly, this model uses Pointnet network as the infrastructure to obtain global features for each point in the sample, uses K-nearest neighbor method to select k neighboring points for each point, and models the relationship between this point and other neighboring points as the local features of the region. Secondly, the SE_block module in the squeeze_excitation_network is used to allocate weights between feature channels, adding an attention mechanism to the modified Pointnet that enables finer and more discriminative feature extraction. Finally, the polymerization was carried out through the mixing pooling layer. The mixing pooling consisted of maximum pooling and average pooling in different proportions. The experimental part of the paper showed the influence of the proportion coefficient.The proposed model achieves 4.1% improvement in object classification accuracy compared with Pointnet on Modelnet40 dataset. Experiments show that the proposed model combining attention and neural network can obtain more discriminative features and effectively improve the classification accuracy of 3D point cloud objects.