改进PU-GAN的点云上采样网络

Improved network of PU-GAN for point cloud upsampling

针对点云分类问题,提出一种改进点云上采样网络,从低密度点云生成密集点云,提高点云分类准确率。针对点云上采样网络PU-GAN细节表示能力不足的问题,在生成器网络中引入Transformer模块,整合提取的特征信息;在特征扩张模块加入门控循环单元和分层上采样单元,重构细粒度特征;以PU-GAN数据集进行训练,构建闽南古建筑数据集作为测试。实验结果表明,改进后网络的上采样效果获得了提升,具有良好的鲁棒性。通过对ModelNet40数据集进行上采样,在PointNet上进行分类实验,验证了该网络对分类准确率的提升。

Aiming at the problem of point cloud classification,an improved point cloud upsampling network was proposed to gene-rate dense point clouds from low-density ones,thereby improving the accuracy of point cloud classification.To address the insu-fficient detailed representation capability of the PU-GAN point cloud upsampling network,a Transformer module was introduced into the generator network to integrate the extracted feature information.Gated recurrent units and hierarchical upsampling units were incorporated into the feature expansion module to reconstruct fine-grained features.The PU-GAN dataset was used for training,while a Minnan ancient architecture dataset was constructed for testing.Experimental results demonstrate that the proposed network achieves improved upsampling performance and exhibits good robustness.Furthermore,by upsampling the ModelNet40 dataset and conducting classification experiments on PointNet,the network’s enhancement on classification accuracy is validated.