Research On Point Cloud Upsampling Based On Deep Learning

Point cloud technology is a 3D data processing technology that has developed rapidly in recent years.It has a wide range of applications in the reconstruction,simulation,analysis,recognition,and classification of 3D objects.However,due to many factors when collecting point clouds,such as light and shade,object surface materials,occlusion,device noise,sensor resolution,etc.,the collected point clouds are usually incomplete,sparse,noisy and uneven Therefore,it is necessary to preprocess and optimize the collected point cloud data to improve the effect of subsequent applications.Most of the existing point cloud upsampling networks are three-tier cascaded networks that combine feature extraction,feature expansion,and coordinate reconstruction.There is a non-negligible deviation between the point cloud generated by this network architecture and the ground truth after upsampling,especially in the details.At the same time,most of the existing point cloud upsampling networks only consider factors of certain integer scales(eg:2x,4x),which limits the application of point cloud upsampling networks in real-world scenarios.In practical applications,it is necessary to upsample the original point cloud into various usercustomized densities to achieve mesh reconstruction,point cloud processing or other requirements,which also means that different scale factors are required to fit the original points of different densities.point cloud.To address the above issues,this paper proposes two point cloud upsampling networks.First,this paper designs a graph convolutional point cloud upsampling network based on spatial refinement.The network adopts a four-tier cascaded network architecture and is mainly divided into four parts:feature extraction,feature expansion,coordinate reconstruction,and spatial refinement.The four-tier cascaded point cloud upsampling network introduces a GCN-based feature expansion module and a spatial refinement module,which further refines the dense point clouds generated by traditional three-tier cascaded networks,obtaining more refined,smooth,and uniform point clouds,thereby better restoring the original scene.Second,to address the problem that existing networks can only upsample to specific multiples,this paper designs a flexible scale upsampling network framework based on feature fusion.The network mainly consists of a feature extraction module and a flexible scale upsampling module.In terms of feature extraction module,this paper designs a multi-scale graph convolution feature extractor called Dense Edge Conv.The feature extractor can fuse local features,global features and input features to form a fusion feature map.This feature fusion method can not only focus on the local geometric structure,but also avoid losing the overall shape structure of the point cloud,thus improving the effect of point cloud upsampling.In terms of the flexible scale upsampling module,this paper designs a module that can handle upsampling factors of various scales.In this way,the point cloud upsampling network designed in this paper can cope with different scale factors after one training,instead of training a separate model for each scale factor,which improves the flexibility and practicality of point cloud upsampling network.To verify the effectiveness of the method in this paper,this paper conducted comparative experiments and ablation experiments on multiple datasets.The experimental results show that both the proposed spatial refinement-based graph convolutional point cloud upsampling network and the proposed feature fusion-based flexible scale upsampling network can achieve better performance than the state-of-the-art methods.