Research And Application Of 3D Point Cloud Data Restoration Technology

Three-dimensional point clouds have broad applications in the field of 3D vision,and the quality of their data significantly affects the subsequent application of point cloud data.However,point cloud data collected by instruments inherently possess sparsity and irregular attributes and require further processing.Existing deep learning-based dense reconstruction methods suffer from excessively smooth reconstruction results and an overabundance of outliers.In deep learning-based point cloud completion reconstruction,the input point cloud data is uneven and noisy,which hampers the accurate representation of the underlying fine structure of the point cloud and the extraction of good point cloud features.This paper addresses these issues and investigates techniques for 3D point cloud dense reconstruction and completion reconstruction.The main work is as follows:(1)Research on point cloud dense reconstruction technology based on PU-GCN.This paper employs an upsampling network framework based on K Nearest Neighbors(KNN)and Graph Convolutional Network(GCN).Specifically,the parallel multi-scale feature extraction module based on graph convolution is optimized to better extract local and global features.After upsampling,a multi-head attention module is added to avoid the bias of a single attention effect,and thus extract valuable fine-grained features with different focuses.The HD of our model has increased by 1.997,and P2 Favg has improved by 1.094,which significantly improves the dense reconstruction performance of the model.(2)Research on point cloud completion reconstruction technology based on dualchannel feature encoding.This paper utilizes a dual-channel feature extraction architecture and constructs a partial point cloud feature extraction module and a spatial refinement module.Specifically,in the partial point cloud feature extraction module,this paper uses a lightweight Dynamic Graph Convolutional Neural Network(DGCNN)to extract the local features of the central points in the point cloud,then embeds these central points into the extracted local features to generate point surrogates.This reduces information loss during the encoding process and better extracts global features.In the spatial refinement module,the obtained coarse point cloud is refined through local feature refinement units,global feature refinement units,and loss function constraints to better represent the fine structure of point cloud features.The CD of our model has increased by 0.04,and F1 has improved by 0.095,which significantly improves the completion reconstruction performance of the model.(3)Design and implementation of the point cloud repair system.This provides a reliable and convenient processing tool for the smooth progression of downstream point cloud work.By using the Django and Bootstrap framework libraries and the Linux operating system,a 3D point cloud repair system is designed and built.With simple operations,the reconstruction work of sparse 3D point clouds and incomplete 3D point clouds is completed,providing reliable data for downstream point cloud workers.