Research For Three-dimensional Point Cloud Reconstruction From A Single Image Based On Deep Learning

Three-dimensional reconstruction is a classic task in the field of computer graphics.In recent years,deep learning has been widely used in two-dimensional image processing and has achieved remarkable results.More and more researchers try to bring the success of deep learning on two-dimensional image processing tasks to the task of 3D reconstruction,which makes the research on 3D reconstruction based on deep learning has gradually become a research hotspot.Compared with traditional 3D reconstruction methods that require precise acquisition equipment and strict calibration of image information,3D reconstruction methods based on deep learning can match 2D images to 3D models through deep neural networks and able to reconstruct 3D models of objects of various categories massively and quickly even using RGB images obtained from ordinary acquisition equipment.Reconstructing high-quality 3D models from a single image is more challenging.How to extract image features,restore 3D model information,and generate high-quality 3D models has become the focus of 3D reconstruction research.Main researches in this thesis are summarized as follows:(1)Analyzing and summarizing the existing three-dimensional reconstruction methods based on deep learning: According to the development of 3D reconstruction based on deep learning,the domestic and foreign related researches of 3D voxel reconstruction,3D point cloud reconstruction and 3D mesh reconstruction are sorted out.After analyzing the advantages and disadvantages of each 3D model representation and the research needs,3D point cloud is taken as our research object.Besides,the 3D model data sets,data processing methods,evaluation methods and basic operations used in the 3D reconstruction neural network are also analyzed in detail.(2)Proposing an attention-based dense point cloud reconstruction method: In view of the fact that most researches of 3D point cloud reconstruction in the past were limited to reconstruct low-resolution point clouds,a method of reconstructing highresolution point clouds in multi-stage was proposed.A low-resolution point cloud is generated as an intermediate result from a single RGB image through an attention-based auto-encoder network,and then is passed into the point cloud high-resolution network to generate a high-resolution point cloud.Compared with the quantitative evaluation and visualization results of the existing high-resolution 3D point cloud reconstruction methods,the effectiveness of the method is verified.(3)Proposing a deformation-based multi-resolution point cloud reconstruction method: Aiming at the problem that the method proposed in(2)can only generate a specific resolution point cloud and the reconstruction of some categories of objects is defective,a method that can flexibly reconstruct multiple resolution point clouds is proposed.First,the input image features are extracted through the image encoding stage.Then the image features are merged with the initial random point cloud in the point cloud encoding stage to extract the point cloud features.Finally,the initial random point cloud shape is deformed into the final reconstructed point cloud using the point cloud features in the point cloud decoding stage.Due to the randomness introduced by the initial point cloud,point coordinates of each reconstructed point cloud shifts randomly slightly.Overlapping multiple reconstruction results gets multiple resolution point clouds.Compared with the quantitative evaluation and visualization results of the existing both low-resolution and high-resolution 3D point cloud reconstruction methods,this method has achieved better performance.