| 3D reconstruction is one of the important tasks in the field of computer vision research.With the emergence of multiple large 3D model standard data sets,more and more scholars try to transfer mature methods from 2D image tasks to 3D reconstruction research.Compared with traditional reconstruction methods that require expensive image information acquisition equipment to calibrate image information,the reconstruction method based on deep learning does not need to use high-precision 3D information acquisition equipment,but only needs to input 2D images into the trained neural network model to reconstruct 3D data of objects.Among the existing research methods for object 3D data reconstruction using a single 2D image,the method with better reconstruction effect has the problems of large number of network parameters and slow training speed,while the method with lighter network has the problems of poor reconstruction effect.This paper induce and summarizes the advantages and disadvantages of PSGN,a benchmark method in the field of reconstructing point clouds,and studies the working principles of autoencoders,densely connected networks,and attention mechanisms.Insufficient,the PSGN network with double-layer autoencoder has a large amount of computation and slow training speed.Based on the network structure of PSGN with a single-layer autoencoder,a point cloud with self-adaptation and higher accuracy of reconstructed point cloud is proposed.The point cloud reconstruction network greatly improves the accuracy of the reconstructed point cloud while keeping the network lightweight.The specific research contents of this paper are as follows:(1)A point cloud reconstruction network(Dense Connection-PSGN,DC-PSGN)based on the idea of dense connection is proposed.Aiming at the problems of large amount of parameters and slow training speed of PSGN network with two-layer autoencoder,this paper improves the structure of PSGN network with single-layer autoencoder.First,we replace the connection between the convolutional layers of a single convolutional module in the encoder part with a dense connection,which improves the accuracy of the reconstructed point cloud and enables the network to have better performance.Second,we use the features obtained by a single densely connected module as the regularization term of the deep features of the network,which improves the integrity of the feature information.Then,we designed a variety of different network structures for the above improved method and selected the network structure with the best adaptability to the task of reconstructing point clouds through the experimental results.Finally,we demonstrate the effectiveness of the DC-PSGN method by observing and comparing the parameters,computation,convergence speed,and reconstructed point cloud accuracy between DC-PSGN and the original PSGN method.(2)In view of the insufficient accuracy of DC-PSGN for reconstructing point clouds of some types of objects,this paper introduces a variety of attention mechanisms representing different feature reconstruction methods to improve DC-PSGN.The method obtains the weight of the input feature of the dense connection module by using the attention mechanism and reconstructs the output feature of the dense connection module with the above weight,so that the network model can adaptively strengthen important information and weaken non-important information in the process of feature extraction.Secondly,we determine the most suitable attention mechanism for the content studied in this paper by measuring the improvement of the accuracy of the network reconstructed point cloud by introducing different attention mechanisms and the increase in the amount of network parameters and computation.Finally,we prove through experimental results that the accuracy of the network reconstructed point cloud is measured in CD(Chamfer Distance)and EMD(Earth Mover’s Distance)when the number of parameters is reduced by 49% compared to PSGN with two layers of autoencoders.It has increased by 22% and 24% respectively. |
