Research For 3D Reconstruction From A Single Image Based On Single Encoder And Multiple Decoders

3D reconstruction is one of classical tasks in the field of computer graphics.Recently,neural networks have been widely used in 2D image processing and have achieved remarkable results.Many researchers have began to try to apply neural networks to 3D reconstruction tasks.However,reconstructing high-quality 3D models from single plots is more challenging than reconstructing from multiple graphs.Compared to ordinary encoder-decoder networks,single-encoder multi-decoder network introduces the concept of parallelization.With multiple decoders,effectively completing the 3D point cloud reconstruction while improving the reconstruction quality,Therefore,how to design 3D reconstruction network with single-encoder multi-decoder structure has become the focus of this paper.The main research work is summarized as follows:(1)Analysis and summarize 3D reconstruction methods based on different forms of characterization: According to the development of 3D reconstruction neural network collect the domestic and foreign related studies of voxel,point cloud and grid.The advantages and disadvantages of the three methods are relatively analyzed,and the 3D point cloud is determined as the study object according to the research needs.In addition,the current common neural network models,the common 3D model data set and the methods of data preprocessing are introduced in detail.(2)An efficient three-dimensional point cloud reconstruction method is proposed:According to the problems of insufficient shape similarity,the three-dimensional reconstruction network based on layered structure is proposed.By designing a new layered structure,extracting the feature information in a single graph,and then using multiple decoders to generate the prediction projection of multiple perspectives in parallel,it is combined to form a dense point cloud.Finally,the pseudo-renderer is used for supervised learning,and the gradient optimization module is introduced to prevent the gradient explosion.The experimental results show that the network can effectively improve the 3D reconstruction of objects,both in both shape similarity and point cloud surface coverage.(3)A 3D point cloud reconstruction method based on the improved feature extraction network is proposed: the method in(2)shows that the enhanced feature extraction of the 2D image can improve the 3D reconstruction effect,Therefore,a 3D reconstruction network based on feature extraction and multi-model fusion is proposed,Fusing the residual unit,the Convolutional Block Attention Module(CBAM),and the Atrous Spatial Pyramid Pooling(ASPP),three modules in the image extraction network.Improve the feature extraction of the input image by increasing the network depth,enhancing the effective features to suppress useless features and expanding the receptive field respectively.The Poly strategy is used during the network training process,which can reduce the training loss.Compare with the mainstream network and the network in the third chapter through the single-category experiment.The results show that the evaluation result of the network in this chapter is better;Ablation experiments are used to explore and analyze the effect of different modules on 3D reconstruction,the results show that the convolutional block attention module and the atrous spatial pyramid pooling module play a dominant role in improving the reconstruction effect.