Font Size: a A A

Research On Technologies Of 3D Reconstruction Based On Deep Learning

Posted on:2020-03-28Degree:MasterType:Thesis
Country:ChinaCandidate:Y GuoFull Text:PDF
GTID:2518306131466044Subject:Electronics and Communications Engineering
Abstract/Summary:PDF Full Text Request
The purpose of 3D reconstruction is to obtain image information by means of visual inspection equipment and establish a 3D model with a sense of reality.It has been widely used in virtual reality,urban planning,medical imaging,and so on.The traditional 3D reconstruction method is based on the principle of binocular vision geometry.In the real world,due to the complexity of natural scenes and the different characteristics of modeling objects,the traditional modeling process is complex and unable to guarantee the sense of reality.Therefore,it is a challenging task to improve the efficiency of 3D reconstruction on the premise of ensuring the accuracy and sense of reality.In this paper,the key problems of 3D reconstruction technology are analyzed in depth,and the key technologies of multi-view 3D reconstruction based on deep learning are studied in order to improve the quality of 3D reconstruction.In this thesis,we implement a 3D recursive reconstruction method based on residual dense structure and joint loss function.Firstly,in the 3D recursive reconstruction model composed of basic codec framework,the residual dense structure is used to encode the 2D image,so as to better extract the features and improve the convergence speed of network training.Secondly,for the loss function of network training,according to the goal of 3D voxel model reconstruction,the two loss functions of mean square error and cross entropy are combined to improve the reconstruction accuracy of 3D voxel model.Experimental results show that the proposed method achieves better reconstruction performance on the open data set.In this thesis,a multiview 3D reconstruction method based on multilevel fusion and mesh deformation is proposed.The method is mainly composed of multi-level fusion feature extraction module and cascading network lattice deformation module.Firstly,based on the attention mechanism,a multi-level fusion feature extraction module is constructed to extract image features of different scales and gradually fuse multi-perspective image information.Then,the mesh deformation module is used as the basic module to construct the network.Finally,the graph-pooling layer is adopted to connect the mesh deformation module,so that the number of vertices and edges can be increased and the detail processing ability can be improved while maintaining the three-dimensional mesh topology.Experimental results show that the proposed method can achieve good reconstruction performance.
Keywords/Search Tags:3D Reconstruction, Deep Learning, 3D Voxel, Multistage Fusion
PDF Full Text Request
Related items