Research On Multi-view Voxel 3D Reconstruction

From the neural network in the early days of computer science,to the once hot expert system later,and now to the booming deep learning,the research of artificial intelligence has experienced ups and downs.Thanks to the explosion of data and the improvement of computer processing capacity,artificial intelligence has ushered in a new development boom.Today,machine learning methods have spread to all areas of artificial intelligence,such as computer vision,natural language understanding,and speech recognition.Among them,computer vision is one of the most important core technologies.With the increasing demand of unmanned driving,robot grasping objects,3D printing,virtual reality and other applications,the automatic reconstruction of 3D models from 2D images has become a demand.People can get more information from threedimensional models than from two-dimensional images.Therefore,this research field is becoming more and more important.The emergence of some large 3D model libraries has promoted the development of this research field.In recent years,due to the rapid development of deep learning,learning-based methods have become the mainstream of 3D reconstruction.The purpose of this thesis is to study the 3D reconstruction technology based on deep learning,and to explore how to get a more accurate 3D model from the 2D RGB image.Main researches in this thesis are summarized as follows:(1)Summarizing and analyzing the three-dimensional reconstruction method of input two-dimensional images: The three-dimensional reconstruction method of inputting two-dimensional images and outputting three-dimensional models is summarized.On the one hand,this thesis summarizes the problems existing in the method of single-view 3D reconstruction,and analyzes the advantages and disadvantages of the existing methods.On the other hand,the problems existing in the three-dimensional reconstruction method using multiple views are summarized,and the advantages and disadvantages of the existing methods are analyzed.In addition,the methods for exploring novel three-dimensional representations are summarized.(2)Proposing three-dimensional fusion hierarchical reconstruction method for any number of views and comparing the performance indicators and visualization results with existing methods: Aiming at the lack of information in single-view 3D reconstruction and some problems in existing multi-view 3D reconstruction,this method uses a feature combination method that treats each input image equally and a hierarchical prediction strategy for improving the reconstruction effect of the thinner parts of the object.The network model can receive any number of images as input and obtain the three-dimensional reconstruction results,and as the number of input images increases,the reconstruction results become more and more accurate.(3)Proposing multi-view 3D reconstruction method based on autonomous view selection and comparing the performance indicators and visualization results with existing methods: Aiming at the lack of information in single-view 3D reconstruction and some problems in existing multi-view 3D reconstruction,this method uses the view selector in the network to autonomously select the next input view according to the current reconstruction state and update the reconstruction state.It overcomes the problems of insufficient view information and unstable reconstruction results caused by randomly selecting views in the existing methods,and obtains more accurate reconstruction results.