Research On 3D Object Reconstruction Technology Of Single View Based On Deep Learning

3D object reconstruction has been widely used in a variety of scenarios such as autonomous driving,3D printing,virtual simulation and VR games,involving the fields of biology,neuroscience,ecology and agriculture.With the development and application of deep learning and the establishment of large-scale 3D object reconstruction datasets,the use of deep neural networks to recover the 3D shape of objects from a single image has attracted more and more attention.Existing deep learning methods have achieved varying degrees of success by using different geometric representations and different deep neural network frameworks for single-view 3D object reconstruction.However,these works have different degrees of deficiencies in the reconstruction accuracy and the generalization ability of different categories of object reconstruction.Based on the existing research work,this thesis continues to study the single-view 3D object reconstruction technology based on deep learning.The main work is summarized as follows:(1)For objects with simple topology(aircraft,cars,benches,sofas,etc.),we summarize and propose a depth neural network model based on free form deformation method.The model first uses the encoder-decoder-predictor structure to extract the image control point matrix from the input image,then an adaptive model dimension reduction network based on Point SIFT is used to extract the model control point matrix from the template model.Finally,the free form deformation method is used to deform the template model,so as to obtain the 3D model similar to the feature of the object in the image.(2)For objects with complex topology(such as tables and chairs,etc.),we propose a skeleton modification network model based on the combination of central skeleton and free-form deformation method.Firstly,the skeleton reconstruction network is used to generate a 3D rough point cloud skeleton through a single image,and then the point cloud skeleton is used as an automatically generated template model.The control point matrix is extracted from the image and the point cloud skeleton through different feature extraction networks.Finally,the free form deformation method is used to modify the point cloud skeleton in detail,thereby achieving high-precision point cloud reconstruction of complex topological objects.(3)A single view 3D object reconstruction software based on Py Qt5 is developed.The software can reconstruct the point cloud of the object in the image through a single image.Under the condition of providing mask,it supports the 3D reconstruction of the real scene,and provides the relevant operation of reconstruction visualization.In the comparison experiments on the public dataset,the method proposed in this thesis has achieved good results in point cloud reconstruction accuracy and generalization ability,which proves its effectiveness.