Research On 3D Reconstruction Of Single Image Based On Deep Learning

3D reconstruction is always one of the main research directions of computer vision,and how to get 3D model from single image reconstruction is a problem that scholars have been exploring.To solve this problem,the most ideal goal is to be able to simulate the human visual experience,associate from the single RGB image,infer from the prior knowledge obtained in the whole life,and get the appropriate three-dimensional model.With the development of deep learning and the improvement of 3D CAD data set,the 3D reconstruction method based on deep learning has made a breakthrough.Aiming at the problems of two-dimensional feature extraction and feature transformation in deep learning three-dimensional reconstruction,this paper proposes a solution of multi-scale feature extraction and RNN integration from the three parts of encoder,transcoder and decoder,and at the same time,it puts forward the judgment of edge vector loss function to improve the three-dimensional reconstruction loss,and finally gets better reconstruction results.First of all,for single image 3D reconstruction,the most important thing is how to extract the features of 2D image effectively.In order to solve this problem,this paper analyzes the existing deep learning 3D reconstruction framework,designs a multi-scale3 D reconstruction network based on the encoder decoder network,and uses the multi-scale convolution neural network as the 2D feature extraction module to obtain more 2D image details.Then,the original designed multi-scale three-dimensional reconstruction network is analyzed by experiments.For the problem of multi-scale feature integration,this paper proposes an improved three-dimensional reconstruction network based on RNN,which integrates the feature matrix using the characteristics of RNN network retention and forgetting.In order to determine the loss and back propagation of 3D mesh data better,this paper considers the loss of edge vector,proposes the loss function of Edge Vector,and realizes the error determination of 3D mesh points,edges and faces.Finally,we use TensorFlow framework to train the 3D reconstruction network designed in this paper under the public ShapeNet data set.Compared with the existing3d-r2n2,PSG net,pixel2 mesh and other 3D reconstruction networks under the three standards of Chamfer distance,F-score and IOU value,it is proved that this paper can achieve better 3D reconstruction effect for most models of aircraft,cabinet,automobile,display and so on.