3D Reconstruction From Images Based On Deep Learning

Nowadays,3D reconstruction technology has attracted more and more attention from researchers in the computer vision,even our daily life is inseparable from the application of 3D reconstruction.For example,many applications such as face recognition,damage identification and prediction of buildings,restoration of ancient relics or archaeological sites need to be based on 3D images.Therefore,3D reconstruction is the most important link.3D reconstruction based on the depth information of the image has become a research highlight owe to its advantages of low cost and wide adaptability.In our approach,we use an indirect method to achieve the 3D reconstruction.First,we take advantage of deep learning to achieve the depth estimation of the monocular image.Subsequently,convert the obtained depth map to the point cloud model on the basis of the coordinate conversion relationship,so as to achieve our task of 3D reconstruction of the image.The accuracy of image depth directly affects the results of 3D reconstruction,so how to obtain high-quality image depth information is the focus of this paperIn the details of our method,firstly,we improve the framework of convolutional neural network,which remove the part of fully connected layer that used for classification,but use the fully convolutional networks to predict the depth of the monocular image.Secondly,we use the "up-sampling + convolution" method to enlarge the resolution of the output feature map,which overcomes the checkerboard artifacts of using deconvolution alone in traditional methods.Thirdly,we use supervised learning and unsupervised learning methods to estimate the depth of indoor and outdoor images.Meanwhile,we improve the encoder part of neural network and training sets,which proved that the deeper network structure and more datasets are helpful to enhance preciseness of the image depth estimation.After that,by combining supervised learning and unsupervised learning methods,we propose a semi-supervised learning method which further enhances preciseness of the image depth estimation.Besides,in order to ensure the depth estimation capability of the neural network,we added the long skip connection between the network layers,which effectively avoids the gradient dispersion phenomenon caused by the deepening of networks.Finally,according to the high-quality depth map that has been obtained in this paper,we convert the depth map into a 3D point cloud model to achieve the goal of 3D reconstruction based on images.Compared with other methods,the final improved results of this paper have achieved different degrees of advantages in both image depth estimation and 3D reconstruction.Our approach is more convenient and relatively easy to implement,the reconstructed 3D model also has better authenticity,which demonstrates the superiority and feasibility of our method.