Research On 3D Reconstruction Algorithm Based On Deep Learning

3D reconstruction is an important topic in the field of computer vision.3D model can help people to better understand the objects in the objective world.However,restoring the three-dimensional structure of an object from a two-dimensional RGB image is a typical inverse problem,which needs the combination of some prior knowledge of the image.In recent years,due to the powerful expression and learning capabilities,deep learning has achieved great results in image-related fields.In addition,it can be used to infer the three-dimensional structure of objects.There are many kinds of expression forms of3 D model in computer,they have their own characteristics and are suitable for different application scenarios.Among them,the representative ones are voxel and mesh model.Therefore,from the perspective of neural network,this paper designs a new voxel reconstruction model and mesh reconstruction model respectively.For the task of voxel reconstruction,this model first uses the image encoder based on vision transformer to extract the structural features of the image,then uses the decoding layer based on 3D transposed convolutional layers to escape the feature information to the voxel space.Finally,the output layer based on 3D vision transformer module and 3D convolutional module is designed to obtain the probability value of the voxel and output the voxel model with the resolution size of32 ×32 ×32.For multi-view tasks,this paper designs an attention module based on three-dimensional convolutional layers to fuse the voxel results of different image outputs.Unlike many fusion modules based on recurrent neural networks,this method can efficiently use images from different perspectives to complete reconstruction.The experiments show that the method in this paper can quickly recover the voxel of the object in single view and multi-view situations.The voxel can well reflect the whole structure of the original object.For the task of 3D mesh reconstruction,with the consideration of the diversity of object structures,this model first uses convolutional layers to generate a coarse3 D voxel model and then transforms the 3D voxel model into a triangular mesh model by using cubify method.Next,this model project the extracted two-dimensional features and three-dimensional features to the vertices of the mesh,then use the multi-layer graph convolutional neural network with residual connection and identity mapping to deform the triangular mesh.The results show that the proposed method can use a single image to restore the triangular mesh model with object texture,lines and other details.