Research On Single-view 3D Mesh Reconstruction Algorithm Based On Deep Learnin

3D reconstruction is one of the core research directions in the field of computer vision,aiming to reconstruct 3D models of people,objects or scenes that can be recognized,and it is of great significance to the implementation of technologies such as augmented reality,mixed reality,robot navigation,autonomous driving and virtual surgery.Single-view based 3D reconstruction can be adapted to real-world reconstruction with lower input costs and higher universality,especially suitable for those scenes where it is difficult to obtain the input information of the object to be reconstructed comprehensively.In single-view 3D reconstruction,it is common to use triangular meshes as the data carrier for the reconstructed object,which can improve the quality of the final reconstruction by taking advantage of its curvilinearity,scalability and richness of shape detail.However,there are still some problems with the practicality and generalizability of using mesh carriers.For example,the topology of the mesh is limited by predefined templates and is not flexible enough,resulting in unnecessary mesh self-intersections and connections when reconstructing complex topologies,thus destroying the surface details;the performance of the deep learning network used for reconstruction is limited by the choice of training method,which cannot take into account the surface fineness of the reconstructed model,the training efficiency and the generalization performance of the network.This paper therefore addresses each of the above-mentioned shortcomings as follows:(1)To address the problem that the reconstructed model is limited by predefined topology,a single-image 3D reconstruction method based on Graph Convolutional Network(GCN)and topology modification is proposed,which guarantees the generation of high-quality mesh surfaces through GCN and improves the flexibility of topology using topology modification techniques.Meanwhile,unlike previous methods that simply vectorize the image,this study proposes a new feature fusion method that uses features at different stages of image convolution,extracts the features of each vertex at the corresponding point on the feature map using bilinear differences and fuses them into the GCN,and uses the multilayer perceptron MLP to fuse the final feature vector and predict the topological error of the mesh as a way to improve intermodule compatibility.In addition,when training the reconstructed network based on the publicly available dataset Shape Net,a new weight term was added to the corresponding loss function in the expectation of focusing more on the backbone of the meshes during training and speeding up the training of the model.(2)To address the problem that the performance of the reconstruction network is limited by the choice of training methods,a multi-dimensional training algorithm for 3D texture mesh reconstruction(P2M2P)is proposed.The method combines the advantages of 3D supervision and 2D supervision by training the 3D reconstruction network in different dimensions in stages.The training efficiency and generalization performance of the reconstruction network are further improved without destroying the overall framework and surface details of the model.At the same time,this study introduces a microscopic rendering technique to solve the problem of non-microscopic 2D supervised real-time rendering.The simulation experimental results show that the method of this study achieves a good balance in the indexes of surface fineness of the reconstructed model,training efficiency of the reconstructed network,and adaptation to the type of reconstruction.