Research On 3D Reconstruction Method Based On Autoencoder And Adaptive Octree Neural Network

3D reconstruction technology has been widely used in various fields such as industry,military,and medicine.Its application fields include 3D modeling,virtual simulation,non-contact measurement,scientific computing,augmented reality,and battlefield environment perception.In recent years of research and exploration,many scholars have developed various 3D reconstruction theories and implementation techniques,but high-efficiency and high-precision 3D reconstruction is still a challenging problem.In order to achieve large-scale,high-precision,and high-efficiency 3D reconstruction,it is necessary to establish a more efficient 3D reconstruction method than existing methods.At the same time,it must have an efficient storage format and computational efficiency,as well as good generation quality.In this dissertation,the method of deep learning is used to reconstruct the fast three-dimensional topography of objects.The main work contents are as follows:(1)Aiming at the problem of 3D shape reconstruction from a single 2D image input,a 3D reconstruction method based on auto-encoder is proposed.The encoder realizes the extraction of three-dimensional features,and at the same time,the two-dimensional image is extracted by two-dimensional convolutional neural network,and the parameters are optimized by the cross-entropy loss function and the L2 loss function at the same time,and finally the three-dimensional image is restored by the encoder.Compared with the current main PCA research methods,the algorithm in this dissertation can reconstruct a three-dimensional image in the case of a single image,and even restore the three-dimensional shape in the case of insufficient input.Our method achieves extremely high performance with an average precision(AP)of 97.6%and outperforms PCA with a 25%reduction in relative average error rate.(2)Aiming at the problems of high memory consumption and high computational cost of all-element convolutional neural networks,an octree-based convolutional neural network method is proposed for 3D shape reconstruction.The method is based on the octree representation of the 3D shape,takes the plane normal vector of the 3D model in the bottom layer of the octree as input,and performs 3D convolutional neural network on the octree nodes occupied by the surface of the 3D shape.network operations.Using the octree data structure,the octree node information and CNN features are efficiently stored into memory,and the entire OCNN training and testing is performed on the GPU.And by limiting computation to octree nodes occupied by 3D surfaces,memory consumption and computational cost are reduced.Compared with the original 3D voxel CNN method,the memory and computational cost are changed from N~3 to N~2,and the average chamfering distance is3.47,which achieves a better reconstruction effect.(3)Aiming at the problem that the reconstructed surface is not smooth enough,an adaptive octree convolutional neural network based on plane patch is proposed.The 3D shape of octree nodes with different levels can be adaptively represented,and the 3D shape within each octree node is modeled using plane patches.The adaptive octree convolutional neural network encoder takes plane patch normals and displacements as input and performs 3D convolution operations only on the octree nodes at each level,while the adaptive octree convolutional neural network decodes The processor infers the shape occupancy and subdivision state of the octree nodes at each level,and estimates the plane normal and displacement of each leaf node.The planar patch-based adaptive octree convolutional neural network not only reduces the memory and computational cost,but also has better shape generation capability than existing 3D CNN methods.In this dissertation,the shape prediction from a single image verifies the efficiency and effectiveness of the adaptive octree convolutional network generation task.The chamfering distance is 3.64,which is lower than the chamfering distance of Atlas Net of 4.46,and a better reconstruction effect is achieved.