Surface Reconstruction Algorithm And Parallelization Research Based On Deep Learning

When the traditional reconstruction algorithm is used for surface reconstruction,the precision and integrity of 3d surface reconstruction will be seriously affected by the quality of point cloud data obtained.When 3d surface reconstruction is carried out,the point cloud data containing the spatial information of the object needs to be acquired first.However,when obtaining the point cloud data on the object surface,the object to be reconstructed may be affected by factors such as occlusion or the defects of the object itself,resulting in the absence of the obtained point cloud data and the appearance of holes on the reconstructed 3d surface.Holes need to be repaired after the reconstruction is completed.In recent years,deep learning has been widely applied in image recognition,target detection and other aspects and achieved good results.This paper introduces the target detection method in deep learning to improve the accuracy and integrity of surface reconstruction.As one of the research directions of computer vision,3d surface reconstruction based on deep learning has been developing rapidly in recent years.Different from the traditional point cloud-based reconstruction method,this paper adopts multi-view dataset for surface reconstruction.In the process of surface reconstruction,CNN convolutional neural network is used to extract feature point clouds of different views,and multidimensional matrix is used to save point cloud data.The network is used to match feature points of different views,and Chamfer algorithm and EMD algorithm are used to improve the matching efficiency and accuracy of point cloud data.Through 3DConv in the network,the three-dimensional point cloud space mapping is carried out and the three-dimensional point cloud model is reconstructed.In order to speed up the computing problems in the process of surface reconstruction,CUDA,a parallel computing technology,is used to optimize the surface reconstruction algorithm,which reduces the time consumption of surface reconstruction and improves the efficiency of surface reconstruction.The main work of the paper is as follows:1.For the surface reconstruction model,this paper applies the convolution network to extract the feature point cloud.Aiming at the feature point cloud corresponding to different views,the network is used to learn the distance between different points cloud and measure the similarity of different points cloud.The learning distance is constrained by the standard threshold distance,and the point cloud smaller than the threshold distance is superimposed into the initial point cluster to form a complete 3D point cloud model.2.In order to improve the accuracy of point cloud matching,this paper introduces the Chamfer distance and EMD algorithm to measure the distance between different cloud distributions,so as to make the smoothness and texture of the reconstructed surface closer to the real surface,and make the integrity of the reconstructed surface reach more than 99%.Compared with traditional point cloud reconstruction,when a small number of views(16 views)are used for surface reconstruction,no holes appear in the reconstruction results based on this method.3.For the time-consuming part in the reconstruction algorithm,this paper uses GPU technology for acceleration training,and uses CUDA programming model to optimize the surface reconstruction algorithm.The average reconstruction time and acceleration ratio before and after using CUDA technology were compared.Experimental results show that the multi-view surface reconstruction algorithm based on CNN convolutional neural network proposed in this paper can reconstruct 3d surfaces with high integrity.The prediction function of convolutional neural network is used to reconstruct the surface without holes.CUDA parallel computing technology is adopted to accelerate the surface reconstruction speed and effectively save time cost.