| Point cloud hole repair in reverse engineering is one of the important research directions in the field of three-dimensional reconstruction in computer vision.It has been a rapid development trend in recent years and has become an important direction for scholars at home and abroad.Under normal circumstances,laser scanners are used to acquire point cloud data.The point cloud data obtained has high accuracy and high density.However,when acquiring point cloud data,due to the obstruction of the object model,incompleteness of the model itself,or inaccuracy of the normal information,the point cloud data is lost,which seriously affects the correctness of the three-dimensional reconstruction result.Therefore,how to efficiently rebuild complex object objects and scenes has become a difficult and hot issue in the current academic research.In addition,in view of the good effects and extensive application of deep learning in image recognition and detection targets in recent years,this paper introduces a deep learning image target detection method.In this paper,the octree tree data structure is used to complete the decomposition of the point cloud data from the whole to the local for the scattered scattered point cloud data,and then the hierarchical operation of the point cloud model is completed.Three-dimensional two-dimensional coordinates are obtained through spatial mapping.The CNN convolutional neural network is used to extract the hole features of the point cloud model.Then,based on the extracted hole features,corresponding sparse representation classification models are constructed,and the residuals of each category are calculated and generated.Point cloud boundary,get pure hole areas.And the filling of the polygon hole is completed.When using the multi-hidden CNN convolutional neural network to repair point cloud holes,the operation of the point cloud hole repair algorithm is low due to the large scale of the network and the complicated operation process.In addition,the calculation of a large amount of scattered point cloud data will cause an increase in the amount of computation.In order to improve the efficiency of the algorithm,combining the commonly used parallel computing technology,such as CPU acceleration,GPU acceleration and CPU + GPU hybrid parallel technology to optimize the point cloud hole repair algorithm,so as to quickly reduce the algorithm's running time and improve the algorithm's operating efficiency.The main work of the dissertation is as follows:1.For the scattered point cloud model,an octree-based data structure is used to divide the space of the scattered point cloud data model and store it into the octree data structure.2.According to the common point cloud hole repair algorithm,a point cloud hole repair algorithm based on the octree-based CNN convolutional neural network is proposed.The point cloud hole repair algorithm can effectively extract model features and adopt a polygon hole filling algorithm.The extracted point cloud holes are filled accordingly.3.Focusing on the hotspots in point cloud hole repair algorithm based on CNN convolutional neural network and combining multi-core CPU and GPU technology for optimization,this paper mainly adopts three optimization methods:OpenMP-based algorithm optimization,CUDA-based algorithm optimization and Hybrid algorithm optimization based on OpenMP+CUDA.4.Analyze and compare the performance of the algorithm before and after the optimization.This paper mainly analyzes the running time of the algorithm,the running efficiency of the algorithm,the acceleration ratio of the algorithm and the accuracy of the algorithm.Experiments show that the point cloud hole repair algorithm based on the multi-hidden layer CNN convolutional neural network proposed in this paper can well repair the voids on the surface of the point cloud model.In addition,using parallel techniques based on OpenMP and CUDA,the performance of the optimized algorithm is significantly improved. |
PDF Full Text Request