Research On Point Cloud Classification Based On Deep Learning

With the popularization and application of 3D data acquisition equipment,the acquisition of 3D data is more convenient.Three-dimensional data contains rich geometric,shape and scale information.How to classify,segment and recognize 3D data effectively and accurately is a research hotspot in the field of computer vision.This paper focuses on the analysis of the working principle of deep learning network models based on different 3D model representation formats,and the advantages and disadvantages of feature extraction methods in the network structure.Aiming at the problems of the complicated training process and the lack of effective local information extraction methods in the existing deep learning network models,according to the unstructured characteristics of 3D data,a point cloud model classification network based on graph convolution neural network is proposed.The work and results of this article are:(1)Two methods for constructing the adjacency matrix required for graph convolution calculation are proposed.They are the method of constructing adjacency matrix based on K-nearest neighbors and the method of constructing adjacency matrix using grid information.The idea of constructing the adjacency matrix method based on K nearest neighbors is: for a vertex in a point cloud,k adjacency is used to construct the adjacency relationship of the vertex;applying this method to all the vertices to obtain the adjacency relationship of all the vertices,and construct the adjacency matrix by this;The idea of constructing the adjacency matrix method based on the grid information is: first construct an n*n(n is the number of vertices)zero matrix,then traverse the patches in the grid data,and establish the adjacency relationship by the vertex index in the patch.For vertices with adjacency relations,the corresponding weight is set to 1 in the zero matrix,and an adjacency matrix is finally constructed.(2)A classification network model is designed for the 3D data feature information extracted by graph convolution.First,the adjacency matrix and features are used as the input of the network.Secondly,the shallow and deep features are extracted by the graph convolutional neural network,and the shallow and deep features are merged into multi-scale feature vectors,and then the feature vectors on the graph are pooled.The globalization feature is used to obtain the global feature,and finally the corresponding classification probability is output through the fully connected classification network.The results of classification experiments on the ModelNet40 dataset show that the point cloud classification model proposed in this paper has the advantages of processing unstructured data and aggregating adjacent vertices to extract local information.Compared with existing point cloud classification models,it achieves higher classification The accuracy also significantly reduces the parameter size of the model.