Point Cloud Classification And Segmentation Methods Based On Local And Nonlocal Features

At present,deep learning has shown great potential in 3D point cloud classification and segmentation,which has attracted the extensive attention of many scholars.In particular,all kinds of convolutional neural networks bloom.However,based on the irregularity and disorder of point clouds,most deep learning methods still face great challenges in designing convolutional neural networks to extract perfect point cloud features.This dissertation is mainly aimed at the task of 3D point cloud classification and segmentation,based on the existing methods of point cloud feature extraction and learning in convolutional neural networks,a series of novel and efficient convolutional modules are designed.In turn,rich point cloud features are extracted and the performance of 3D point cloud classification and segmentation tasks is improved.The main research works are summarized as follows:1.A local and nonlocal interactive convolutional neural network for3 D point cloud classification is proposed.The innovation of this work is that the local and nonlocal convolution module is designed to improve the feature extraction ability of the point cloud.Firstly,a local and nonlocal interactive convolution module is used to obtain local similar features and nonlocal similar features.Then,an interactive augment operation is designed to alleviate the redundancy problem of a single neighborhood when representing a closed area,augment the hierarchy and stability of the network,and alleviate the degradation of the network at the same time.Then,the convolutional neural network is built with this module as the basic unit.In addition,an adaptive feature fusion module is constructed to fully utilize different hierarchical features,so as to realize the classification of3 D point cloud.Experimental results show that the proposed method can accomplish the task of point cloud classification excellently.2.A local self-augment convolutional neural network for 3D point cloud classification and segmentation is constructed.The innovation of this work is that a local self-augment convolution module and a local aggregation module are proposed to slow down the translation transformation of the point cloud.Specifically,firstly,the point cloud is represented by the graph structure to avoid information loss while keeping the number of point clouds unchanged.At the same time,thanks to the graph structure,the local and global features can be learned.Then,a local self-augment convolution module is designed to reduce the translation variance of the point cloud in the graph structure by relying on the characteristics of the module.Moreover,a local aggregation module is designed in the graph structure to combine the overview and details in the graph neighborhood.Experimental results show that the proposed network can produce good results in classification and segmentation tasks.3.A serial attention convolutional neural network for 3D point cloud classification is designed.The innovation of this work is that a neighborhood attention convolution module and a multi-channel attention module are proposed.Specifically,firstly,the neighborhood attention convolution module is designed to represent the point cloud through the graph structure.Then,a multi-spatial index is proposed to optimize the selection of neighborhood points in the graph structure to avoid the limitations of a single spatial form.At the same time,the score weight is calculated through the attention mechanism,and then the neighborhood features are aggregated to overcome the disorder of point cloud and smooth the structural information in the local neighborhood.Then,a multi-channel attention module is designed to reduce the channel redundancy in the embedded space and improve the fusion and discrimination of features.Finally,a serial attention convolution neural network is built to realize the task of 3D point cloud classification based on these two modules.Experimental results show that the designed network can effectively improve the performance of classification.