Structural Attention Convolutional Neural Networks Methods For 3D Point Clouds Classification

3D point clouds classification is one of the fundamental yet crucial topics of 3D vision research.So far,many deep neural network-based methods have emerged,showing powerful performance advantages of deep learning.But some recent methods only focus on the perception of local or global almost,yet ignore the structural hierarchy of point clouds.However,the structural hierarchy of point clouds is crucial for classification performance.In this thesis,by means of the powerful learning and approximation ability of deep neural networks,convolutional neural networks methods for3 D point clouds classification will be designed.Specifically,these methods mainly focus on the effective extraction and utilization of structural attributions of 3D point clouds,with a good trade-off between accuracy and efficiency simultaneously.Our major works are as follows.1.An object decoupling network for 3D point clouds classification is designed.We propose an operator to measure the point feature energy.By measuring the energy response of each point in point clouds after initial feature extraction,the point clouds are decoupled into several structural sub-blocks.After that,a criterion of structural affinity is proposed to evaluate the spatial relationships among the points in sub-blocks,and a selfattention model is established to study the impact of the structural affinity for feature interaction.A great number of experimental results verify the effectiveness of the network for point clouds classification,especially the distinction of resemblance categories.2.A progressively aware multiscale network is constructed to classify the point clouds objects.We point out that not only the multi-scale point clouds resolution,the receptive field varying of the convolution operator should also be concerned.This progressive architecture endows the network extracting more scales of point clouds feature information,ranging from the small local detail scale to the global scale of the entire point clouds outline.In addition,a regional convolution operator is constructed to fuse the distribution information and geometric structure attributes of point features among the region.The experimental results show that this network could reach excellent classification performance,especially for real-world point clouds objects.3.A lightweight binary tree-type network for 3D point clouds classification is established,which presents a method to trade off accuracy and network complexity.Specifically,the network realizes the capability of down-sampling the point clouds and abstracting their shape-structural features by means of organizing the point clouds into the structured form of a binary tree,and then the lightweight is realized.On the basis above,in order to further mine the structure information extracted by each layer of the network,a structural feature interaction module is designed to alleviate the information degeneration phenomenon during the feature abstraction process.Experiments on classification datasets show that the network could reach a competitive performance while being lightweight.