Research On Deep Learning Model Of Point Cloud Classification With Attention Mechanism

With the increasing maturity of 3D laser scanning,airborne LiDAR and other technologies,the quality,acquisition efficiency and cost performance of 3D point cloud data have been significantly improved,providing a new expression for accurately describing the real 3D space.Handling presents huge challenges.The realization of point cloud classification and semantic segmentation based on deep learning methods is the basis of 3D scene understanding and semantic analysis.Due to the irregularity,non-uniformity and disorder of 3D point cloud data,how to effectively and accurately realize point cloud classification has become a challenging task.Based on the deep learning network models of Point Net and Point Net++,the paper improves the problems of insufficient local feature extraction and optimization of classification performance,and builds a deep learning model for point cloud classification that integrates attention mechanism,and analyzes the improved model through experiments.The effect of point cloud classification accuracy.The main research work is as follows:(1)Research on point cloud classification methods and convolutional neural networks.Different 3D point cloud data representation forms(multi-view,voxel,discrete point cloud)are studied,and the current 3D point cloud classification methods based on deep learning models are analyzed from the perspectives of indirect classification and direct classification.Secondly,the basic structure and characteristics of the convolutional neural network are studied,and the functions of the convolutional layer,multi-layer perceptron,activation function,pooling layer,classifier and other components,as well as the basic principle and calculation process of the attention mechanism are described.The network and main modules of feature extraction of the classic point cloud classification models Point Net and Point Net++ are analyzed.(2)Improve Point Net point cloud classification model based on dual attention mechanism.Aiming at the problem that deep learning point cloud classification ignores the correlation between high-level global single point and low-level local point cloud features,which leads to low classification accuracy,a dual-attention mechanism is proposed based on the Point Net model to build a DATPoint Net point cloud classification network model.The model realizes the feature weighting calibration in the channel dimension and the feature aggregation in the spatial dimension through the fusion layer of the double attention mechanism,deepens the deep information transmission of the network structure,mines the fine-grained features of the local area of the point cloud,and improves the overall performance of the network architecture.Compared with Point Net,the overall accuracy of the model using the fusion attention mechanism on the Model Net40 dataset is improved by 1.6%;in the S3 DIS indoor spatial dataset Area5,the block classification accuracy reaches 81.5%,which is 3.0%higher than that of Point Net.At the same time,the airborne LiDAR point cloud data provided by the International Association of Photogrammetry and Remote Sensing can obtain a more robust ground object expression effect,and improve the classification accuracy of the airborne LiDAR point cloud.(3)An end-to-end encoding-decoding point cloud classification network model is constructed based on Point Net++.The model fully combines the channel attention mechanism and the self-attention mechanism to design a feature extraction encoding and decoding module.The encoder and decoder can filter irrelevant and redundant information during propagation,and adjust the sampling strategy to realize the dynamic relationship between points.Update figure.The channel attention mechanism completes the exploration of the relationship between the deep point cloud features and reduces the problem of low classification accuracy caused by the useless channel feature learning process.On the other hand,a self-attention mechanism is designed and embedded to aggregate contextual information in the feature space,strengthen longdistance dependencies,more directly achieve global connectivity,and improve the integrity of the network architecture.Taking the airborne LiDAR point cloud data provided by Model Net40,S3 DIS and ISPRS as the experimental object,the validity of the proposed algorithm is verified by the accuracy analysis.The generalization experiment results based on the GML(B)dataset also show the feasibility of the improved algorithm.There are 35 figures,18 tables,and 71 references.