Research And Application Of Point Cloud Multi-scale Classification Network

The point cloud is a data representation form of 3D computer vision.At present,the classification technology of point cloud is very cutting-edge research in the field of computer vision.Existing point cloud classification methods based on deep learning are not very effective in feature learning of point cloud targets because they ignore high-frequency information and pay little attention to local features,resulting in low classification accuracy and poor system robustness.In this paper,the public data set is used as experimental data to study the point cloud classification problem,and two classification methods are proposed.First,the point cloud target classification problem based on random Fourier feature mapping is studied.The end-to-end point cloud classification model determines the classification result of a target according to the features extracted by the multi-layer perceptron of the target,but cannot evaluate the target category according to the high-frequency features of the target.Therefore,this paper proposes a neural network algorithm based on random Fourier feature mapping to improve the accuracy of target classification.In this method,the key high-frequency features are extracted,and the features extracted based on Gaussian mapping are combined with those extracted by the neural network,so that the fused features have more effective information,to improve the accuracy of point cloud target classification.The experimental results show that the high-frequency information of the point cloud is effective and feasible in improving classification accuracy.Secondly,a point cloud classification Network BFNet(Blending Feature Pyramids)is proposed to improve the robustness of classification.Firstly,the basic Feature extraction Network is constructed,and the features of different levels are obtained and fused with FPN(Feature Pyramid Network)structure.Then,the full connection layer corresponding to dimension input is used for classification.Secondly,according to the structure of the Blending algorithm,five parallel classification networks were set up as classifiers using the basic network.Finally,a result fusion layer is set up to fuse the results of five classifiers to improve the accuracy and robustness of classification.The BFNet is trained and tested by using the Model Net40 object classification data set.Experiments show that the accuracy and robustness are improved compared with Point Net and Point Net++.Thirdly,a point cloud classification system based on the above two classification methods is designed and implemented.Firstly,the requirements of the system are analyzed,and then the overall framework of the system is designed according to the requirements,and the language and tools for development are determined.Specific functional modules include the user management module,point cloud data upload and preprocessing module,classification network prediction module,and result display module.Finally,the main interface and four-module interface of the system is displayed in the form of an application.