Research On Deep Learning Method Of 3D Point Cloud Classification Based On Graph Convolution And Attention

With the wider use of 3D point clouds,the demand for object representation and classification technology based on 3D point clouds has become larger and larger,and it has also promoted the development of 3D point cloud classification and recognition.3D point cloud classification has been widely used in robot navigation and scene understanding.However,existing deep learning methods for 3D point cloud classification usually suffer from insufficient feature extraction and ignore the global context relation of point clouds.And many neural networks used for point clouds tend to ignore the attention of important regions on point cloud objects,resulting in some useful information not being extracted,which leads to the problem of low classification accuracy.Therefore,the paper mainly focuses on obtaining global features and contextual features of point clouds and focusing on important areas and useful information of 3D point cloud,focusing on designing a classification method for 3D point cloud.The specific research work is as follows:(1)In view of the problems of insufficient feature extraction and neglect of global context in existing 3D point cloud classification methods based on deep learning,a 3D point cloud classification method(DAM-Point Net VLAD)based on dual attention mechanism was proposed.By using a dual attention network to extract local structural features as well as global context information,the graph attention mechanism is mainly used to learn the local ensemble representation of 3D point clouds,and the self-attention mechanism can check the spatial relationship between all points,thereby mining the contextual global information of the 3D point cloud.And through the VLAD layer to indirectly describe the high-level semantic information of each point by associating its geometric descriptor with visual words.The experimental results show that the DAM-Point Net VLAD network framework proposed in this chapter achieves a classification accuracy of 92.69% on the Model Net40 dataset.(2)In view of the limited attention to some important areas on the point cloud in the 3D point cloud classification method based on deep learning,some useful information in the neighborhood features is usually ignored when aggregating features,a 3D point cloud classification method(Att-Adapt Net)based on attention and adaptive graph convolution is proposed.The adaptive convolution kernel is mainly combined with the graph convolution to form the adaptive graph convolution to obtain the global features of the 3D point cloud,and the global attention mechanism is used to calculate the attention mask of each point,which is combined with the self-Adaptive graph convolution is combined to focus on feature regions and channels,and the Att-Adapt Net network is constructed.Experimental results show that the model achieves a classification accuracy of 93.3% on the Mode Net40 open dataset.