Research On Point Cloud Classification Network Based On Attention Mechanism And Multi-layer Feature Fusion

With the continuous development of deep learning,point cloud data,like image data,has gradually become a basic data of deep learning.The classification of point cloud data is the key technology for the acquisition and three-dimensional reconstruction of point cloud data information,which has a wide range of applications,because point cloud data has the characteristics of disorder and irregularity,making it challenging for point cloud to carry out high-precision classification tasks.After continuous development,point cloud classification has completed the classification of point cloud data based on traditional methods to the existing classification of point clouds using deep learning methods,but the existing point cloud classification methods based on deep learning have problems such as insufficient extraction of local information of point clouds,lack of attention to point cloud channels and spatial feature information,ignorance of the fusion of multi-layer features in the global context of point clouds,and failure to pay attention to the two-dimensional feature information of point clouds when extracting point cloud features.In view of the above problems,the main research work is as follows:(1)Aiming at the problems of missing the extraction of local features of point clouds,ignoring global context features,and not paying attention to point cloud channels and spatial feature information in the existing three-dimensional point cloud classification network model based on deep learning,a point cloud classification network model based on graph convolution and fusion attention mechanism is proposed.Firstly,the three-dimensional point cloud is regarded as a node on the graph,the KNN algorithm is used to compose the map,and the information between points is dynamically captured by stacking multiple graph convolutional layers to strengthen the local feature extraction ability,and then,the fusion attention mechanism module is introduced to extract the point cloud channel and spatial feature information,and finally,the point cloud global context feature is obtained by multilayer feature fusion.Experimental results show that the network model achieves 92.5%classification accuracy on the Model Net40 dataset.(2)Aiming at the lack of extraction of two-dimensional feature information of point cloud in the existing three-dimensional point cloud classification network,a point cloud classification method 3D CLIP based on three-dimensional point cloud projection image is studied and designed with the help of two-dimensional CLIP experience.This method extracts the text description features of the point cloud through the text encoder,extracts the two-dimensional projection image features of the point cloud through the image encoder,and classifies the point cloud features by calculating the maximum similarity of the relevant features.The experimental results show that adding 3D CLIP to different point cloud classification network models can improve the classification accuracy of the network,and3 D CLIP can be used as a plug-in for the 3D point cloud classification network model to make up for the lack of point cloud two-dimensional feature information extraction in the3 D point cloud classification network,so as to improve the classification accuracy of the network.(3)For the two proposed point cloud classification network models,the GFANet network and the 3D CLIP network can be fused to obtain better point cloud classification results.The point cloud classification method of GFANet plus 3D CLIP can not only make the network pay attention to the global context feature information of the point cloud and strengthen the local feature extraction capability of the network,but also enable the network to extract the two-dimensional feature information of the point cloud and improve the classification accuracy of the network model.The experimental results show that the point cloud classification method of GFANet plus 3D CLIP achieves a classification accuracy of93.6% on the Model Net40 dataset,indicating that the point cloud classification method of GFANet plus 3D CLIP can significantly improve the classification accuracy of the 3D point cloud classification network.