Towards Point Cloud Classification Network Based on Multilayer Feature Fusion and Projected Images

Learning (DL) based point cloud classification techniques now in use suffer from issues such as disregarding local feature extraction, missing connections between points, and failure to extract two-dimensional information features from point clouds. A point cloud classification network that utilizes multi-layer feature fusion and point cloud projection images is suggested to address the aforementioned problems and produce more accurate classification outcomes. Firstly, the network extracts local characteristics of point clouds through graph convolution to strengthen the connection between points. Then, the fusing attention mechanism is introduced to aggregate the useful characteristics of the point cloud while suppressing the useless characteristics, and the point cloud characteristics are fused by multi-layer characteristic fusion. Finally, a 3D point cloud network plug-in model based on point cloud projection image (3D CLIP) is proposed, which can make up for the defects of other 3D point cloud classification networks that do not extract two-dimensional information characteristics of point clouds, and solve the problem of low accuracy of similar category recognition in datasets. The ModelNet40 dataset was used for classification studies, and the results show that the point cloud classification network, without the addition of a 3D CLIP plug-in model, achieves a classification accuracy of 92.5%. The point cloud classification network with a 3D CLIP plug-in model achieved a classification accuracy of 93.6%, demonstrating that this technique can successfully raise point cloud classification accuracy.