Research on Point Cloud Classification and Segmentation Algorithm Based on Improved Dynamic Graph Convolutional Network

TGraph Convolutional Networks (graph CNN, GCN) are widely used for segmentation and classification in point cloud technology. One of the most representative is dynamic graph CNN (DGCNN), but it only considers the deep model to extract global features without considering the problem of gradient vanishing caused by too deep model, which affects the accuracy and robustness of segmentation and classification. This paper proposes an improved dynamic graph convolutional network for point cloud segmentation and classification, which adds a residual convolution module (Residual Conv, Res) between the input and the output, and uses the operations of one-dimensional convolution and pooling combined with a multilayer perceptual machine to increase the extraction of the point cloud features while reducing the amount of computation; the feature extraction is enriched with the use of the design of the bottleneck structure, and the residual structure is used to provide a deep fusion of the local features and the global features are deeply fused, which significantly improves the performance of the network model. Experiments show that the classification accuracy is improved on the ModelNet40 standard public dataset, and the segmentation accuracy is improved on the S3DIS standard public dataset.