关键词: 点云, 分类分割, 深度学习, 注意力机制, 特征融合

Abstract: Existing networks are difficult to effectively learn the local geometric shape information of point clouds, and there are problems such as being unable to effectively focus on important feature structures and insufficient fusion. Therefore, a point cloud classification and segmentation network based on dual attention mechanism and multi-scale fusion was proposed. Firstly, in the data feature extraction stage, the geometric position and weight of the convolution kernel are dynamically adjusted using Geometric Adaptive Convolution (GAC), so that it can dynamically adapt to the local geometric structure of point cloud data, thereby more effectively capturing local features. Secondly, in order to further improve the feature expression ability, the Dual Attention Mechanism (DAM) was introduced to automatically learn and adjust the weights of feature channels and spatial information, thereby enhancing the feature representation of key points. Finally, feature information of different scales was connected for effective fusion to enhance the feature learning effect, making the final feature representation richer and improving the classification and segmentation accuracy of the network. Experimental results on ModelNet40, ShapeNet and S3DIS datasets show that the proposed network improves the Overall Accuracy (OA) and mean Intersection over Union (mIoU) compared with PointNet++ and Dynamic Graph Convolutional Neural Network (DGCNN), and effectively improves the performance of point cloud classification and segmentation.

Key words: point cloud, classification and segmentation, deep learning, attention mechanism, feature fusion