Point Cloud Classification Based on Multi-scale Fusion and Residual Dual Attention
Due to the disorder,sparsity,and limited information of point clouds,deep learning methods find it difficult to fully capture the complex correlations between points,resulting in limited classification accuracy and poor robustness.A new point cloud classification method DGCMR is proposed for this purpose,including using multi-scale dynamic graph convolution to fuse features from multiple scales,extracting more refined and comprehensive features,construct a dual attention module with residual structure,enhancing useful features and suppressing redundant information on important channels and key local neighborhoods,improving feature expression ability,and solving network degradation problems,combining the results of maximum pooling and average pooling to compensate for the information loss caused by single pooling,and obtaining the final classification result through a fully connected layer.The experimental results show that the overall accuracy of the model on the datasets ModelNet40 and ModelNet10 reaches 93.7％and 95.0％,respectively,with stronger robustness and better performance than current advanced methods.In terms of parameter generation,the model has decreased by 52.3％and 7.9％compared to PointNet and DGCNN,respectively,achieving better lightweight results and can be better applied to embedded 3D scanning devices.
point cloud classificationrobustnessmulti-scale featuresdouble attentionfeature enhanceme
万方数据
