FDA-PointNet++: A Point Cloud Classification Model Based on Fused Downsampling Strategy and Attention Module

In recent years, the use of deep learning models for point cloud classification and segmentation tasks has increasingly become a hot topic in 3D point cloud research. However, the sparsity and inhomogeneity of point cloud data make it difficult to extract point cloud features. Meanwhile, how to effectively extract fine-grained local features becomes crucial in point cloud understanding. Therefore, in this study, we propose a novel FDA-PointNet+ + point cloud classification model based on fusion downsampling strategy and attention module. Firstly, the method proposes a fusion downsampling strategy, which performs hierarchical downsampling on the initial point cloud data, and then repeats the downsampling operation on the sampling results and performs feature fusion to form feature maps with multi-scale information to enhance the richness of local spatial point cloud feature information. Secondly, we incorporate a channel attention mechanism into PointNet+ + and propose a Local Feature Aggregation (LFA) module for point cloud local feature extraction. This method improves the local feature extraction capability of the network model by amplifying the relevant local features and suppressing the non-relevant features. Experimental results on the ModelNet40 dataset demonstrate that FDA-PointNet+ + achieves higher classification accuracy and robustness, with a 1.3% increase in overall accuracy (OA) and a 1.4% improvement in class accuracy.