LG-Former: Local-Global Dual-Branch Structural Transformer for Point Cloud Classification

To address the issues of insufficient local and global feature extraction, limited geometric structure modeling capability, and inefficient multi-scale information fusion in point cloud classification, this paper proposes a novel classification network (LG-Former) that integrates a local-global interaction framework with a structural Transformer. The core innovations of the method include: 1) Designing a dual-branch interaction architecture, where the local branch employs a multi-scale graph convolutional feature aggregation module (LFA) to enhance geometric structure modeling and multi-scale fusion, while the global branch strengthens global representation through average pooling and a position-aware structural Transformer (iteratively updating positional information and feature interactions); 2) Coordinating dual-branch features to drive classification decisions. Experiments on the ModelNet40 and ScanObjectNN datasets demonstrate that the algorithm achieves overall accuracy (OA) rates of 93.8% and 88.6%, respectively, significantly outperforming mainstream methods. Ablation studies validate the effectiveness of the LFA module, dual-branch interaction, and structural Transformer. Complexity analysis confirms that the model achieves a balance between computational efficiency and performance. This research proves that jointly optimizing local geometric modeling, global structural perception, and feature interaction mechanisms can significantly improve the accuracy and robustness of point cloud classification.