基于全局感知和局部增强的三维点云分类算法

【摘要】 由于点云数据具有无序性和离散分布的特点，传统点云分类方法在处理局部特征提取和全局信息建模方面存在一定的局限性。为此提出了一种融合全局感知和局部增强的三维点云分类网络GALE-Net(Global Awareness and Local Enhancement Network)，旨在提升点云分类的准确率。该网络由GA(Global Awareness)模块和LE(Local Enhancement)模块组成。通过融合多头自注意力机制和多层感知器构建GA模块，用于捕获全局上下文信息，提升了模型对长距离依赖关系的捕捉能力，有效增强点云的整体结构感知。此外，引入自适应的局部几何权重，并与采样与分组策略结合构建LE模块，能够对特征进行更细粒度的提取，专注于局部几何特征的细化，提高模型对局部区域的细致表达能力。实验结果表明，GALE-Net在ModelNet40数据集上取得出色性能，达到93.2%的总体准确率（OA）、91.9%的平均准确率（mAcc），且有着较高鲁棒性。更多还原

【Abstract】 This paper proposes a 3D point cloud classification network based on global awareness(GA) and local enhancement(LE), namely GALE-Net, for enhancing the classification accuracy of the 3D point cloud. The GA is developed to capture global contextual information using a multi-head self-attention mechanism and a multi-layer perceptron, thereby enhancing the model’s ability to recognize long-range dependencies and improving overall structural awareness of the point cloud. Additionally, the LE incorporates adaptive local geometric weights combined with sampling and grouping strategies to extract more fine-grained feature, which can refine local geometric features.更多还原