基于PointCloudTransformer和优化集成学习的三维点云分类

Point cloud classification based on PointCloudTransformer and optimized ensemble learning

针对三维点云的不规则性和无序性所导致的难于提取特征并进行分类的问题,提出了一种融合深度学习和集成学习的三维点云分类方法。首先,训练深度学习点云分类网络PointCloudTransformer,并使用主干网络提取点云特征,进而训练基分类器,获得基分类器集合;然后,针对集成学习算法设计基分类器选择模型,模型的优化目标为基分类器组合的差异性和平均总体精度。为了降低集成规模,本文基于增强后的白鲸优化算法提出了二元多目标白鲸优化算法,并使用该算法优化基分类器选择模型,获得集成剪枝方案集合;最后,采用多数投票法集成每个基分类器组合在测试集点云特征上的分类结果,获得最优基分类器组合,从而构建基于多目标优化剪枝的集成学习点云分类模型。在点云分类数据集上的实验结果表明,本文方法使用了更小的集成规模,获得了更高的集成精度,能够对多类别三维点云进行准确分类。

Aiming at the difficulty of extracting features and classifying 3D point clouds due to their irregularity and disorder,a 3D point cloud classification method that fuses deep learning and ensemble learning is proposed.Firstly,the deep learning model PointCloudTransformer is trained to extract point cloud features and train base classifiers to establish a base classifier set.Subsequently,a base classifier selection model is designed for ensemble learning,and optimization objectives of the model are diversity and average overall accuracy of base classifiers.To reduce ensemble scale,binary multi-objective beluga optimization algorithm based on improved beluga optimization algorithm is proposed to optimize the base classifier selection model and obtain an ensemble pruning scheme set.Finally,the majority voting is used to ensemble the classification results of each base classifier combination on the test set to obtain the optimal base classifier combination,and an ensemble learning model of point cloud classification based on multi-objective optimization ensemble pruning is obtained.Experimental results on the point cloud classification dataset demonstrate that the method in this paper achieves higher ensemble accuracy with a smaller ensemble scale and can accurately classify multi-class 3D point clouds.