提出一种基于嵌入特征优化的口腔点云异常区域分离的方法，去除每个点云中的异常部分，保留准确的部分。首先，采集2 000个口腔点云数据，并标注出需要去除的异常区域，制作出用于训练、验证和测试的数据集；然后，使用基于期望最大化算法的迭代优化方法及对比学习的思想，在点云降噪模型上进行特征优化改进，训练200次，并在测试集上验证结果；最后，以基于平均交并比R _mIoU的定量分析和基于点云可视化的定性分析作为评价指标与几种常见的点云分割和点云降噪模型进行对比。结果表明：基于嵌入特征优化改进的方法的R _mIoU与其他方法相比提高了1.08%，对口腔点云异常区域分离这一问题效果更好。

A method for separating abnormal regions in oral point clouds based on embedded feature optimization is proposed， where abnormal parts of each point cloud are removed while accurate parts are preserved. Firstly， 2000 oral point cloud datasets are collected， and the abnormal regions to be removed are labeled， creating datasets for training， validation， and testing. Subsequently， an iterative optimization approach based on the Expectation-Maximization （EM） algorithm and the concept of contrastive learning is employed to enhance feature optimization in a point cloud denoising model， with training conducted over 200 iterations and results validated on the test set. Finally， quantitative analysis based on mean Intersection over union R _mIoUand qualitative analysis based on point cloud visualization are used as evaluation metrics to compare the proposed method with several common point cloud segmentation and denoising models. The results demonstrate that the R _mIoU of the method improved by 1.08% compared to other approaches， indicating superior performance in separating abnormal regions in oral point clouds.