基于Dental-Net的口内3D扫描点云补全  

Point Cloud Patching from Intraoral 3D Scanning Based on Dental-Net

作　　者：李杰[1] 陈胜[1] 左小五[1] 

机构地区：[1]上海理工大学,上海

出　　处：《软件工程与应用》2023年第1期68-77,共10页Software Engineering and Applications

摘　　要：目的:为了解决口腔内3D扫描由于受口腔内复杂环境的影响(牙釉质、口水、唾液、金属牙、牙齿填充物),导致重建的3D点云存在大量缺失孔洞的问题。方法:本文提出一种基于真实口内3D点云的特征信息,补全3D扫描过程中部分缺失点云的Dental-Net深度学习算法模型,实现了口腔牙齿3D精细化完整建模。Dental-Net整体为编解码器架构,编码器处的特征提取模块负责对不同分辨率下的牙齿点云进行特征提取,将提取到的特征输入到解码器,解码器依序输出补全的缺失点云,过程中结合使用了生成对抗网络思想。结果:使用评估指标倒角距离(Chamfer distance, CD)对最终补全结果进行评价,本文提出的方法在口内3D扫描点云数据集上CD = 0.292。结论:最终点云补全结果说明了本文模型具备精确的细节补全能力和不同缺失条件下的泛化能力。Objective: To solve the problem of missing holes in 3D point cloud reconstruction due to the influence of complex oral environment (enamel, saliva, metal teeth, tooth fillings). Methods: This paper proposed a Dental Net depth learning algorithm model based on the feature information of the real 3D point cloud in the mouth to complete part of the missing point cloud in the 3D scanning process, and realized the 3D fine and complete modeling of oral teeth. Dental Net is an overall codec architecture. The feature extraction module on the encoder side is responsible for feature extraction of tooth point clouds at different resolutions, inputs the extracted features to the decoder, and the decoder outputs the missing point clouds to complete step by step. In the process of completion, the idea of generation countermeasure network is combined. Results: The evaluation index chamfer distance is used to evaluate the final completion result. By the method proposed in this paper CD = 0.292 on the 3D scanning point cloud data set in the mouth. The final point cloud completion results show that the model in this paper has accurate detail completion ability and generalization ability under different missing conditions.

关 键 词：3D扫描 深度学习 生成对抗网络 倒角距离 点云补全 

分 类 号：R78[医药卫生—口腔医学]