摘要
针对复杂牙颌模型的三维网格结构特征相似、语义混淆等问题,提出一种基于共形映射参数化和DeepLabv3+的三维牙颌语义分割方法。首先,引入最小二乘共形映射,约束边界顶点并扩展颜色映射,实现牙颌网格的二维参数化与逆向重构;其次,设计基于DeepLabv3+的语义分割模型,采用MobileNetV2作主干网络,构建密集空洞空间金字塔模块,增强多尺度特征融合和特征学习能力;最后,引入条形池化模块增强线性结构感知,并通过批归一化提升训练稳定性和泛化能力。结果表明,在自建牙颌数据集上,平均交并比(mIoU)和平均像素准确率(MPA)达到78.71%和88.22%,较当前方法分别提升了3.61个百分点和3.15个百分点,模型参数量仅5.8M;在公共数据集Teeth3DS上的使用结果表明,mIoU和MPA达到了73.61%和82.05%,较未改进的DeepLabv3分别提升了10.42个百分点和11.15个百分点,具有更好的分割效果。引入最小二乘共形映射与优化的DeepLabv3+网络,可实现高精度、轻量化的三维牙颌自动分割。
To address the challenges of similar 3D mesh structural features and semantic ambiguity in complex dental arch models,a 3D dental arch semantic segmentation method based on conformal mapping parameterization and DeepLabv3+was proposed.First,the least squares conformal mapping(LSCM)was introduced to constrain boundary vertices and extend color mapping,enabling the 2D parameterization and inverse reconstruction of the dental arch mesh.Second,a semantic segmentation model based on DeepLabv3+was designed,employing MobileNetV2 as the backbone and constructing a dense atrous spatial pyramid module to enhance multi-scale feature fusion and representation learning ability.Finally,a strip pooling module was incorporated to improve the perception of linear structures,while batch normalization was applied to enhance training stability and generalization ability.The results show that the proposed method achieves a mean intersection over union(mIoU)of 78.71%and a mean pixel accuracy(MPA)of 88.22%on the self-constructed dental arch dataset,representing improvements of 3.61%and 3.15%over existing methods,with a model parameter size of only 5.8M.The results on the public Teeth3DS dataset show that the mIoU and MPA reach 73.61%and 82.05%,respectively,representing improvements of 10.42%and 11.15%over existing methods,demonstrating superior segmentation performance.The results demonstrate that the integration of least squares conformal mapping and the optimized DeepLabv3+network enables high-precision and lightweight 3D dental segmentation.
作者
马天
冷應庆
马子楠
李远成
MA Tian;LENG Yingqing;MA Zinan;LI Yuancheng(College of Artificial Intelligence and Computer Science,Xi’an University of Science and Technology,Xi’an 710054,China)
出处
《广西大学学报(自然科学版)》
2025年第6期1182-1194,共13页
Journal of Guangxi University(Natural Science Edition)
基金
科技创新2030-重大项目(2022ZD0119005)
陕西省自然科学基础研究计划项目(2022JM-508)
陕西省自然科学基础研究计划项目(2025JC-YBMS-754)。
