基于深度学习的颅面点云生成模型

Craniofacial Point Cloud Generation Model Based on Deep Learning

针对实际应用中对初始头骨数据完整性要求高、重建复原时间长、扩展性和实用性不强等问题,论文设计提出一种基于深度学习的形状修补以及形状学习颅面点云复原模型。模型通过使用补全自编码器(Repair AutoEncoder,rAE)从缺失的头骨点云数据中复原完整头骨点云,并通过形状生成自编码器(Generation AutoEncoder,gAE)来生成初始头骨所对应的头面部点云。实验表明该模型补全生成的头骨点云和学习生成的头部点云的复原精度能很好地支撑对头部模型的三维重建。论文提出的深度学习方法与传统方法相比较能够由不完整的头骨数据,快速、逼真和准确地重构三维头部形状,在经典颅面复原任务如:出土残缺古人类头骨补全、面貌复原、刑事案件侦破等任务中展现出良好的稳定性与可扩展性。

Aiming at the problems that the existing craniofacial restoration methods have high requirements on the integrity of the initial skull data,long reconstruction and restoration time,and poor scalability and practicability in practical applications,this paper designs and proposes a shape inpainting and shape-learning craniofacial point cloud restoration model based on deep learning.The model uses the rAE to restore the complete skull point cloud from the missing skull point cloud data,and uses the gAE to generate the head and face point cloud corresponding to the original skull.Experiments show that the restoration accuracy of the model complementing the generated skull point cloud and the learned head point cloud can well support the 3D reconstruction of the head model.Compared with traditional methods,the deep learning method proposed in this paper can quickly,realistically and accurately reconstruct the three-dimensional head shape from incomplete skull data.It shows good stability and scalability in classical craniofacial restoration tasks,such as the completion of unearthed incomplete ancient human skulls,face restoration,and criminal case detection.