三维人脸测量与分割系统研究  

Research on three-dimensional face measurement and segmentation system

作　　者：余宗璞 杨懿昕 王俨铮 刘雪岩 周平[1] 周光泉[1] YU Zongpu;YANG Yixin;WANG Yanzheng;LIU Xueyan;ZHOU Ping;ZHOU Guangquan(School of Biomedical Science and Medical Engineering,Southeast University,Nanjing 210096,China)

机构地区：[1]东南大学生物科学与医学工程学院,江苏南京210096

出　　处：《应用光学》2021年第4期664-670,共7页Journal of Applied Optics

基　　金：国家自然科学基金(61771130);中央高校基本科研业务费专项基金资助(2242020K40074)。

摘　　要：三维人脸的测量与分割有着广泛的应用需求,是目前重要的研究方向,但三维人脸数据的庞大、无序等问题制约了其快速发展。首先开发了基于结构光方法的三维人脸测量系统,获得以点云形式存储的高精度三维人脸数据。其次,采用保角变换对三维人脸数据进行了预处理,并采用二维卷积神经网络分割结合三维反映射的思路,实现了三维人脸的分割,解决了三维数据的无序性以及旋转性,降低了三维数据分割的时间消耗。实验结果表明,本文提出的三维人脸测量系统精度可达0.5 mm,三维分割的平均交并比可达0.78,二维分割结合三维反映射的整体效率明显高于直接进行三维分割的效率。The three-dimensional face measurement and segmentation have extensive application requirements and are the important research direction at present.However,the rapid development is constrained as the threedimensional face data is huge and unordered.Firstly,the three-dimensional face measurement system based on the structured light method was developed,and the three-dimensional face data with high precision saved in the form of point cloud was obtained.Secondly,after the conformal transformation was adopted to preprocess the three-dimensional face data and the two-dimensional convolutional neural network segmentation combined with the three-dimensional inverse mapping was adopted,the three-dimensional face segmentation was realized,the disorder and rotation of three-dimensional data were solved,and the time consumption of threedimensional data segmentation was reduced.The experimental results show that the proposed precision of three-dimensional face measurement system can reach to 0.5 mm,and the average intersection-over-union(IoU)of three-dimensional segmentation can reach to 0.78.The overall efficiency of two-dimensional segmentation combined with the three-dimensional inverse mapping is obviously higher than that of the threedimensional segmentation.

关 键 词：人脸三维测量 结构光 人脸分割 卷积神经网络 

分 类 号：TN209[电子电信—物理电子学]