高精度三维人脸图像数据库

High Precision 3D Facial Image Database

3D人脸图像数据库广泛应用于计算机视觉、动画绘图设计、医学等很多领域。在法庭科学领域,采集三维人脸图像并建立数据库,可进行人像特征分类、统计人像特征的分布以及训练人像模型,这些分析是人像比对和识别的基础。与传统的二维数据库相比,三维(3D)人脸图像库能够提供更多信息,例如,三维人脸图像的空间结构和形状包含多视角轮廓。3D人脸图像采集方法包含多视角几何信息的方法、结构光的方法和3D扫描仪的方法,这些方法有不同的采集设备和环境。国内外已经建立了几个有代表性的3D人脸图像数据库,例如MPI实验室的MPI和BJUT的BJUT-3D,但这些库在分辨率和精度方面尚有不足。本文首先回顾了MPI、BJUT-3D数据库和它们的采集环境,然后对建立中国人的高精度3D人脸图像数据库进行探索研究。用彩色手持三维扫描仪(Artec Spider)采集了1100个3D人脸图像,这些图像包含彩色纹理和深度信息(几何形状和点云),每个人脸图像的几何形状的采样点数目超过2000万,三角面片数目超过4000万。与BJUT-3D人脸数据库在人脸形状、分辨率和纹理等方面的比较结果显示,本研究采集的人脸图像有更高的精度,在嘴巴、鼻子、眼睛等方面比其他数据库中的人脸图像显示了更多的细节。建立的数据库将会支持在3D人像识别和算法评估方面的进一步工作。

Three-dimensional（3D） facial database can be wildly used in computer vision, animated drawing designing, medicine and etc. In forensic science, facial image database is used to classify face features, calculate the distribution of face features, and train the model. These analyses are the base for face identification and recognition. 3D facial image database can provide more information than traditional 2D database. For example, the special structure and shape in 3D facial image include multi-view profile. The 3D facial image capturing methods include multi-view geometry information, structured light and 3D scanning. These methods have different capturing devices and environments. There are several typical 3D facial image databases such as the MPI from MPI lab and the BJUT-3D from BJUT, which have some disadvantages in resolution and precision. In this paper, we first reviewed MPI and BJUT-3D databases and their capturing environments. Then we established our high precision 3D facial image database of Chinese people. 1100 3D facial images had been captured, including their color texture and depth（geometry shape or point cloud） information with Artec Spider, a portable 3D scanner. The number of sampling points of the geometry shape was more than 20 million. The number of triangle surfaces of 3D image was more than 40 million. The results showed that the images in our database exhibited higher precision than those in BJUT-3D database when considering face shape, resolution and texture. Meanwhile, these images presented more details in mouth, nose, eyes etc. than the images in the other databases. The established database would support our further work in 3D facial identification and algorithm evaluation.