基于散斑立体匹配的快速三维人脸重建

Fast 3D face reconstruction by speckle stereo matching

本文提出一种高效的人脸三维重建方法。该方法将散斑投影至人脸表面以增加其特征信息,并采用一种由粗到精的时空立体匹配算法来提高三维人脸重建的精度。进一步,该算法利用时空积分图对立体匹配的代价函数进行加速计算,进而提高了重建效率。此外,所提出方法通过人脸检测去除无关背景,使后续的三维重建算法能够高效地作用在人脸区域上。实验表明,所提出方法对3D打印人脸(精度为0.01 mm)模型的重建平均误差为0.32 mm,对哑铃规球心测距和直径测距(精度皆为0.01 mm)其误差皆低于1个百分点,以上结果优于同类产品。与现有立体匹配算法相比,本文方法所得视差图面部无空洞且视差变化均匀,更真实地反映出被测人脸的三维形状。

This paper proposes an efficient method for 3 D face reconstruction.Speckle patterns are projected onto a face to enhance its discriminative feature information,and a coarse-to-fine stereo matching scheme is investigated to improve the accuracy of reconstruction.Furthermore,spatio-temporal integral image is employed to accelerate the computation of cost function during stereo matching,which subsequently improves the efficiency of reconstruction.In addition,the proposed method removes unrelated background regions by using face detection,so that the following 3 D reconstruction process is able to focus only face area for the purpose of efficiency.Experimental results show that the proposed method achieves the average error of 0.32 mm on reconstructing a face model that is obtained by 3 D printer(printing precision 0.01 mm).When measuring two standard spheres(precision 0.01 mm),measurement errors on the distance of sphere centers and the sphere diameters are lower than 1%.Such results are better than those of some existing 3 D camera products.Compared with existing stereo matching approaches,disparity maps generated by our method show almost no cavity and vary more smoothly,which show more consistency with the real shape of human face.