为了从单张图片中自动快速重建高细节人脸表情模型,提出一种从粗到细逐步优化的方法.首先从单张照片检测到的特征点中,通过多初值迭代方法优化求解对应的三维头部姿态和大尺度的人脸表情;其次以检测到的人脸特征点为依据对不准确的人脸表情进行矫正,并使用非刚性的迭代最近点方法对齐模型上的特征点和图像特征点,使用拉普拉斯坐标影响其余非特征点位置;最后使用带有常量假设的明暗重建形状方法为人脸模型重建细尺度的几何细节,以增加重建模型的逼真度.实验结果表明,文中方法能够在头部大幅度摆动的情况下生成更准确的带有几何细节的人脸模型;该方法不需要对图片的摄像机进行标定,也不需要预先训练或预设用户的特征模型和混合形状,同时不强制约束室内光照以及单调的背景环境.%This paper proposes a coarse-to-fine method to reconstruct a detailed facial model from a single image.Based on detected feature points from the image,our method starts by solving a rigid head pose and a coarse expression using multiple initials.Then we correct the 3D shape using 2D feature points.This step is combined with a non-rigid ICP method and a Laplacian deformation method.Finally,we generate detailed geometry from the coarse shape using a shape-from-shading method.In this step,our method uses a constant hypothesis to solve the boundary problem.Experimental results demonstrate that the method can reconstruct an accurate facial shape with details even when the head swings significantly.Our method does not need camera calibration nor any pre-processing like per-user training or manual preparation.In addition,the method can achieve good results under poor lighting conditions or in a variegated background.
展开▼
