单张图片自动重建带几何细节的人脸形状

发布时间：2018-02-21 23:33

本文关键词： 图像重建人脸模型多线性模型方法非刚性迭代最近点方法拉普拉斯变形明暗重建形状　出处：《计算机辅助设计与图形学学报》2017年07期 　论文类型：期刊论文

【摘要】：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.
[Abstract]:In order to automatically and quickly in a single image reconstruction of facial expression in high detail model, put forward a method from coarse to fine. First, gradually optimize the feature points from a single photograph detected by facial expression, multiple initial iterative methods for solving the corresponding optimization of 3D head pose and large scale; secondly to face feature points detected on the basis of facial expression is not accurate correction, and using feature point iterative closest point method for non rigid alignment on the model and the image feature points, using the Laplasse coordinate affect the rest of non feature points; finally using geometric detail method with constant brightness reconstruction shape assumption for facial model reconstruction of fine scale reconstruction, in order to increase the fidelity of the model. The experimental results show that this method can generate face model with geometric details more accurately in the head greatly under the condition of the swing; The method does not need to calibrate the camera of the picture, nor does it need to pre train or preset user's feature models and mixed shapes, while it does not restrict indoor lighting and monotonous background environment.

【作者单位】：中国科学院计算技术研究所前瞻研究实验室/移动计算与新型终端北京市重点实验室;中国科学院大学;
【基金】：中国科学院计算技术研究所创新课题(20166040)
【分类号】：TP391.41

【参考文献】