基于全局信息的人脸特征点精确定位

摘要

空间中物理位置的线性变化在图像空间的变化往往是非线性的，人脸特征点定位受到头部姿态较大的影响。提出一种改进的基于头部姿态估计的条件回归森林方法，该方法有效地将原非线性问题转换为分段线性问题。使用局部保持投影（LPP）得到全局的姿态信息标签，通过非线性回归（NLR）得出头部姿态，训练并使用条件回归森林对全局特征条件下的人脸特征点进行一个精确定位。实验结果表明，该方法有效地降低了头部偏转等图像空间中的非线性变化引起的特征估计误差，提高了人脸特征点定位的精确度。%Due to changes in the reality are linear physics position changes while in the image space are nonlinear, thus head pose has great influence for facial feature detection. In this paper, it proposes a precision head pose estimation method as a condition of regression forests, the algorithm modifies the performance of the nonlinear problem into a linear one. The main idea is used by Locality Preserving Projection and Nonlinear Regression(LPP+NLR)for obtaining the global information of pose and label, then, it utilizes trained conditional regression classifier to identify the feature points in global characteristics. Experimental results show that the algorithm decreases the missing rate caused by head deflection, and increases the precision of facial feature point detection.