利用形状估计的人脸特征点定位算法

摘要

针对鲁棒级联姿势回归算法(robust cascaded pose regression,RCPR)缺乏形状约束,对复杂人脸图像的定位精度差、成功率低的问题,提出一种利用形状估计的分块特征点定位算法.为提高定位成功率和准确度,对人脸特征点进行分块,对每一块进行形状估计作为约束;为保证形状估计的精度和连续性,在传统核回归的基础上,学习得到图像特征与目标形状间的联合概率分布函数,称做匹配函数,并求取最大值作为形状估计;为提高算法性能,只需对部分点的位置进行回归,减少了回归器的数量,并引入了形状索引特征的采样先验.实验表明,算法对复杂人脸图像具有更高的定位准确度和鲁棒性,定位成功率可达86%,同时计算速度可以实现实时处理.%To improve the localization accuracy and success rate of complex facial image which lack of shape constraint in RCPR algorithm,this paper proposed a novel part-based cascaded regression algorithm combining shape estimation.Firstly,it divided the facial landmarks into several areas and used the shape estimation as shape constraint.Secondly,in order to ensure the accuracy and continuity of the shape estimation,it learned a joint probability distribution function of image features and target shapes,known as matching function.The estimated shape of the target could be obtained by maximizing the matching function.Thirdly,to improve algorithm performance,it exerted a prior constraint of the distance between the pixels pair during sampling shape-indexed features,meanwhile it reduced the number of regressors.Experiments demonstrate that the algorithm outperforms in localization accuracy and achieves a better robustness for occlusion and complex facial image,a localization success rate of 86% is obtained and can realize real-time processing.