Lasso整脸形状回归的人脸配准算法

摘要

人脸配准是人脸识别、美化和面部表情分析等人脸相关应用的重要组成部分，这些应用通过人脸配准以精准定位人脸五官及面部轮廓特征点。在整脸形状回归的人脸配准框架基础上，使用 Lasso 回归来解决人脸配准问题，提出基于 Lasso 的整脸回归人脸配准算法。首先对人脸配准过程中的回归系数施加 L1模惩罚，以在不牺牲效果的基础上减少模型大小；然后提出人脸变换比例调整方法，在回归过程中使用人脸变换比例对特征点位置进行调整，用于解决小规模样本条件下不同尺度样本相互干扰的问题。在相关数据集上的实验结果表明，该算法配准精确度高，可以达到实时的速度，且适用于不同姿态下的人脸配准问题。%Face alignment focuses on the problem for localizing facial landmarks and it is an important topic of face recognition, beautification and facial expression analysis etc. Based on Explicit Shape Regression (ESR) framework, we proposed a face alignment algorithm using Lasso regression by adding aL1 norm con-straint on the model parameters to reduce the size of shape model while maintaining the resultant perform-ance. Meanwhile, a face ratio transformation method is proposed by using the face ratio to adjust the estima-tion result of face landmarks. The method can be used to solve the problem of mutual influence among various training samples in different scales. Experimental results show that the overall system performs well in related dataset and it can run in real-time and be insensitive to human faces in multiple poses.