基于嘴巴状态约束的人脸特征点定位算法  

A facial feature point locating algorithm based on mouth-state constraints

作　　者：师亚亭 李卫军[1] 宁欣[1] 董肖莉 张丽萍[1] SHI Yating LI Weijun NING Xin DONG Xiaoli ZHANG Liping(Institute of Semiconductors, Chinese Academy of Sciences, Beijing 100083, China)

机构地区：[1]中国科学院半导体研究所高速电路与神经网络实验室,北京100083

出　　处：《智能系统学报》2016年第5期578-585,共8页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61572458)

摘　　要：嘴巴区域特征点的精确定位对于特征匹配、表情分析、唇形识别、驾驶行为分析等应用具有极其关键的作用。然而,用现有的人脸特征点定位算法进行人脸形状估计时,嘴巴区域特征点的定位误差相对较大。针对这一问题,提出了基于HSV颜色空间和基于卷积神经网络的两种嘴巴状态分类器以及一种基于局部特征点位置关系的强形状约束策略,并在此基础上提出了基于嘴巴状态约束的人脸特征点定位算法,根据嘴巴状态标签对显式形状回归ESR算法的估计结果进行约束以获得更加准确的特征的位置。相比传统的ESR算法,该方法在保障人脸形状定位鲁棒性的同时,在Helen数据库和LFPW数据库上的嘴巴特征点定位准确度均明显提高。The precise locations of the feature points of the mouth critically influence applications which use feature matching, expression analysis, lip recognition and driving behavior analysis, etc. However, when estimating facial shapes using current facial landmarks detecting methods, the locating error of feature points around the mouth re- gion is relatively large. In order to solve this problem, two kinds of ＇ mouth-state＇ classifiers were proposed, one was based on HSV color space and the other on a convolutional neural network, with a strong shape constraint strat- egy focusing on the spatial relationship between local facial landmarks. Furthermore a facial feature point locating method was presented based on the mouth-state constraint, which constrains the predicted explicit shape regression （ESR） result and is more accurate as regards locating facial landmarks. Compared with the original ESR algorithm, this method significantly improves the accuracy of locating landmarks for the mouth for both the Helen and LFPW datasets, and has no impact on the robustness of facial shape prediction.

关 键 词：人脸特征点定位 ESR 嘴巴状态分类器 强形状约束 HSV颜色空间 卷积神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]