受限局部模型在人脸特征点定位中应用综述  

Facial Feature Point Detection with Constrained Local Models: a Survey

作　　者：金鑫[1] 谭晓阳[1] 

机构地区：[1]南京航空航天大学计算机科学与技术学院,南京210016

出　　处：《小型微型计算机系统》2017年第2期371-375,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61373060)资助;国家"九七三"重点基础研究发展计划项目(2015CB352401)资助;江苏省科创基金项目(KYLX_0289)项目

摘　　要：人脸特征点定位(又称人脸对齐)的目标是定位出人脸图片中可以描述人脸部件的特征点,如眼角、鼻尖、嘴角等.受限局部模型(Constrained Local Models)是一类重要的人脸特征点定位算法,其拟合过程通常可以形式化为两步:首先使用局部检测器扫描图片,为每一个人脸特征点生成一张响应图;然后基于响应图优化一个全局人脸形状模型,从而得到人脸形状(即所有人脸特征点)的一次更新.文献中的受限局部模型默认使用主成分分析(PCA)法建立参数人脸形状模型;而非参数形状模型如基于样本(exemplar-based)的方法虽然也被成功的应用于人脸特征点定位,但却没有被纳入受限局部模型的范畴.在这样的背景下,本文的贡献主要有3点:1)成功将基于参数(PCA-based)和非参数(exemplar-based)人脸形状模型的方法统一到概率形式的受限局部模型的框架之下;2)在3个公开的数据集LFW、LFPW、HELEN上汇报了几种典型的参数和非参数受限局部模型的性能;3)总结并讨论了参数和非参数受限局部模型的优缺点,如基于PCA的参数模型拟合速度更快,而基于exemplar的非参数模型精度更高.Facial feature point detection, which is usually called face alignment, aims to localize the facial landmarks such as the eye comers, the tip of the nose and the mouth comers that can describe the shapes and locations of facial components. Constrained Local Models （ CLMs ） are one of the most prominent approaches for face alignment with impressive results. Fitting in most CLM methods is typically formulated as a two-step process during which discriminatively trained local detectors are first used to scan the image to yield a filter response for each facial point and then global shape optimization is performed over these filter responses. While CLMs typical- ly model the face shape parametrically by the Principle Component Analysis （PCA} ,some non-parametric（ exemplar-based} methods also achieve good performance in face alignment, which however have not been categorized as CLMs in literature. The contribution of this paper are three-fold： 1 ） we propose a generic probabilistic framework for CLMs that include both parametric （PCA-based） and non-parametric（exemplar-based ） shape model based methods ;2 ） we report the performances of several representative parametric and non-parametric CLMs on three public face datasets ;3） we present discussions about the advantages and limitations of both parametric and non-parametric CLMs, demonstrating some interesting conclusions such as that PCA-basd parametric CLMs are more efficient, while the exemplar-based non-parametric CLMs are more accurate and robust in fitting.

关 键 词：受限局部模型 人脸特征点定位 人脸对齐 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]