基于特征融合和流形增强的视频人脸识别  

Video Face Recognition Based on Feature Fusion and Manifold Enhancement

作　　者：黄存东[1] 刘仁金[2] 杨思春[3] 

机构地区：[1]安徽国防科技职业学院信息工程系,安徽六安237011 [2]皖西学院信息工程学院,安徽六安237012 [3]安徽工业大学计算机学院,安徽马鞍山243032

出　　处：《计算机工程》2012年第9期193-196,共4页Computer Engineering

基　　金：安徽省高校教学研究基金资助重点项目(20101689);安徽省自然科学基金资助项目(11040606M150);安徽省高校自然科学研究基金资助重点项目(KJ2009A054;KJ2011A048)

摘　　要：提出一种在视频环境下的人脸识别算法(FFME)。将人脸划分为不同的区域,并融合针对人脸不同区域变化特点的不同类型局部特征,建立K-NN模型,根据sum rule划分人脸分类,利用流形建立参考人脸图集,以此重排分类结果,增强人脸识别准确率。在视频人脸数据库Mobo数据集和Honda/UCSD数据集上的实验结果表明,FFME的识别性能优于主成分分析、线性鉴别分析、隐马尔科夫模型、局部线性嵌入,以及流形距离等方法。A new approach of face recognition in videos, named FFME, is proposed. Faces are divided into different regions. Different local feature descriptors, which are selected against the variation of some region, are combined for face representation. K-NN model is built for feature vectors classification. A sum rule is followed by FFME to combine the individual classification results. Manifold is introduced for constructing a set of reference face images to re-rank the top candidate images, which is approved later to enhance the face recognition rates. Experimental results on the video-based face database Mobo and Itonda/UCSD show that FFME is superior to other methods such as principal components analysis, linear discriminant analysis, hidden Markov model, locally linear embedding and manifold-manifold distance.

关 键 词：特征融合 流形 视频人脸识别 局部特征 全局特征 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]