ComparisonofLinearand Non-linearData Projection TechniquesinRecognizingUniversal
B. Bullot · 2005
Thispapercompares theperformances oflinear and non-linear dataprojection techniques inclassifying sixuniversal facial expressions fromonlyvisual data. Threedifferent data projection techniques, namely, Principal Component Analysis (PCA), Non-negative MatrixFactorization (NMF)andLocal Linear Embedding (LLE)wereexperimented toproject the facial motion ontolowerdimensional subspace toestimate the intrinsic dimension of thefacial expression data.The classification approach relies onatwo-step strategy onthetopof projected facial motion information obtained fromsequence of facial expression images. First, abankoflinear classifier was applied onprojected dataanddecision madebythelinear classifiers wascoalesced toproduce acharacteristic signature foreachuniversal facial expressions. Thesignatures thus computed fromthetraining datasetwereusedtotrain discrete HiddenMarkovModels(HMMs)tolearn theunderlying model foreachfacial expression. Theperformances ofeachdata representations inclassifying thefacial expressions were compared using five fold cross validation onadatabase of488 videosequences thatinclude 97subjects. Tofurther illustrate theefficacy oftheproposed approach andtobetter understand theeffects ofanumberoffactors that aredetrimental tothe facial expression recognition usingonlyvisual datawere conducted. Thefirst empirical analysis wasconducted on a database consisting of108facial expressions collected fromTV broadcastandlabeled byhumancoder forsubsequent analysis. Thesecond experiment wasconducted onfacial expressions obtained from21subjects byshowing thesubjects sixdifferent clips ofmovies chosen ina mannertoarousespontaneous emotional reactions thatwouldproducenatural facial expression.