Face Recognition with APCA in Variant Illuminations

PCA (Principal Component Analysis) and FLD (Fisher Linear Discriminant) methods for face recognition perform well in controlled laboratory conditions but encounter difficulties under variant illuminations or facial expressions. We propose a new method called Affine PCA to overcome these limitations with respect to lighting variations. This technique distinguishes contributions between features in the affine PCA derived eigenspace according to the correlation between eigenfeatures and illumination. This technique is applied to the Asian Face Image Database PF01, which consists of 535 images - 107 faces under 5 different illuminations. We perform three-fold cross-validation on this database to show that the proposed Affine PCA method achieves 96% recognition rate compared to standard PCA and FLD with only 51% and 65% respectively.