Tharshini Gunendradasan, Chinthaka Dinesh, Roshan I. Godaliyadda, and Mervyn P.B. Ekanayake
Face recognition, expression-variant faces, expression negation, prin-cipal component analysis (PCA), cosine similarity, single sample perperson
In many face recognition applications, facial expression is a signiﬁcant complication, especially when the system has limited samples per person. In this paper, a method that handles single sample per subject problem is proposed. This method neutralizes the expressive
images and allows face recognition to be performed on the neutral component of the faces. In this approach, a given expressive face is considered as an aggregation of neutral and expression components. The neutral element is determined by subtracting the expression component, assuming the deformation incurred on faces under given
expression is alike between individuals. To rationalize this assumption, images are warped to their corresponding standard expression templates to coordinate the features of similar expressive faces. Generic training images are used to analyse the prior expression information, and the expression component for each expression is learned using a principal component analysis-based approach. In
order to make the neutral domain classiﬁcation approach precise and robust, ﬁrst a few similar images of given probe images are sorted out from the gallery by means of estimated neutral components. Then, speciﬁc constituent components of the face are used to determine the correct face. Experiments on the Cohn–Kanade database demonstrate the eﬀectiveness of the proposed approach.