Face Recognition Using Class Specific Space Model

In this paper, we investigate the face recognition problem via clustering of frontal face images represented in frequency domain by low frequency discrete cosine transform (DCT) coefficients. Our approach termed as class specific space model (CSSM) is based on the assumption that faces of different subjects are clustered in different low dimensional subspace of the feature space. The proposed approach uses 2D-DCT for feature extraction, each of the class clusters in the feature space are later modeled under Gaussian mixture model framework by a set of parameters which best fit the data. The proposed approach is tested on AR face database and its effectiveness in terms of identification rate is compared with the conventional IPCA and DLDA-SVM based classifiers.