A two stage algorithm for face recognition: 2DPCA and within-class scatter minimization

Abstract

The paper proposes a two-phase algorithm using 2DPCA and Gram-Schmidt Orthogonalization Procedure for better representation of face images with reduced dimension. While minimizing the within-class scatter, maximization of the total scatter is taken into account. The proposed method obtains the covariance matrix as in 2DPCA, and applies eigenvalue-eigenvector decomposition to this covariance matrix. Feature extraction is achieved using only d eigenvectors corresponding to largest d eigenvalues. The algorithm computes orthonormal bases by applying Gram-Schmidt Orthogonalization Procedure. Using these orthonormal bases, a common feature vector is calculated for each space in a class. A common feature matrix, which is used for image recognition, is then obtained for each class by gathering d common feature vectors of this class in a matrix form. Ar-Face database is used for experimental study. The proposed method produced better recognition rates compared to Eigenface, Fisherface and 2DPCA.