Robust Generalized Low Rank Approximation of Matrices for image recognition

For a set of 2D objects such as image representations, a 2DPCA approach that computes principal components of row-row and column-column covariance matrices would be more appropriate. The Generalized Low Rank Approximation of Matrices (GLRAM) approach has proved its efficiency on computation time and compression ratio over 1D principal components analysis approaches. However, GLRAM fails to efficiently account noise and outliers. To address this problem, a robust version of GLRAM, called RGLRAM is proposed. To weaken the noise effect, we propose a non-greedy iterative approach for GLRAM that maximizes data covariance in the projection subspace and minimizes the construction error. The proposed method is applied to face image recognition and shows its efficiency in handling noisy data more than GLRAM does. Experiments are performed on three benchmark face databases and results reveal that the proposed method achieves substantial results in terms of recognition accuracy, numerical stability, convergence and speed.