Bayesian regularized nonnegative matrix factorization based face features learning

This paper proposes a novel technique for learning face features based on Bayesian regularized non-negative matrix factorization with Itakura-Saito (IS) divergence (B-NMF). In this paper, we show, the proposed technique not only explicitly incorporates the notion of `Bayesian regularized prior´ which imposes onto the features learning but also holds the property of scale invariant that enables lower energy components in the learning process to be treated with equal importance as the high energy components. Real test has been conducted and the obtained results are very encouraging.