Abstract: Aiming at the efficiency and efficacy of online feature learning for face recognition, a novel
incremental discriminant learning algorithm of non-negative matrix factorization is proposed. Compared
with former unsupervised online matrix factorization methods, our approach takes good advantage of the
label information of training samples. Coefficient vector means of the same class were used for
initialization, then the algorithm iterates with constraints on minimizing Euclidean distance of in-class
samples. Which results more discriminative features and less iterations in computation. Experiments on
ORL and PIE face datasets demonstrated that the proposed algorithm achieved better classification
accuracy and converged faster than former batch based non-negative matrix factorization algorithms.

Key words: face recognition, supervised learning, non-negative matrix factorization, incremental learning