Skip to Main Content
Non-negative matrix factorization (NMF) as a part-based representation method allows only additive combinations of non-negative basis components to represent the original data, so it provides a realistic approximation to the original data. However, NMF does not work well when directly applied to face recognition due to its global linear decomposition; this intuitively results in a degradation of recognition performance and non-robustness to the variation in illumination, expression and occlusion. In this paper, we propose a robust method, random subspace sub-pattern NMF (RS-SpNMF), especially for face recognition. Unlike the traditional random subspace method (RSM), which completely randomly selects the features from the whole original pattern feature set, the proposed method randomly samples features from each local region (or a sub-image) partitioned from the original face image and performs NMF decomposition on each sampled feature set. More specially, we first divide a face image into several sub-images in a deterministic way, then construct a component classifier on sampled feature subset from each sub-image set, and finally combine all of component classifiers for the final decision. Experiments on three benchmarks face databases (ORL, Yale and AR) show that the proposed method is effective, especially to the occlusive face image.