A Novel Enhanced Nonnegative Feature Extraction Approach

Nonnegative matrix factorization (NMF) is a promising approach for image-data representation and classification under the nonnegativity constraint. However, the traditional NMF does not exploit the data-label information and its performance will be degraded in classification tasks. To overcome the flaw of NMF, this paper proposes a novel enhanced nonnegative matrix factorization (ENMF) method for learning powerful discriminative feature. The basic idea is that the training data from the same class are forced to embed into their own feature subspaces, where these subspaces are mutually orthogonal. The ENMF update rule is obtained by means of auxiliary function technique and Cardano's formula. We prove and analyze the convergence of the proposed ENMF algorithm theoretically and empirically. Experimental results on face recognition show that the proposed method outperforms the existing state-of-the-art NMF-based algorithms.