A discriminant graph nonnegative matrix factorization approach to computer vision

This paper proposes a novel dimensional reduction method, called discriminant graph nonnegative matrix factorization (DGNMF), for image representation. Inspired by manifold learning and linear discrimination analysis, DGNMF provides a compact representation which can respect the original data space. In addition, In addition, the within-class distance of each class in the representation is very small. Based on these characteristics, our proposed method can be viewed as a supervised learning method, which outperforms some existing dimensional reduction methods, including PCA, LPP, LDA, NMF and GNMF. Experiments on image recognition have shown that our approach can provide a better representation than some classic methods.