Abstract:

A generalized discriminant orthogonal non-negative matrix factorization algorithm is proposed. Unlike traditional non-negative matrix factorization (NMF) algorithms, this algorithm adds to orthogonal constraint to guarantee the nonnegativity of the low-dimensional features and it is also different from traditional discriminant non-negative matrix factorization algorithms which add discriminant constraints in low-dimensional weights. Because low-dimensional features involve in pattern recognition directly, the algorithm generalizes the discriminant constraints to low-dimensional features and improves the recognition accuracy. The algorithm is derived in detail and it is applied to face verification and facial expression recognition. The experiments indicate that the algorithm enhances the discrimination ability of low-dimensional features and has better performance.