Application and method of feature extraction for approximate orthogonal non-negative matrix factorization
In order to get effectively feature vectors of samples, a method of feature extraction is proposed based on non-negative matrix factorization (NMF) with approximate orthogonal constraint. An objective function is defined to impose approximate orthogonal constraint, in addition to the non-negative constraint in the standard NMF. An algorithm of iterative update for basis matrix and weight matrix in non-negative matrix factorization is derived, and then computational method of feature vectors is gotten. A proof of the convergence of the algorithm is provided. Empirical results of handwritten character recognition show that there is high recognition accuracy if only the rank of the basis matrix is set correctly. The method of feature extraction is useful to get essential feature of sample, and is effective to recognition when the structures of the handwritten characters are obviously different.