Orthogonal Nonnegative CP Factorization for Image Representation and Recognition

An orthogonal non-negative CANDECOMP/PARAFAC factorization algorithm (ONNCP) is proposed. With the orthogonal constrain, the low-dimensional presentations of samples are kept non-negative in ONNCP. The relationship between NNCP and NMF is analyzed theoretically. The solution process and the convergence of the algorithm are discussed. The experiments indicate that, compared with other non-negative factorization algorithms, ONNCP can reduce the redundancy of the base images and enhance the sparseness of the base images due to its orthogonality. It also ensures the low-dimensional feature is non-negative. The algorithm can achieve better recognition rate in facial expression recognition and will convergence a fixed point. Furthermore, the algorithm can be generalized to any order tensor.