Sparse Non-negative Pattern Learning for image representation

In this paper, we propose sparse non-negative pattern learning (SNPL) based on self-taught learning framework. In the algorithm, visual patterns are first learned from unlabeled data by non-negative matrix approximation with sparseness constraints, and then features are extracted by the second part of the algorithm, a conjugate family based non-negative sparse feature extraction method. By combining sparse and non-negative constraints of patterns together, SNPL model gives a better representation for images than state-of-art methods. Beyond that, we give an analytical solution for feature extraction although it is approximate, and thereby we extract the features for self-taught learning framework in a faster and more stable way. We apply the new model to various areas, including pattern coding, feature extraction, and recognition. Experimental results show the advantages of SNPL model.