On-line learning parts-based representation via incremental semi-supervised multi-label image annotation

In this paper, an incremental algorithm which is derived from nonnegative matrix factorization (NMF) is proposed for semi-supervised multi-label image annotation, is named (ISSML). by using Incremental non-negative matrix factorization (INMF) instead of NMF, our algorithm can learn a linear part-based subspace in an online fashion. INMF preserves dimension reduction capability of NMF without increasing the computational load and also stays constant the space complexity without residing the entire new data in the memory and thus can be applied to large-scale or streaming datasets. experimental results on three benchmark datasets show that efficiency of our proposed algorithm improves accuracy of image annotation and also decreases time complexity.