Discriminant analysis via jointly L2,1-norm sparse tensor preserving embedding for image classification

There exits an increasing interest on sparse subspace learning (SSL) for dimensionality reduction and pattern recognition. In this paper, we propose a novel sparse subspace learning method named discriminant sparse tensor neighborhood preserving embedding (DSTNPE) which incorporates discriminant information into tensor sparse neighborhood preserving embedding to perform robust image classification. DSTNPE introduces the L2,1-norm to sparse neighborhoods and criterion, in which the within-neighborhood tensor scatter and between-neighborhood tensor scatter are defined for sparse regression. One virtue of DSTNPE is that it can avoid selecting the scale of local neighborhood of the manifold learning algorithms. Additionally, DSTNPE can iteratively obtain the transformation matrices by the sparse tensor neighborhoods preservation. Furthermore, by means of virtue of maximum margin criterion(MMC), the discriminant performance of DSTNPE is further enhanced. To evaluate the proposed method, extensive experiments conducted on five public databases demonstrate that our proposed algorithm outperforms many state-of-the-art algorithms.