Sparse Low-Rank Preserving Projection for Dimensionality Reduction

The representation-based learning methods, such as sparse representation-based classification and low-rank representation, show effective and robust for image clustering and classification. However, these methods essentially belong to the transductive methods and they cannot deal with the new samples. Meanwhile, the original high-dimensional data contains a large amount of redundant information. If the original data are directly performed, it will not only degrade the performance of the algorithm but also lead to a sharp increase in the amount of computation. Therefore, a novel robust sparse low-rank preserving projection (SLRPP) is presented for dimensionality reduction, in which both the essential similarity structure of the observed data and the optimal feature representation are simultaneously obtained. By alternatively iterating the augmented Lagrangian multiplier method and the eigendecomposition, the framework of the SLRPP can be solved. The experimental results on six image databases proved that our SLRPP algorithm can achieve a competitive performance compared with the state-of-the-art subspace learning methods.