Jointly Learning Non-negative Projection and Dictionary with Discriminative Graph Constraints for Classification

Dictionary learning (DL) for sparse coding has shown impressive performance in classification tasks. But how to select a feature that can best work with the learned dictionary remains an open question. Current prevailing DL methods usually adopt existing well-performing features, ignoring the inner relationship between dictionaries and features. To address the problem, we propose a joint non-negative projection and dictionary learning (JNPDL) method. Non-negative projection learning and dictionary learning are complementary to each other, since the former leads to the intrinsic discriminative parts-based features for objects while the latter searches a suitable representation in the projected feature space. Specifically, discrimination of projection and dictionary is achieved by imposing to both projection and coding coefficients a graph constraint that maximizes the intra-class compactness and inter-class separability. Experimental results on both image classification and image set classification show the excellent performance of JNPDL by being comparable or outperforming many state-of-the-art approaches.