Supervised dictionary learning with multiple classifier integration

• Multiple classifier learning is integrated into sparse dictionary learning.
• The proposed algorithm simultaneously updates dictionary and classifiers.
• The proposed method can largely improve the discriminability of sparse codes.
• An interesting insight into label consistency from the view of ensemble learning.
• The experiments show the excellent performance of the proposed method.

Supervised sparse coding has become a widely-used module in existing recognition systems, which unifies classifier training and dictionary learning to enforce discrimination in sparse codes. Many existing methods suffer from the insufficient discrimination when dealing with high-complexity data due to the use of simple supervised techniques. In this paper, we integrate multiple classifier training into dictionary learning to overcome such a weakness. A minimization model is developed, in which an ensemble of classifiers for prediction and a dictionary for representation are jointly learned. The ensemble of classifiers is constructed from a set of linear classifiers, each of which is associated with a group of atoms and applied to the corresponding sparse codes. Such a construction scheme allows the dictionary and all the classifiers to be simultaneously updated during training. In addition, we provide an interesting insight into label consistency from the view of multiple classifier learning by showing its relation with the proposed method. Compared with the existing supervised sparse coding approaches, our method is able to learn a compact dictionary with better discrimination and a set of classifiers with improved robustness. The experiments in several image recognition tasks show the improvement of the proposed method over several state-of-the-art approaches.