Class-oriented discriminative twin reconstructions dictionary pair learning for visual recognition

Abstract

Projective dictionary pair learning (DPL) and its variations have shown particular effectiveness for pattern classification due to their powerful discriminative capability and efficient computational performance. However, to boost up classification capacity, most of these methods mainly focus on adding more complicate discriminative terms, which not only sacrifices computational cost but also results in overfitting. To release the aforementioned issues, in this paper we propose a novel DPL-based classification framework which adopts an effective reconstruction strategy termed discriminative twin reconstructions (DTR) and a simple fisher-like discriminative term. Moreover, the proposed DTR strategy reconstructs data in a cross-data form with $l_{2,1}$-norm constraint. The strategy benefits to simultaneously reduce the feature distribution gap of the coefficients in the same class and deliver more accurate reconstruction. But $l_{2,1}$-norm sparse constraint is pivotal to achieve more efficient optimization yet discriminative capacity. The experimental results compared with 13 dictionary learning-based methods validate the impressive effectiveness of our proposed DTR-based method.