Discriminative dictionary learning via mutual exclusion

We apply our recently developed concept of mutual exclusivity [1] in the context of discriminative coding, to the problem of learning dictionary for representing signals drawn from N classes in a way that optimizes their discriminability. We first briefly review our mutual-exclusivity concept and then deploy it a simple discriminative dictionary learning algorithm that directly generalizes the well-known KSVD algorithm which is addressed for the traditional problem of signal coding. We demonstrate performance improvements over traditional KSVD based feature extraction schemes and conclude by describing avenues for future research.