Improved Image Discrimination Using Fast Non-linear Orthogonal Dictionary Learning

Most real-world signals or images have an intrinsic non-linear similarity measure and can be harder to discriminate. Kernel dictionary learning with applications to signal classification offers a solution to such a problem. However, decomposing a kernel matrix for large datasets is a computationally intensive task. Existing papers on dictionary learning using optimal kernel approximation method improve computation run-time but learn an over-complete dictionary. In this paper, we show that if we learn a discriminative orthogonal dictionary instead then learning and classification run-time can be significantly reduced. The proposed algorithm, Kernelized simultaneous approximation, and discrimination (K-SAD), learns a single highly discriminative and incoherent non-linear dictionary on small to medium-scale real-world datasets. Extensive experiments result in > 97% classification accuracy and show that the algorithm can scale both in space and time when compared to existing dictionary learning algorithms.