Sparse Discriminative Tensor Dictionary Learning for Object Classification

Dictionary learning methods aim to learn atoms that can best represent a signal class. In recent years, these methods have been extended to learn discriminative dictionaries such that the learned atoms are specific to each class enabling better classification accuracy. With the increase of high dimensional and multi-aspect data, there's a growing need to extend dictionary learning algorithms for tensor type data. In this paper, we propose an efficient, separable and orthogonal dictionary structure for learning class-specific dictionaries for tensor objects. The proposed cost function tries to minimize the representation error as well as within-class scatter while putting a sparsity constraint on the learned representation. The algorithm is applied to different tensor object classification tasks with extensive evaluations of the effect of sparsity, discriminability and reconstruction error on classification accuracy.