Image set classification using discriminative dictionary learning

Abstract

In this thesis, we propose a strategy for classification of image sets using discriminative dictionary learning. The proposed strategy was motivated from two ideas: (i) a redundant dictionary can be regarded as a collection of subspaces and sparse coding over a discriminative dictionary as the selection of a subspace that corresponds to the class of an input signal; (ii) we can expect that more robust result can be achieved by using sets of vectors rather than vectors themselves because they can cover various characteristics of images or image objects. We propose a way to combine two ideas by defining an objective function as well as an algorithm to find the optimal dictionary. We verify the proposed formulation and the algorithm on two publicly available datasets and compare the result with a comparative method. The results on the datasets show that superiority of our algorithm in terms of recognition rate increases as the number of classes in dataset gets larger.