IK-SVD: Dictionary Learning for Spatial Big Data via Incremental Atom Update

Abstract

A large group of dictionary learning algorithms focus on adaptive sparse representation of data. Almost all of them fix the number of atoms in iterations and use unfeasible schemes to update atoms in the dictionary learning process. It's difficult, therefore, for them to train a dictionary from Big Data. A new dictionary learning algorithm is proposed here by extending the classical K-SVD method. In the proposed method, when each new batch of data samples is added to the training process, a number of new atoms are selectively introduced into the dictionary. Furthermore, only a small group of new atoms as subspace controls the current orthogonal matching pursuit, construction of error matrix, and SVD decomposition process in every training cycle. The information, from both old and new samples, is explored in the proposed incremental K-SVD (IK-SVD) algorithm, but only the current atoms are adaptively updated. This makes the dictionary better represent all the samples without the influence of redundant information from old samples.