Coherence regularized dictionary learning

Abstract

Sparse representations over redundant learned dictionaries have shown to produce high quality results in various image processing tasks. An important characteristic of a learned dictionary is the mutual coherence of dictionary that affects its generalization performance and the optimality of sparse codes generated from it. In this paper, we present a dictionary learning model equipped with coherence regularization. For this model, two novel dictionary optimization algorithms based on group-wise minimization of inter- and intra-coherence penalties are proposed. Experimental results demonstrate that the proposed algorithms improve the generalization properties and sparse approximation performance of the trained dictionary compared to several incoherent dictionary learning methods.