Adaptive weighted orthogonal matching pursuit for robust pattern classification

Abstract

Orthogonal Matching Pursuit has been a popular sparse representation algorithm due to its simplicity and competitive performance. However, the original orthogonal matching pursuit algorithm and most of its variant utilize the mean square error (MSE) criterion and assign the same weight to each entry of the data. In the presence of large impulsive noise, their performance may severely degrade due to the effect of corrupted entries with large noises. In this work, we propose a novel adaptive weighted orthogonal matching pursuit method for sparse representation with application to robust classification. We show that the proposed method can adaptively assign large weights on clean entries of data and small weights on severely corrupted ones. The experiments on benchmark real-world databases demonstrate the efficacy of the proposed method for handwritten digit and face recognition.