Fisher Discriminative Coupled Dictionaries Learning

摘要

As a recently proposed technique, dictionary learning (DL) has been extensively studied in the field of pattern recognition. Most scholars use a sparse representation as the basic formula for DL while incorporating other techniques into the DL process for obtain an expected dictionary, and exploring a problem with an \( l_{0} \)-norm or \( l_{1} \)-norm. However, these strategies increase the time complexity and require additional classifier-aided classification work. In this paper, we propose a novel form of DL called Fisher discriminative coupled dictionaries learning based on general dictionary learning. We use an \( l_{2} \)-norm to improve the training speed. On embedding the Fisher discrimination into the process of DL, the updated dictionary contains the discriminant information. We update the sample dictionary and coefficient projection dictionary simultaneously as a “dictionary pair”. The sample dictionary is used directly for image classification. The superiority of the proposed method is proven through exhaustive experiments on the AR, extended Yale-B, Scene 15, and Caltech-101 databases.