Class-Aware Analysis Dictionary Learning for Pattern Classification

Abstract

Dictionary learning (DL) plays an important role in pattern classification. However, learning a discriminative dictionary has not been well addressed in analysis dictionary learning (ADL). This letter proposes a Class-aware Analysis Dictionary Learning (CADL) model to improve the classification performance of conventional ADL. The objective function of CADL mainly includes two parts to promote the discriminability. The first part aims to learn a discriminative analysis subdictionary for each class instead of a global dictionary for all classes. The learned analysis dictionary is class-aware, generating a block-diagonal coding coefficient matrix. The second part aims to enhance the discrimination of coding coefficients by integrating a max-margin regularization term into our proposed framework. This term ensures the coefficients of different classes to be separated by a max-margin, which boosts the confidence of classification. A theoretical analysis is also given to support the max-margin regularization term from the perspective of preserving the pairwise relations of samples in coding space. We employ an alternating minimization algorithm to iteratively find the convergent solution. By evaluating our method on four pattern classification datasets, we demonstrate the superiority of our CADL method to the state-of-the-art DL methods.