Robust Projective Dictionary Learning by Joint Label Embedding and Classification

Abstract

In this paper, we propose a new discriminative dictionary learning framework, called robust Label Embedding Projective Dictionary Learning (LE-PDL), for data classification. LE-PDL can learn a discriminative dictionary and the blockdiagonal representations without using the l0-norm or l1-norm sparsity regularization, since the l0 or l1-norm constraint on the coding coefficients used in the existing DL methods makes the training phase time-consuming. To enhance the performance, we also consider label information of the dictionary atoms in the learning process of LE-PDL to encourage the intra-class atoms to deliver similar profiles and enforce the coefficient matrix to be block-diagonal. Besides, our LE-PDL also involves an underlying projection to bridge data with their coefficients by extracting special features from given data. Then, we can train a classifier based on the extracted features so that the classification and representation powers are jointly considered. So, the classification approach of our model is efficient, since it avoids the extra time-consuming sparse reconstruction process with trained dictionary for each new test data as most existing DL methods. Besides, a robust l2,1-norm is regularized on the classifier and the non-negative constraint is used for the coding coefficients to enhance the performance. Experimental results show the effectiveness of our formulation.