Abstract

Dictionary learning (DL) for sparse coding has shown  promising results in classification tasks, while how to  adaptively build the relationship between dictionary atoms  and class labels is still an important open question. The  existing dictionary learning approaches simply fix a dictionary atom to be either class-specific or shared by all  classes beforehand, ignoring that the relationship needs to  be updated during DL. To address this issue, in this paper  we propose a novel latent dictionary learning (LDL)  method to learn a discriminative dictionary and build its  relationship to class labels adaptively. Each dictionary  atom is jointly learned with a latent vector, which  associates this atom to the representation of different  classes. More specifically, we introduce a latent  representation model, in which discrimination of the  learned dictionary is exploited via minimizing the  within-class scatter of coding coefficients and the  latent-value weighted dictionary coherence. The optimal  solution is efficiently obtained by the proposed solving  algorithm. Correspondingly, a latent sparse representation based classifier is also presented. Experimental results  demonstrate that our algorithm outperforms many recently proposed sparse representation and dictionary learning  approaches for action, gender and face recognition.