This paper addresses the problem of learning a discriminative dictionary from training signals. Given a structured dictionary, each atom of which has its corresponding label, one signal should be mainly constructed by its closely associated atoms. Besides the representations for the same class ought to be very close to form a cluster. Thus we present out-of-label suppression dictionary model with cluster regularization to amplify the discriminative power. Upon out-of-label suppression, we don't adopt $l_0$-norm or $l_1$-norm for regularization. Meanwhile, two simple classifier are developed to take full advantage of the learnt dictionary. The effectiveness of the proposed dictionary model has been evaluated on two popular visual benchmarks.