Discriminative components play an important role in computer vision and pattern recognition. In this paper, we propose a novel approach for learning the discriminative components of sparse representation. To achieve this goal, the information theoretic perspective of term frequency - inverse document frequency (TF-IDF) measure is introduced to learn a compact dictionary, which is composed of discriminative components such as patch, color and human priors, for salient motion detection. Experimental results on benchmark dataset suggest that the proposed method works effectively on reconstructing discriminative components.