This paper proposes a sparse discriminant representation (SDR) model. The model utilizes information entropy and term frequency-inverse document (tf-idf) frequency to analyze and reconstruct the discriminative information carried by atoms. SDR is a general model; its result is dependent on the input data and hence can be applied to various applications.The paper includes two parts. The first half revisits the history of sparse representation and its applications in computer vision, introducing its fundamental principal, algorithms and state-of-the-art discriminative sparse representation models. Moreover, the paper proposes a novel sparse discriminant representation (SDR) under the sparse representation framework.The second half applies the proposed SDR to three specific applications:moving object detection, face recognition and saliency detection. In moving object detection, SDR is firstly utilized to reconstruct coarse motion areas, which are subsequently refined via a modified mean filtering and robust sparse coding; In face recognition, SDR is used to reconstruct facial images, then subspace learning is employed for face recognition; in saliency detection, SDR is used to compute prior maps as background prior for a similarity difference based framework, improving the performance. Finally, experiments are conducted on international benchmark datasets; the experimental results demonstrate the effectiveness of SDR in the three applications. |