In this paper, a novel semi-supervised dictionary learning and sparserepresentation (SS-DLSR) is proposed. The proposed method benefits from thesupervisory information by learning the dictionary in a space where thedependency between the data and class labels is maximized. This maximization isperformed using Hilbert-Schmidt independence criterion (HSIC). On the otherhand, the global distribution of the underlying manifolds were learned from theunlabeled data by minimizing the distances between the unlabeled data and thecorresponding nearest labeled data in the space of the dictionary learned. Theproposed SS-DLSR algorithm has closed-form solutions for both the dictionaryand sparse coefficients, and therefore does not have to learn the twoiteratively and alternately as is common in the literature of the DLSR. Thismakes the solution for the proposed algorithm very fast. The experimentsconfirm the improvement in classification performance on benchmark datasets byincluding the information from both labeled and unlabeled data, particularlywhen there are many unlabeled data.
展开▼