Dictionary learning algorithms have been successfully used in bothreconstructive and discriminative tasks, where the input signal is representedby a linear combination of a few dictionary atoms. While these methods areusually developed under $\ell_1$ sparsity constrain (prior) in the inputdomain, recent studies have demonstrated the advantages of sparserepresentation using structured sparsity priors in the kernel domain. In thispaper, we propose a supervised dictionary learning algorithm in the kerneldomain for hyperspectral image classification. In the proposed formulation, thedictionary and classifier are obtained jointly for optimal classificationperformance. The supervised formulation is task-driven and provides learnedfeatures from the hyperspectral data that are well suited for theclassification task. Moreover, the proposed algorithm uses a joint($\ell_{12}$) sparsity prior to enforce collaboration among the neighboringpixels. The simulation results illustrate the efficiency of the proposeddictionary learning algorithm.
展开▼
