Transfer Kernel Sparse Coding Based on Dynamic Distribution Alignment for Image Representation

Sparse coding based domain adaptation methods aim to learn a robust transfer classifier by utilizing the knowledge from source domain and the learned new representation of both domains. Most existing works have achieved remarkable results in solving linear domain shift problems, but have poor performance in nonlinear domain shift problems. In this paper, we propose a Transfer kernel sparse coding based on dynamic distribution alignment (TKSC-DDA) approach for cross-domain visual recognition, which incorporates dynamic distributed alignment into kernel sparse coding to learn discriminative and robust sparse representations. Extensive experiment on visual transfer learning tasks demonstrate that our proposed method can significantly out-perform serval state-of-the-art approaches.