Abstract:
To deal with the problem that the data samples with non-linear distribution are poorly separable in the original feature space, in this paper, a kernel transfer sparse soding algorithms is proposed for cross-domain image classification.Firstly, the image features and the dictionary are mapped to a high-dimensional reproducing-kernel hilbert space, which makes the problem of linear inseparable problem become a linear separable problem. Then, the samples are individually represented in high-dimensional feature space.The proposed algorithm not only effectively handles the the nonlinear structure data, but also considers the distribution differences and geometric structure information between source and target domains, which gain more robust sparse representation and improve cross-domain image classification accuracy.