L2,1-norm and graph-regularization based low-rank transfer subspace learning

A novel L2,1-norm and graph-regularization based low-rank transfer subspace learning method was proposed in this paper. Firstly, by applying the L2,1-norm constraint on the reconstruction matrix during low-rank reconstruction, the key features embedded in the target domain can be better explored. In addition, the rotation invariant characteristic of L2,1-norm will gives the algorithm the capability of handling the images with different poses. Secondly, the graph-regularization was integrated in the object function to better utilize the local geometric information embedded in the training data. As a result, the classification performance can be further enhanced. Finally, to tackle the problem of incomplete feature space coverage problem result from the insufficient source domain data and ensure the robustness of reconstruction, we advocate grouping the target domain and source domain data to form a joint "dictionary". Extensive experiments on Caltech256, Office, CMU-PIE, COIL20, USPS, MNIST, VOC2007 and MSRC dataset validate the effectiveness and robustness of our algorithm. © 2018, Editorial Department of Control Theory & Applications South China University of Technology. All right reserved.