Distribution Regularized Nonnegative Matrix Factorization for Transfer Visual Feature Learning

Transfer visual feature learning (TVFL), which learns compact representations for images such that we can build accurate classifier for target domain by leveraging rich labeled data in the source domain, has attracted increasingly attention recently. Previous methods mainly focus on reducing the distribution difference between domains but ignore the intrinsic hidden semantics in data. In this paper, we put forward a novel method for TVFL, called Distribution Regularized Nonnegative Matrix Factorization (DRNMF). Specifically, we employ Nonnegative Matrix Factorization (NMF) to uncover the intrinsic information in visual data, and regularize it with geometrical distribution, marginal probability distribution and conditional probability distribution. Thus, DRNMF can discover the intrinsic information, preserve the manifold structure and reducing both marginal and conditional probability distribution difference simultaneously, which all perspectives above are important for TVFL. We also propose an effective and efficient algorithm for the optimization of DRNMF and theoretically prove the convergence. Extensive experiments on three types of cross-domain image classification tasks in comparison with several state-of-the-art methods demonstrate the superiority of our DRNMF, which validates its effectiveness.