Common Dictionary and Domain-Specific Dictionary based Cross-Domain Image Classification

Traditional image classification methods require the independence and the same distribution of training and testing data. However, this requirement cannot always be satisfied in some real-world applications, especially for cross domain image classification tasks. In this paper, we propose to deal with this problem by combining transfer learning with sparse coding and dictionary learning. In dictionary learning, based on the idea of "low-coupled dictionary learning" of source domain and target domain data, we train a specific and discriminative dictionary for each domain, and a common dictionary for both domains. In addition, we keep these dictionaries independent and obtain the domain-invariant sparse coefficients as well as the domain-specific coefficients. Finally, the extensive experiments on the Office, Caltech256, USPS and MNIST datasets verify the effectiveness of our algorithm.