Visual domain adaptation based on modified [formula omitted]distance and sparse filtering

Abstract

Domain adaptation studies how to build a robust model to solve pattern recognition problems when training in a source domain while testing in a related but different target domain. The existing methods focus on how to shorten the distance between the two domains, however, they have limited considerations on the preservation of data structures. In this paper, we propose a novel model for unsupervised domain adaptation. For the reduction of domain discrepancy, we propose modified A−distance, which is computationally fast and can be optimized using gradient information. Moreover, in order to capture the internal structures of target samples, within-domain normalization based sparse filtering is raised, which proved to be more powerful for domain adaptation. Extensive experiments compared to both shallow and deep methods demonstrate the effectiveness of our approach.