Class Discriminative Maximum Mean and Covariance Discrepancy for Unsupervised Domain Adaptation

Unsupervised domain adaptation is a hot topic in transfer learning, which aims to transfer knowledge from the labeled source domain to the related but unlabeled target domain. For reducing the difference between the distributions of both domains, it is important to choose a suitable distribution discrepancy metric. Recently, a new distribution discrepancy metric named maximum mean and covariance discrepancy (MMCD) was proposed, which combined traditional maximum mean discrepancy (MMD) with maximum covariance discrepancy (MCD). Since the MMCD captures both the first and second order statistics in a reproducing kernel Hilbert space, it can give better results than using MMD alone. However, existing MMCD based methods ignored the label information from the source domain, thus limited their classification performance. In this paper, a class discriminative MMCD domain adaptation (CD-McDA) method is proposed, which not only reduces the MMCD of marginal and class-conditional distributions between both domains, but also enhances the class discriminative ability, i.e. maximizing the inter-class dispersion and minimizing the intra -class scatter. Comprehensive experiments are carried out on two visual and one pressure vessel crack cross-domain classification tasks, and the results demonstrate the effectiveness of the proposed CD-McDA method.