Domain adaptation has been widely used in the field of computer vision. Current methods of domain adaptation mainly aim to reduce the difference of the marginal and conditional distributions of the source and target domains in a centralized manner. However, most of the existing domain adaptation methods ignore the correlation information of the two domains, or doesn't take it very seriously. That making it difficult to learn related features from the source domain for the target task. A new method, canonical correlation cross-domain alignment (CCCA), is proposed to effectively reduce the cross-domain distribution difference by combining the least squares formula of CCA with domain adaptation. In CCCA, a common latent subspace with the maximum correlation is learned to ensure that the learned features are from the two domains with maximum correlation. A Laplace graph is learned to maintain the structural consistency of CCCA. To verify the performance of our method, we conduct experiments on several benchmark visual databases. The experimental results illustrate its superiority to several other methods.
展开▼
