Hybrid adversarial network for unsupervised domain adaptation

Recent advances suggest that adversarial domain adaptation has been embedding into deep neural networks to learn domain-transferable representations, which reduces distribution divergence in both the training and test samples. However, previous adversarial learning algorithms only resort to learn domain-transferable feature representation by bounding the feature distribution discrepancy cross-domain. These approaches, however, may lead to misalignment and poor generalization results due to without further exploiting class information and task-special adaptation. To cope with this issue, a joint adversarial learning with class information and domain alignment deep network architecture, is proposed which is named Hybrid Adversarial Network (HAN). Specifically, it incorporates a classification loss to learn a discriminative classifier, and a domain adversarial network learns a domain-transferable representation to reduce domain shift. Meanwhile, a CORAL loss is used to minimize the discrepancy between the covariance matrices in the two domains. Additionally, we introduce an adaptation layer for further boosting the performance of HAN model. Comprehensive cross-domain visual recognition experiments validate that our method exceeds the state-of-the-art methods on three real-world benchmark including Office-31, Office-Home, and ImageCLEF-DA datasets.