Improving Feature's Capability of Carrying Category-specific Information for Adversarial Domain Adaptation

Recent research has shown that generative adversarial networks (GANs) have been successfully applied in aligning features for domain adaptation. However, the extracted features might lose category-specific information, because they are distinguished as either a source or a target during the adversarial training. To address this issue, a two-stage training framework consisting of two sets of GANs and a dedicated classifier is proposed in this study. In the pretraining stage, we use an encoder-decoder-classifier structure to obtain discriminative and representative features of the source domain and use it as a reference in the subsequent training. In the adversarial training stage, two sets of GANs are used to align target-domain features with those of the source domain and transfer samples of the target domain to the source domain simultaneously. A dedicated classifier is trained along with the adversarial loss to force the generated features of the target domain to carry category-specific information, which significantly improves classification performance. The features of the source domain stay intact in the adversarial training stage. Thus, our approach can alleviate the training burden of GANs. The proposed method has been validated on digital datasets and office31 datasets. Experimental results demonstrate that an average accuracy of 96.5% and 95.1% is achieved, which leads to superior or comparable performance to state-of-the-art results.