SADD: Generative Adversarial Networks via Self-attention and Dual Discriminator in Unsupervised Domain Adaptation

Image classification is an active research in the field of computer vision. There are still significant challenges in the classification accuracy of cross-domain images due to the privacy, human, and material cost issues involved in labeled data collection, and the distribution differences among the collected images of the same category. To address the above problems, unsupervised domain adaptation (UDA) methods emerge, which transfer prior knowledge from the labeled source domain to the unlabeled target domain. In this work, we propose a new UDA architecture, SADD, which performs feature-level and pixel-level discrimination in a self-attention generative adversarial network. Specifically, we use the self-attention mechanism in extracting features to obtain globally dependent embeddings. In addition, we apply pixel-level distribution consistency loss on the embedding-generated images to mitigate the pixel-level distribution shifts due to unstable image style shifts. Further, we use discriminators for embedding reconstruction to assist the feature extractor in aligning features and enhancing the classification ability of the classifier. We evaluate our approach on the DIGITS classification dataset and the OFFICE-31 recognition dataset, and the results demonstrate the robustness and superiority of our approach.