Center-aligned domain adaptation network for image classification

Abstract

For a target task where the labeled data are unavailable, unsupervised domain adaptive learning performs transfer learning from labeled source data to unlabeled target data. Previous deep domain adaption methods mainly learned the global domain shift between different domains, the global distributions are aligned without considering the correspondence information between the same class data of different domains. Recently, more and more researchers pay attention to semantic alignment that focuses on accurately aligning the distributions of the same class data from different domains. However, most of them ignore two points: the learning of the global distribution of the target domain data; the compactness of intra-class domain data and the discrimination of inter-class domain data, which lead to unsatisfying transfer learning performance. To resolve this problem, we propose a Center-aligned Domain Adaptation Network (CenterDA) to facilitate the semantic alignment, In this study, for each class in label space, we learn a common class center for all data with the same class label in the source and target domains, which allows us to learn the global distribution of the target domain data under the supervised learning of the source domain data. Furthermore, we minimize the distance between the deep features and its common class center to compact the feature representations of data. In this manner, we achieve the desired goals: The global distribution of the target domain data is learned by common class center. Second, the source and the target domain data of the same class are aligned near the common center. Third, we model the intra-class compactness and the inter-class separability modeling. Extensive experiments on three datasets show that our method achieves remarkable results on image classification and has comparable performance with the latest methods.