Adversarial domain adaptation with semantic consistency for cross-domain image classification

Abstract

In the cross-domain image classification scenario, domain adaption aims to address the challenge of transferring the knowledge obtained from the source domain to the target domain that is regarded as similar but different from the source domain. To get more reliable domain invariant representations, recent methods start to consider class-level distribution alignment across the source and target domains by adaptively assigning pseudo target labels. However, these approaches are vulnerable to the error accumulation and hence unable to preserve cross-domain category consistency. Because the accuracy of pseudo labels cannot be guaranteed explicitly. In this paper, we propose Adversarial Domain Adaptation with Semantic Consistency (ADASC) model to align the discriminative features across domains progressively and effectively, via exploiting the class-level relations between domains. Specifically, to simultaneously alleviate the negative influence of the false pseudo-target labels and get the discriminative domain invariant features, we introduce an Adaptive Centroid Alignment (ACA) strategy and a Class Discriminative Constraint (CDC) step to complement each other iteratively and alternatively in an end-to-end framework. Extensive experiments are conducted on several unsupervised domain adaptation datasets, and the results show that ADASC outperforms the state-of-the-art methods.