Self-Training Cross-Domain Image Classification Model Via Label Adaptation

Unsupervised domain adaptation focuses on transferring knowledge from a labeled source domain to a completely unlabeled target domain, with the goal of enhancing classification performance in the target domain. Recently, the teacher-student self-training framework, a popular semi-supervised learning method, has shown its effectiveness in domain adaptation. In this framework, the student model is trained using pseudo-labels generated by the teacher model. However, during the initial training phase, the student model relies solely on the source domain, which can create a bias in the generator’s feature space towards the source. This bias makes it challenging for the teacher model to produce high-quality pseudo-labels, which hampers the student model's learning process in the target domain. To resolve this issue, this paper proposes a self-training domain adaptation model that incorporates label adaptation. Specifically, we introduce a label adaptation module within the student model, enabling it to integrate target domain information into the source domain. This helps the student model improve cross-domain learning and addresses the problem of domain shift. Additionally, a clustering module is introduced in the teacher model to learn more compact features, facilitating better classification of features that lie on the boundary in the target domain. Experimental results demonstrate that our method significantly enhances classification performance.