In the realm of Unsupervised Domain Adaptation (UDA), adversarial learning has achieved significant progress. Existing adversarial UDA methods typically employ additional discriminators and feature extractors to engage in a max-min game. However, these methods often fail to effectively utilize the predicted discriminative information, thus resulting in the mode collapse of the generator. In this paper, we propose a Dynamic Balance-based Domain Adaptation (DBDA) method for self-correlated domain adaptive image classification. Instead of adding extra discriminators, we repurpose the classifier as a discriminator and introduce a dynamic balancing learning approach. This approach ensures an explicit domain alignment and category distinction, thus enabling DBDA to fully leverage the predicted discriminative information for an effective feature alignment. We conducted experiments on multiple datasets, therefore demonstrating that the proposed method maintains a robust classification performance across various scenarios.

Unsupervised domain adaptation (UDA) aims to leverage labeled source domain knowledge to improve the target domain's performance. Source-free domain adaptation (SFDA), a recent research focus, addresses challenges such as data privacy by relying solely on a pretrained source model and unlabeled target data, thus eliminating the need for direct access to source domain data. Although many studies have proposed methods such as generating a source domain, using a proxy source domain, or using pseudo-label training, these approaches directly fine-tune the source model, overlooking the excessive bias towards the source domain data in SFDA. The source domain model contains numerous source domain-specific features, and directly updating it to shift towards the target domain is hindered by these domain-specific features. To address this issue, we propose an inverse distillation-based SFDA method. By constructing an initial target domain model, the method extracts pure target domain features and distills them back into the source model, facilitating a smoother transition towards the target domain. Additionally, it identifies stable and active target samples from both structural and scoring perspectives, applying distinct matching strategies for pseudo-label selection. Extensive experiments and ablation studies on public datasets (Digits, Office-31, Office-Home and VisDA-2017) demonstrate the superior performance of our approach in SFDA tasks.