Abstract:Unsupervised domain adaptation (UDA) aims to apply a trained model in the source domain to the target domain with only unlabeled data. Current UDA approaches learn domain-invariant features by aligning the source domain and target domain feature spaces via statistical difference minimization or adversarial learning. However, these constraints may result in the distortion of semantic feature structures and loss of class discriminability. To this end, this study proposes a new method called DAMPL. This method utilizes the CLIP model to inject textual descriptive information to deeply mine the semantic content of the image, and adopts a prompt learning paradigm for domain characteristics to effectively retain information specific to different domains, thus avoiding information loss. Additionally, the pseudo-labelling of the target domains are corrected via a semantic bootstrapping mechanism to reduce the inter-domain differences and enhance the generalization ability of the model. Finally, mutual information maximization loss (IML) is also introduced to preserve the feature distinguishability of the target domains. The final DAMPL method demonstrates optimal performance by achieving 83.8%, 79.7%, and 89.8% classification accuracy on the Office-Home, miniDomainNet, and VisDA-2017 datasets, respectively.