Abstract: Most of the source-free domain adaptive methods are self-supervised learning based on pseudo-labels to solve the source-free limitation.However,these methods overlook the impact of the clustering structure of target sample feature distributions and the uncertainty of samples near the decision boundary on pseudo-label noise during the generation process,which reduces the model’s performance.Therefore,this paper proposes a source-free domain adaptive method based on pseudo-label uncertainty estimation.Firstly,multiple perturbations are introduced to the model’s feature extractor parameters to simulate variations in source knowledge resulting from data fine-tuning.The similarity of target sample feature distributions under different perturbed models is utilized to assess the generalization uncertainty of the source knowledge.Moreover,an extreme value entropy is proposed to quantify the latent information uncertainty within the target domain,where distinct entropy calculation methods are applied based on the difference between the highest and second-highest prediction probabilities.Secondly,the target samples are divided into reliable samples and unreliable samples according to two kinds of uncertainty.For reliable samples,self-supervised learning is employed,using their prediction probabilities as weights to update sample features into class prototypes.Furthermore,historical class prototypes are incorporated to improve the stability of the class prototypes.Contrast learning is applied to unreliable samples to bring them closer to similar class prototypes.Compared with several baseline models,the classification accuracy is improved on three public benchmark datasets:Office-31,Office-Home and VisDA-C,and the effectiveness of the method is verified.

Key words: Domain adaptive, Source-free data, Uncertainty estimation, Information entropy, Contrastive Learning