摘要
Abstract
Objective Unsupervised domain adaptation(UDA)aims to adapt knowledge from labeled source domains to unlabeled target domains,achieving comparable performance in target tasks.Previous UDA methods face risks of privacy leakage and confusion among classes with similar visual appearances in the target domain.This paper proposed a source-free domain adaptation method based on masked image consistency,called Masked Hypothesis Transfer(MHT).Methods MHT adopted the idea of hypothesis transfer(HT)by freezing the classifier module(hypothesis)of the source model and learned the target feature extractor through information maximization and self-supervised pseudo-labeling to implicitly align representations between the target and source domains.Additionally,a masked image consistency(MIC)module was introduced to explicitly enforce the model to learn spatial contextual relationships in the target domain to enhance hypothesis transfer(HT).MIC forced consistency between the predictions on masked target images and the pseudo-labels,so the model must learn to infer predictions from the context of the masked region.Results The algorithm was extensively tested on four public benchmarks under closed-set UDA,partial-set UDA,and multi-source UDA settings.It achieved 87.6％accuracy on VisDA-C,outperforming SHOT by 5.3％,and 72.6％accuracy on Office-Home,surpassing SHOT by 0.9％.Conclusion Experimental results demonstrate that the target of masked image consistency serves as an additional clue to enhance source-free domain adaptation and MHT outperforms other comparative methods.关键词
无监督域适应/无源域适应/掩蔽图像建模/迁移学习Key words
unsupervised domain adaptation/source-free domain adaptation/masked image modeling/transfer learning分类
信息技术与安全科学
