四川师范大学学报（自然科学版）2025，Vol.48Issue(6)：711-722,12.DOI:10.3969/j.issn.1001-8395.2025.06.001
测试时适应综述:从方法到应用
A Survey on Test-Time Adaptation:From Methods to Applications
摘要
Abstract
Under the strong assumption of independent and identically distributed(i.i.d.)data,deep learning models perform well when test data share similar distributions with training data.However,in real-world applications,models often encounter test data that deviate from the training distribution,leading to a problem of the distribution shift that significantly degrades model performance.In recent years,Test-Time Adaptation(TTA)has gained increasing attention as an effective technique to address distribution shift and improve domain generalization.Specifically,TTA refers to the technique of adaptively adjusting the model or feature distributions using unlabeled test data during the inference phase,aiming to enhance model performance in the test domain.This paper starts with the bas-ic concepts of TTA,reviews its development history,categorizes mainstream TTA methods,explores existing applications in various subfields,discusses current challenges in TTA researches,and provides an outlook on future directions.关键词
迁移学习/领域自适应/测试时适应/无源域适应Key words
transfer learning/domain adaptation/test-time adaptation/source-free domain adaptation分类
信息技术与安全科学引用本文复制引用
龚勋,吕金荣..测试时适应综述:从方法到应用[J].四川师范大学学报（自然科学版）,2025,48(6):711-722,12.基金项目
国家自然科学基金(62376231)和四川省自然科学基金(24NSFC1070) 中央高校基本科研业务费(2682025ZTPY052、2682023ZDPY001)对本文给予了资助,谨致谢意. （62376231）
