文献类型：期刊文献

中文题名：基于无监督域自适应的计算机视觉任务研究进展

英文题名：A survey on unsupervised domain adaptation in computer vision tasks

作者：孙琦钰[1];赵超强[1];唐漾[1];钱锋[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237

年份：2022

卷号：52

期号：1

起止页码：26

中文期刊名：中国科学：技术科学

外文期刊名：Scientia Sinica(Technologica)

收录：CSTPCD;;EI(收录号：20220511543430);Scopus;北大核心:【北大核心2020】；CSCD:【CSCD2021_2022】；

基金：国家自然科学基金基础科学中心项目(批准号:61988101);国家杰出青年科学基金(批准号:61725301);上海市优秀学术带头人计划项目(编号:20XD1401300);高等学校学科创新引智计划(编号:B17017);中央高校基本科研业务费专项资金(编号:222202117006)资助。

语种：中文

中文关键词：无监督域自适应;计算机视觉;深度学习;迁移学习;自主系统

外文关键词：unsupervised domain adaptation;computer vision;deep learning;transfer learning;autonomous system

摘要：作为工业互联网的典型实例之一,车联网技术近年来飞速发展,其核心在于信息的互联互通.因此,精准、可迁移的环境信息感知能力是其稳定运行的前提之一.深度学习的进步推动了计算机视觉任务的发展,但基于传统深度学习的方法仍存在训练过程对人工标注数据依赖强、场景泛化能力较差的弊端.而对于计算机视觉任务来说,训练数据的真值标签获取较难,因此如何提升模型的迁移能力,缓解训练对人工标注的依赖受到了学界的广泛关注.无监督域自适应方法使用深度学习模型进行特征提取和对齐,使得深度学习模型在不同域间迁移时仍能保证良好的性能,在计算机视觉任务中发挥了重要作用.因此,本综述主要聚焦无监督域自适应方法在一些典型计算机视觉任务中的挑战和应用.首先,介绍了基于深度学习的无监督域自适应方法的定义、重要意义、应用难点、基本方法和相关数据集.然后,分别针对典型计算机视觉任务介绍了无监督域自适应方法在其中的应用.最后,进行了总结和展望.
As one of the typical applications of Industrial Internet,Internet of Vehicles(IoV)develops rapidly in recent years.It relies on the interconnection of information,where the transferability of the accurate perception is of great importance.Though deep learning has accelerated the development of computer vision tasks,traditional deep learning-based methods still have strong reliance on manually annotated training data and are poor at generalizing knowledge to new environments.For computer vision tasks,since it is difficult to collect training data with ground truth,it is agent to improve the generalization ability of deep learning models and alleviate their dependence on manually annotated labels.Unsupervised domain adaptation(UDA)methods apply deep learning models to extract and align features from data in different domains,which ensures the satisfactory generalization performance of deep learning-based computer vision algorithms.This paper focuses on the challenges and applications of UDA in some typical computer vision tasks.Firstly,the definition,significances,application difficulties,basic methods and relevant datasets of deep learning-based UDA methods are introduced.Then,the mainstream UDA methods in typical computer vision tasks are introduced separately.Finally,the prospective technical development trends and a summary are given.