图像分布外检测研究综述  

Research on Image Out-of-Distribution Detection:A Review

作　　者：郭凌云 李国和[1,2] 龚匡丰[1,2] 薛占熬 GUO Lingyun;LI Guohe;GONG Kuangfeng;XUE Zhan′ao(College of Information Science and Engineering,China University of Petroleum,Beijing 102249;Beijing Key Laboratory of Petroleum Data Mining,China University of Petroleum,Beijing 102249;College of Computer and Information Engineering,Henan Normal University,Xinxiang 453007)

机构地区：[1]中国石油大学(北京)信息科学与工程学院,北京102249 [2]中国石油大学(北京)石油数据挖掘北京市重点实验室,北京102249 [3]河南师范大学计算机与信息工程学院,新乡453007

出　　处：《模式识别与人工智能》2023年第7期613-633,共21页Pattern Recognition and Artificial Intelligence

基　　金：克拉玛依科技计划项目(No.2020CGZH0009);中国石油大学(北京)克拉玛依校区科研基金项目(No.RCYJ2016B-03-001)资助。

摘　　要：分类器学习一般假设训练样本和预测样本具有独立同分布.由于该条件过强,实践中当分类器面向分布外(Out-of-Distribution,OOD)样本时容易导致预测错误.因此,对OOD检测进行深入研究就显得尤为重要.文中首先介绍OOD检测的概念及其相关研究领域,根据网络训练方式的差异性对有监督的检测方法、半监督的检测方法、无监督的检测方法和异常值暴露的检测方法进行系统概述.然后按照关键技术从神经网络分类器、度量学习和深度生成模型三方面总结现有OOD检测方法.最后讨论OOD检测未来的研究方向.Classifier learning assumes that the training data and the testing data are independent and identically distributed.Due to the overly stringent assumption,erroneous sample recognition of classifiers for out-of-distribution examples is often caused.Therefore,thorough research on out-of-distribution(OOD)detection becomes paramount.Firstly,the definition of OOD detection and the relevant research are introduced.A comprehensive overview of supervised detection methods,semi-supervised detection methods,unsupervised detection methods and outlier exposure detection methods is provided according to the difference of network training methods.Then,the existing OOD detection methods are summarized from the aspect of three key technologies:neural network classifiers,metric learning and deep generative models.Finally,research trends of OOD detection are discussed.

关 键 词：机器学习 深度学习 分布外(OOD)检测 图像识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]