一种基于主动学习的开放集图像识别方法  

Open Set Image Recognition Method Based on Active Learning

作　　者：王慧敏 王智强[1,2] 郭婷[1,2] 梁吉业 WANG Huimin;WANG Zhiqiang;GUO Ting;LIANG Jiye(School of Computer and Information Technology,Shanxi University,Taiyuan 030000,China;Key Laboratory of Computational Intelligence and Chinese Information Processing(Shanxi University),Ministry of Education,Taiyuan 030000,China)

机构地区：[1]山西大学计算机与信息技术学院,太原030000 [2]计算智能与中文信息处理教育部重点实验室(山西大学),太原030000

出　　处：《小型微型计算机系统》2024年第10期2442-2448,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(62272285,62376141)资助.

摘　　要：开放集识别(Open Set Recognition,OSR)的主要目的是识别未标记数据中的新类样本,同时对已见类样本进行正确分类.现有的大多数识别方法对未标记数据的评估和伪标记信息的利用不足.本文提出一种基于主动学习的开放集图像识别方法(Open Set Image Recognition Method Based on Active Learning,AC-OSIR),充分利用未标记数据提升开放集识别性能.通过引入已见类别的语义知识,构建语义知识和图像特征的映射关系.对于未标记数据,利用阈值选择策略区分开放集样本和已见类样本,通过主动学习模型迭代地识别高置信度开放集样本和已见类样本,并将高置信度已见类样本添加到标记数据集中.本文在图像分类数据集CIFAR-10、TIN和LSUN,以及两个合成数据集的实验结果表明了基于主动学习的开放集图像识别方法的有效性.The main objective of Open Set Recognition(OSR)is to identify new class samples in unlabeled data and at the same time to correctly classify the seen class samples.Most of the existing recognition methods do not make sufficient use of the evaluation of unlabeled data and pseudo-labeled information.Therefore,in this paper,we propose an Open Set Image Recognition Method Based on Active Learning(AC-OSIR)to improve the open set recognition performance by making full use of unlabeled data.By introducing the semantic knowledge of seen classes,the mapping relationship between semantic knowledge and image features is constructed.For unlabeled data,we use a threshold selection strategy to distinguish open-set samples from seen-class samples,and then iteratively identify high-confidence open-set samples and seen-class samples by an active learning model,and add high-confidence seen-class samples to the labeled dataset.In this paper,experimental results on image classification datasets CIFAR-10,TIN and LSUN,and two synthetic datasets demonstrate the effectiveness of the open-set image recognition method based on active learning.

关 键 词：开放集识别 语义知识 主动学习 阈值选择 图像识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]