基于伪相关反馈的主动样例挑选.pdfVIP

基于伪相关反馈的主动样例挑选 张龙飞曹元大 北京理工大学计算机科学与工程学院北京100081 {longfeizhang，ydcao@bit．edu．cn} 摘要：相关反馈技术提供了填补图像检索中低级特征与高层语义之间的巨大鸿沟的方法。本文从解决 反馈图像信息量不足入手，提出了一种利用簇的主动学习构建伪相关反馈样例挑选算法。首先利用KNN 将样例聚类成簇，在簇挑选后进行基于簇主动学习的伪相关反馈方法构建SVM超平面，挑选超平面附近样 例进行反馈，并完成SVM检索器的构造。利用多例主动学习思想减小了主动学习中分类器抖动问题，在检 索中表现优于伪相关反馈算法和主动学习算法。 关键词：伪相关反馈 主动样例挑选相关反馈 图像检索 Active BasedonPseudoRelevanceFeedback Sampling Yuan-da ZhangLong-fei，Cao of and Instituteof Science 0008 (DepartmentComputer Engineering，BeijingTechnology,11) {longfeizhang，ydeao@bit．edu．cn) Abstract：Anovel relevancefeedbackbasedactive is inthissectiontoo． pseudo samplingalgorithmproposed Afterthefirstretrieval is toclusterthe clustersareusedtotraina retriever,KNN byrough exploited samples，and relevancefeedback needtobe formthe of SVM retriever．The labeled，isselected pseudo samples，which neighbor the ofSVMretriever．ThefinalretrieverSVMistrainedtheselabeled hyperplain by samples．Single-sample retrievalis into Pseudorelevancefee曲ack aetive changed algorithm．andMulti·samplelearning multi-sampleby reducedthe oftheshakeofclassifiers．The ofthefinalretrieval isbetterthan probability performance experiment relevancefeedbackretrieval andactiverelevancefeedback purepseudo algorithm algorithm． active relevancefeedback retrieval Words：Pseudorelevancefeedback Key