图像检索中基于记忆与半监督的主动相关反馈算法  

Memorization and Semi-Supervision Based Active Relevance Feedback Algorithm for Content-Based Image Retrieval

作　　者：周艺华[1] 曹元大[1] 魏本杰[1] 张洪欣[2] 

机构地区：[1]北京理工大学计算机科学技术学院,北京100081 [2]北京邮电大学通信网络综合技术研究所,北京100876

出　　处：《北京理工大学学报》2006年第1期45-48,共4页Transactions of Beijing Institute of Technology

基　　金：国家自然科学基金资助项目(60271018)

摘　　要：为快速提高相关反馈算法的效率,提出一种记忆与半监督相结合的主动相关反馈算法.在检索初期,利用记忆信息获得较多的正训练样本,利用用户已标记样本与数据库内未标记样本有效地解决训练样本不平衡问题,获得准确的初始SVM分类器;在检索后期,利用主动学习算法寻找数据库内对优化学习过程中最有用的样本请求用户标记,减少用户标记的样本量,加快收敛速度.对5 000幅Corel图像数据库的实验表明,与传统相关反馈算法相比,新算法能够显著提高学习器的效率和性能,并快速收敛于用户的查询概念.To improve the efficiency of relevance feedback quickly, an integrated memorization and semi-supervision active relevance feedback algorithm is presented. In its early stage, more positive samples are obtained through memorization. The problem of biased training samples is solved efficiently through labeled and unlabeled training samples and accurate initial SVM classifier is obtained; in the later stage, samples required for labeling by users reduced largely and convergent rate improved greatly by the active learning algorithm which selects the most useful samples in database to solicit the user for labeling. Experimental results on 5 000 Corel images library showed that the proposed algorithm can greatly improve the efficiency and accuracy and converge to user＇s query concept quickly.

关 键 词：基于内容的图像检索 半监督学习 长期学习 主动学习 相关反馈 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]