用于图像检索的三种分类器方法及其性能评价*

作者：王卫伟1，刘 伟1*，徐伟栋1，张 娟2，邵国良2 日期：2010-09-08/span> 浏览：3507 查看PDF文档

用于图像检索的三种分类器方法及其性能评价*

王卫伟1，刘 伟1*，徐伟栋1，张 娟2，邵国良2

（1.杭州电子科技大学 自动化学院，浙江 杭州 310018； 2.浙江省肿瘤医院 放射科，浙江 杭州 310022）

摘要：为了研究基于不同分类器的基于内容图像检索（CBIR）方法检索结果之间的关系，针对3种基于不同分类器的CBIR方法 — 基于解析特征相似性的k近邻方法、基于学习特征相似性的BP神经网络方法和基于信息论的互信息方法，分析研究了它们各自的检索性能以及它们之间检索结果的相关度和权重相关度（相关度描述不同CBIR方法检索到相同ROI占返回ROI总数中的比例信息，权重相关度则描述这些相同的ROI在各自检索结果中的不同排序位置信息）。实验结果表明，KNN，BPANN和MI之间检索结果相关度较差，当返回15个ROI时，平均查准率分别为72.6%，70.7%和68.9%，KNN与MI，KNN与BPANN以及MI与BPANN之间检索结果相关度分别为7.09%，9.60%和14.37%，权重相关度分别为0.011，0.023和0.039。这表明，由于基于不同分类器，不同CBIR方法可能会检索到视觉上和排列顺序上非常“不同的”相似图像。
关键词：基于内容图像检索；分类器；性能评估；相关度；权重相关度
中图分类号：TP391.4
文献标识码：A文章编号：1001-4551(2010)07-0047-06

Three classifiers and their performance evaluation for image retrieval
WANG Weiwei1, LIU Wei1, XV Weidong1, ZHANG Juan2, SHAO Guoliang2
(1. School of automation, Hangzhou Dianzi University, Hangzhou 310018, China; 
2. Department of Radiology, Zhejiang Cancer Hospital, Hangzhou 310022, China)

Abstract: Aiming at studying the retrieval results relation of different classifiers based contentbased image retrieval(CBIR) methods, a preliminary analysis study of the retrieval performance, the association degree and the weight association degree of three CBIR methods (namely analytical feature similarity based KNN method, learning feature similarity based BPANN method and information theoretic similarity based mutual information method) based on different classifiers for CBIR were presented. The association degree described the proportion of the same ROIs in retrieved ROIs for different CBIR methods, while the weight association degree showed the different sort position information for these ROIs in their retrieval results. The experimental results demonstrate that that association degree of retrieval is poor. The average precision of KNN, BPANN and MI are 72.6%, 70.7% and 68.9% respectively, the association degree for KNN vs MI, KNN vs BPANN and MI vs BPANN are 7.09%, 9.60% and 14.37% respectively and the weight association degree for them are 0.011，0.023 and 0.039 respectively when top 15 most similar ROIs are selected. The study indicates that different CBIR methods can retrieve quite different visually “similar” and order ROIs due to the difference in classifiers.
Key words: contentbased image retrieval(CBIR); classifier; performance evaluation; association degree; weight association degree