CBIR中一种基于最近邻的改进相关反馈算法  

Improved relevance feedback algorithm based on nearest neighbor approach in CBIR

作　　者：易唐唐[1] 黄立宏[1,2] 

机构地区：[1]湖南女子学院信息技术系,长沙410004 [2]湖南大学数学与计量经济学院,长沙410082

出　　处：《计算机应用研究》2015年第8期2326-2330,共5页Application Research of Computers

基　　金：国家科技支撑计划资助项目(2012BAH08B00);湖南省教育厅科学研究青年项目(12B066)

摘　　要：针对CBIR中基于NN的相关反馈技术在迭代时由于信息缺乏而产生的小样本问题,提出了一种基于最近邻的改进相关反馈算法。首先,根据检索图像和相关后验概率定义随机变量,从而获得相关性正确概率;然后,将与检索图像距离最近的样本映射到概率估计;最后,利用正则化得到平滑估计,解决了小样本问题。将算法在多种不同的情况下进行测试,实验结果表明,相比标准的NN方法和其他现有的相关反馈机制,在大多数情况下该算法具有显著的性能改善。For the small simple size problem caused by the lack of information of relevance feedback technology based on NN （nearest neighbor） when iterating in CBIR（ content-based image retrieval）, this paper proposed an improved relevance feedback algorithm based on the nearest neighbor approach. Firstly, it defined a random variable by retrieve images and related posteriori probability so as to getting correlation correctly probability. Then, it mapped the sample nearest to retrieve images to probability estimate. Finally, it used regularization to get smoothed estimate, overcoming the small sample size problem. The proposed algorithm has been tested in a number of different situations. Experimental results evidence significant improvements compared to the standard NN approach and other existing relevance feedback mechanisms in most cases.

关 键 词：相关反馈技术 基于内容的图像检索 最近邻 小样本问题 交互式 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]