一种基于群稀疏特征选择的图像检索方法  

Group sparse based feature selection for image retrieval

作　　者：郑秋中 徐军[1] 

机构地区：[1]南京信息工程大学信息与控制学院,南京210044

出　　处：《计算机应用研究》2014年第9期2867-2872,共6页Application Research of Computers

基　　金：国家自然科学基金资助项目(61273259);南京信息工程大学引进人才启动基金资助项目

摘　　要：针对图像检索问题,提出一种基于群稀疏特征选择的图像检索方法。利用特征具有的群聚与稀疏的特性,构建一个L2,1范数正则化逻辑回归问题,运用自适应谱梯度算法(ANSPG)有效地求解权重,根据这个权重选择出有效特征。最后运用所选择的有效特征在基于内容的检索框架上进行图像检索。在Core15K与IAPR IC12图像库上进行的实验结果表明,提出的方法具有非常良好的性能。Based on group sparse representation technique, this paper proposed a new image feature selection algorithm on searching image by using the selected feature for image retrieval. First of all ,it proposed a regulafization-based logistic regres- sion by utilizing both the sparsity and clustering properties. The weight could be effectively worked out by the adaptive spectral gradient algorithm （ANSPG）. Then it selected the most effective features according to the size of the weights. Finally,it em- ployed the selected effective features for CBIR framework. The experimental results on both CorelSK and IAPR IC12 image setsshow that the proposed algorithm has very good performance.

关 键 词：基于内容的图像检索 特征选择 逻辑回归 群稀疏表示 L2 1范数正则化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]