基于结构量化直方图的图像检索方法  

Image Retrieval Based on Structure Quantification Histogram

作　　者：金莲芳[1] 覃团发[1] 王海霞[1] 帅勤[1] 

机构地区：[1]广西大学计算机与电子信息学院,南宁530004

出　　处：《中国图象图形学报》2006年第2期180-185,T0001,共7页Journal of Image and Graphics

基　　金：广西科学研究与技术开发计划资助项目(桂科攻0330010-1);广西"新世纪十百千人才工程"专项资金资助项目(2002222)

摘　　要：颜色结构描述符虽然保留了颜色的结构信息,但是对8×8像素结构化元素内的颜色内容处理较粗糙,为了区分具有相同颜色结构直方图的完全不同的两幅图像,提出了一种基于结构量化直方图的检索方法,指出颜色结构直方图和全局直方图分别是结构量化直方图取量化步长为1和0时的特例。对3者在HMMD(hue-m ax-m in-d ifference)颜色空间采用不同量化维数(32,64,128,256)的2 600张花卉图库检索实验中的平均检索率(RARR)以及平均归一化修正检索等级(GANMRR)性能进行了分析,检索结果表明量化步长取0.5时的结构量化直方图的性能最优。另外还详细介绍了MPEG-7标准的颜色评价准则并以此作为评判检索结果优劣的依据。Although color structure descriptor contains the structural information of color, it is crude for processing the color contents of the 8 × 8 structural elements to distinguish two different pictures with the same color structure histogram. In this paper, a method for image retrieval based on structure quantification histogram is presented. The color structure t histogram is one specific case of structure quantification histogram, which is endowed an extreme quantification step as 1. While the color histogram is the other one of structure quantification histogram, which is endowed an extreme quantification step as 0. The performances of the three types of histogram are discussed through average retrieval rate（RARR ） and average normalized modified retrieval rank（ GANgss ） including 2 600 flower pictures with different quantification dimensions（32,64, 128,256） in HMMD color space. The results show that the performance of structure quantification histogram of step size 0. 5 is the best. Evaluation criterion used in MPEG-7 is introduced in detail and is used to evaluate the experimental results.

关 键 词：MPEG-7 结构量化直方图 颜色结构直方图 全局直方图 平均检索率 平均归一化修正检索等级 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]