摘要： 针对图像自动标注应用,提出一种基于金字塔匹配核(PKM)的颜色直方图特征标注方法。图像以金字塔方式划分成越来越细的格子,统计每个格子的颜色直方图并进行对数处理,以平衡不同颜色对直方图的贡献。所有格子的直方图经过加权处理后连接成单一向量,称为颜色直方图金字塔,通过直方图相交距离度量向量之间的相似度。在corel5k数据集上的实验结果表明,该方法在全局颜色直方图、分块颜色直方图和尺度不变特征变换特征方面优于传统方法,与同样基于PKM的空间金字塔匹配(SPM)方法相比,F度量值提高了10%。

关键词: 颜色字典, 颜色直方图, 金字塔匹配核, 图像标注, 直方图相交距离, 区域特征

Abstract: This paper presents a color histogram feature annotation method based on Pyramid Match Kernel(PKM) for automatic image annotation.This technique works by partitioning the image into increasingly fine grids and computing color histogram within each grid,then applying a binary logarithm transformation to color histogram for regularizing the contribution of each color in the final descriptor.The histograms of all grids are connected to a single vector after weighted processing,named Color Histogram Pyramid(CHP).The similarity between vectors can be measured by histogram intersection distance.Experimental results on corel5k dataset show that the proposed method outperforms traditonal methods on global color histogram,block-based color histogram,and features exacted by scale-invariant feature transform.Compared with Spatial Pyramid Matching(SPM) method based on PKM,F-measure is improved by 10%.

Key words: color dictionary, color histogram, Pyramid Match Kernel(PKM), image annotation, histogram intersection distance, region feature

中图分类号: 

• TP391.41