基于显著区域的图像自动标注  

Automatic image annotation based on salient regions

作　　者：尹文杰[1] 韩军伟[1] 郭雷[1] 贺胜[1] 许明[1] 

机构地区：[1]西北工业大学自动化学院,西安710129

出　　处：《计算机应用研究》2011年第10期3925-3928,3933,共5页Application Research of Computers

基　　金：西北工业大学基础研究基金资助项目(JC201041);西北工业大学引进高层次人才科研启动费资助项目

摘　　要：为了提高图像自动标注的准确率,提出了一种基于图像显著区域的自动标注方法。首先提取图像的显著区域,然后提取图像的SIFT特征,利用K-均值聚类得到视觉词汇,并根据训练图像的SIFT特征是否位于显著区域进行不同的加权运算得到视觉词汇的词袋表示,最后利用支持向量机训练分类模型实现图像分类和标注。在一个包含1 255幅Corel图像的数据库进行实验,所提方法标注的准确率与整体考虑整幅图像特征相比有很大提高,表明提出的算法优于传统方法。In order to improve the performance of automatic image annotation,this paper presented a novel algorithm based on image saliency analysis.Firstly,computed image saliency and obtained the salient objects.Then,the SIFT features were extracted for every image,the visual words were derived through K-Means cluster algorithm,and thus the visual Bag-of-Words model weighted by the saliency information was built to characterize images.Finally,SVM was trained to classify and annotate images automatically.Compared with unweighted algorithm,experimental results on 1255 images from Corel database show that this algorithm improves the annotation accuracy.It demonstrates the proposed approach is promising.

关 键 词：图像自动标注 显著区域 SIFT特征 K-均值聚类 视觉词袋 支持向量机 

分 类 号：P37[天文地球—地球物理学]