文献类型：期刊文献

中文题名：多模型融合的多标签图像自动标注

英文题名：Multi-model Fused Framework for Image Annotation

作者：张静[1,2];胡微微[1];陈志华[1];袁玉波[1]

机构：[1]华东理工大学计算机科学与工程系,上海200237;[2]南京大学计算机软件新技术国家重点实验室,南京210093

年份：2014

卷号：26

期号：3

起止页码：472

中文期刊名：计算机辅助设计与图形学学报

外文期刊名：Journal of Computer-Aided Design & Computer Graphics

收录：CSTPCD;;EI(收录号：20141417541448);Scopus;北大核心:【北大核心2011】；CSCD:【CSCD2013_2014】；

基金：国家自然科学基金(61370174);上海市自然科学基金(11ZR1409600);浙江大学CAD&CG国家重点实验室开放课题(A1213)

语种：中文

中文关键词：图像标注;多模型;MNKDA;区域语义分析;语义相关分析

外文关键词：image annotation; multi-model; multi-feature based saliency analysis; multiple Nystr（o）m-approximating kernel discriminant analysis （MNKDA） ; region semantic analysis

摘要：为了实现更为准确的复杂语义内容图像理解,提出一种融合多模型的多标签图像自动标注方法.该方法采用3个不同的模型分别对图像语义内容进行分析:在前景语义概念检测中,提出一种基于多特征的视觉显著性分析方法,并利用多Nystrm近似核对前景对象的语义进行判别分析;对于背景概念检测,提出一种区域语义分析的方法;通过构造基于潜语义分析的语义相关矩阵来消除标注错误的标签.根据前景和背景的语义和视觉特征,分别采用不同的模型提取前景和背景标注词,而语义相关分析能够有效地提高标注的准确性.实验结果表明,该多模型融合标注方法在图像的深层语义分析以及多标签标注方面具有较好的效果;与同类算法相比,能够有效地减少错误标注的标签数目,得到更加准确的标注结果.
Automatic image semantic annotation has become an important research topic that attracts widespread attention due to the existence of semantic gap.We propose a new framework for image multi-label annotation,in which image annotation is divided into two parts：foreground concepts detection and background labels annotation.A new multi-feature fusion based visual saliency analysis algorithm is proposed for foreground region detection in this paper,which is the basis for accurate foreground annotation.We also propose a new region semantic analysis algorithm for background labels annotation.Furthermore,a semantic correlation model based on latent semantic analysis is proposed to remove the wrong labels to achieve more accurate annotation results.Our automatic image annotation framework which fuses three different models has been evaluated on the Corel 5K and Pascal VOC2007 databases,and compared with previous algorithms.Experimental results show that the proposed multi-model method can achieve promising performance and significantly outperforms previous algorithms.