基于扩展生成语言模型的图像自动标注方法  

Image Auto-Annotation via an Extended Generative Language Model

作　　者：王梅[1] 周向东[1] 张军旗[1] 许红涛[1] 施伯乐[1] 

机构地区：[1]复旦大学计算机与信息技术系,上海200433

出　　处：《软件学报》2008年第9期2449-2460,共12页Journal of Software

基　　金：国家自然科学基金;国家重点基础研究发展计划(973)~~

摘　　要：使用最大权匹配算法,结合统计平滑技术,提出图像区域特征生成概率估计方法,并进一步对训练集中标注词之间的语义相关性(correlation)进行分析与度量,给出一种基于生成模型的图像标注算法.算法使用所提出的基于最大权匹配的图像生成概率估计方法得到较好的起始点,进而设计启发式迭代函数对词与词的相关性加以利用,最终提高标注词与图像的语义相关性.在现实世界图像数据库上的实验结果验证了所提出标注方法的有效性.In this paper, based on the statistical smoothing strategy, a image region feature generative prooability estimation method is proposed by exploiting maximum weight matching algorithm. By further analyzing and measuring the semantic correlations between words based on the training set, a novel image annotation algorithm for adopting the generative model is presented. The ftrst annotation keyword is obtained by using the proposed image region feature generative probability estimation algorithm. Then, a heuristic iterate function is proposed to exploit the keyword semantic correlation. Finally, the semantic correlation between the annotation and the image can be improved by our annotation algorithm. The proposed annotation approach is tested on a real-world image database, and promising results are achieved.

关 键 词：图像标注 生成模型 连续特征估计 最大权匹配 语义相关性 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]