基于上下文相关模型的图像语义标注  

Contextual Model-based Semantic Image Annotation

作　　者：田东平[1,2] 李乃乾[1] 

机构地区：[1]宝鸡文理学院计算机软件研究所,陕西宝鸡721007 [2]宝鸡文理学院计算信息科学研究所,陕西宝鸡721007

出　　处：《小型微型计算机系统》2016年第4期855-860,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61202212)资助;陕西省自然科学基础研究计划项目(2014JQ2-6036)资助;陕西省教育厅专项科研计划项目(15Jk1038)资助

摘　　要：针对现有标注算法精度不高的问题,提出一种基于上下文相关模型的图像语义标注方法.首先,根据语义概念在训练数据集中的共现频率,对每个概念构造马尔科夫随机场图结构;其次,从已标注图像的文本信息出发构建一个非对等模态的概率潜语义分析(PLSA)模型,计算图像和语义概念的联合概率,并将其作为马尔科夫随机场(MRF)中点的观察值.与此同时,基于PLSA设计马尔科夫随机场模型的点势函数和边势函数;最后,通过正则化最大伪似然估计学习MRF的模型参数,利用迭代条件模式进行模型推理,从而获得未知图像的精确化语义标注结果.实验表明,所提出方法的性能明显优于若干经典的自动图像标注方法,而且具有更好的检索性能.Since the precision of existing annotation algorithms is not very high,a contextual model-based semantic image annotation method is proposed（ M RFP）.To begin with,the corresponding graph structure of Markov random fields（MRF） is constructed for each semantic concept respectively,which can effectively distinguish the annotations between each other.Subsequently,a probabilistic latent semantic analysis（ PLSA） model with asymmetric modalities is constructed from textual information of the annotated images and is employed to estimate the joint probabilities of images and semantic concepts,which will serve as observation values for each site of M RF.In the meanwhile,the site potential and edge potential functions are defined on this basis.Finally,the maximum pseudo-likelihood with regularization is utilized to learn the model parameters for each M RF and the iterative conditional mode（ ICM） for inference so as to achieve the refining semantic annotations for the unseen images.The conducted experiments of M RFPcompare favorably with several state-of-the-art approaches.Furthermore,it can yield best retrieval performance compared with other methods.

关 键 词：图像语义标注 马尔科夫随机场 概率潜语义分析 图像检索 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]