融合语义与图像的大规模图像集检索算法  

Retrieval Algorithm of Big Scale Image Set by Probabilistic Fusion of Semantic Content and Visual Content

作　　者：解姗姗[1] 神显豪[2] XIE Shanshan;SHEN Xianhao(School of Information Management,Minnan Institute of Technology,Shishi 362700,China;Department of Information Science and Engineering,Guilin University of Technology,Guilin 541004,China)

机构地区：[1]闽南理工学院信息管理学院,福建石狮362700 [2]桂林理工大学信息科学与工程学院,广西桂林541004

出　　处：《重庆理工大学学报（自然科学）》2019年第7期178-186,共9页Journal of Chongqing University of Technology：Natural Science

基　　金：国家自然科学基金资助项目(61741303);广西自然科学基金资助项目(2016GXNSFAA380153)

摘　　要：目前的大数据图像检索算法大多仅支持单一的关键词或者图像查询,为此提出了一种语义与图像概率融合的社交媒体图像检索算法。该算法提取图像的空间位置特征与颜色特征,并提取文字标注信息,将两种特征基于概率进行融合。为了解决图像标注缺失与标注噪声的问题,设计了新的主题模型,根据共生的标注信息与视觉特征提取图像的语义主题。主题模型中基于视觉特征生成的最近主题能够有效地增强图像与文字标注之间的相关性。此外,主题模型能够有效地补全缺失的文字标注信息,同时删除噪声标注。基于不同规模的数据集进行了仿真实验,结果显示:该算法支持单一的关键词查询、图像查询以及两者的组合查询,并实现了较高的检索准确率。Most of the current image retrieval algorithms of big data only support keywords or image request,a retrieval algorithm of big scale image set by probabilistic fusion of semantic content and visual content is proposed.The proposed algorithm extracts the spatial location feature,color feature and text annotation information and the abstracted features are probabilistically fused.In order to fix the issue of annotation loss and annotation noise,a new topic model is designed and the semantic topic of the image is extracted based on the co-occurrence annotations and visual features.The latent topic generated based on visual features enhances the correspondence between tags and the image content,besides.The topic model guarantees the text annotation completion effectively,and deletes noisy annotations.Simulation experimental results based on different scales of datasets indicate that the proposed algorithm supports the query of keywords,image or combination of both.At the same time,it realizes an ideal retrieval accuracy.

关 键 词：社交图像 图像视觉特征 主题模型 文字标注 半监督学习 图像检索 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]