基于标注词语义与图像视觉的标签丰富算法  

Tagging Enrichment Algorithm Based on Tag Semantic and Image Visual

作　　者：孙登第[1,2] 葛美玲[1] 丁转莲[1] 罗斌[1,2] 

机构地区：[1]安徽大学计算机科学与技术学院,合肥230601 [2]计算智能与信号处理教育部重点实验室,合肥230039

出　　处：《小型微型计算机系统》2017年第4期886-890,共5页Journal of Chinese Computer Systems

基　　金：国家"九七三"重点基础研究发展计划项目(2015CB351705)资助;国家"八六三"高技术研究发展计划项目(AA0151042014)资助;国家自然科学基金项目(61402002)资助;安徽省自然科学基金项目(1408085QF120)资助;安徽大学信息保障技术协同创新中心2015年开放课题(ADXXBZ201511)资助

摘　　要：针对社会化图像初始标注不完备、不准确的问题,提出了一种基于标注词语义与图像视觉信息的社交图像标签丰富算法.该算法融合词频共现和Word Net语义层级两种方式挖掘标签之间的潜在联系,并由此根据已有的标注信息获取候选标签;再综合图像视觉特征和高层语义主题计算图像内容相似度,确定图像间的近邻关系;最后基于近邻图像和候选标签构造了一种新的软近邻投票规则,以刻画候选标签是否真正能描述图像内容.在数据集MIRFlickr-25k上进行实验,系统地比较了不同语义相似度和图像相似度度量方法对实验结果的影响,通过分析实验数据发现,所提出的方法能够有效实现标签丰富,提升标签质量.In order to solve the problem that user-provided tags are incomplete and inaccurate, an image tagging enrichment algorithm based on tag semantic and image visual is proposed. In this approach,co-occurrence probability and WordNet are both utilized to ex- plore latent tag correlations, for selecting candidate tags for images. Then visual features and high-level semantic are also both consid- ered when calculating image similarity, to determine the neighborhood relationship between images. Finally based on the neighbor ima- ges and the candidate tags, a new neighbor soft-voting strategy is adopted and the tags with high voting score are reserved. Experiments taken on datase MIRFlickr-25k,systematically analyzing the influence of different semantic similarity and image similarity measure- ments, experimental results demonstrate that the proposed method can solve the problem of incomplete tags effectively.

关 键 词：社会化标签 语义相似度 软近邻投票 标签丰富 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]