基于视觉和语义互学习的标签完备标注  

Tag Completion Annotation Based on Mutual Learning Using Visual and Semantic

作　　者：杨凯婷 张素兰[1] 张继福[1] 胡立华[1] 杨海峰[1] YANG Kai-ting;ZHANG Su-lan;ZHANG Ji-fu;HU Li-hua;YANG Hai-feng(School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China)

机构地区：[1]太原科技大学计算机科学与技术学院,太原030024

出　　处：《小型微型计算机系统》2022年第3期632-637,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61373099,U1931209)资助。

摘　　要：大多数图像标签完备方法仅依据待完备图像的视觉特征搜索近邻图像,获取候选标签实现完备,往往忽略候选标签与初始标签隐含的语义和视觉关系.为有效完备图像标签,本文依据信息的相互学习机制提出一种基于视觉和语义互学习的标签完备标注方法.首先,依据待完备图像的视觉特征和初始标签语义,搜索其近邻图像集并获取候选标签;其次,根据候选标签与初始标签的视觉内容相似性和语义相关性度量候选标签与待完备图像的关联度;最后,结合候选标签频率,选取关联程度高的候选标签实现标签补全.在基准数据集Corel5K以及真实数据集Flickr30Concepts上进行实验验证,结果表明该方法在图像标签完备的平均准确率、平均召回率以及覆盖率上均表现出更优的性能.Most image tag completion methods usually search for neighboring images by using the incomplete images′visual features,and obtain candidate labels to achieve tag completion,which often ignore the semantic and visual relationship between candidate tags and initial tags.To effectively complete image tags,this paper presents a tag completion method based on mutual learning by using visual and semantic according to the mutual learning mechanism of information.Firstly,the visual features and initial label semantics of an incomplete image are fused to search for its nearest neighbor image set and obtain candidate tags.Secondly,according to the visual content similarity and semantic relevance between the candidate tag and the initial tag,the degree of association between the candidate tag and the incomplete image is measured.Finally,the candidate tags with the high correlation degree are selected to realize tag completion through combing the frequency of the candidate tags.Experiments on the benchmark dataset Corel5k and real-world dataset Flickr30concepts show that our method has better performance on the average accuracy rate,average recall rate and coverage rate of image tag completion.

关 键 词：图像标签完备 互学习 视觉相似性 语义相关度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]