摘要
如何自动判断社会化标签与图像内容之间的相关性是社会化多媒体内容检索领域一个重要的研究问题.近邻投票算法是已知的计算标签相关性的最有效方法之一.但该算法采用硬投票策略,并未考虑近邻图像的权重以及近邻图像自身标签的质量.针对上述问题,文中提出一种一般性的软近邻投票框架,通过考察近邻权重和近邻标签权重这两个维度,系统性地比较了23种软近邻投票实现方案.以近120万张Flickr图像为训练集、约两万张图像为测试集的实验表明,软近邻投票策略要优于硬投票策略:平均查准率从0.764提升到0.783,且软近邻投票对于近邻个数这一重要参数的选取相对不敏感.
How to automatically determine the relevance of a social tag with respect to image content is crucial for social multimedia retrieval.The neighbor voting algorithm is known to be effective for tag relevance estimation.However,it uses a hard voting strategy,which does not take into account the importance of neighbor images,nor the quality of the tags associated with the neighbors.As an improvement,we propose in this paper a generic soft neighbor voting framework.By exploiting multiple methods for computing the importance of the neighbors and the importance of the tags of the neighbors,we systematically compare up to 23 distinct soft neighbor voting solutions.We conduct experiments using 1.2 million Flickr images as training data and another set of 20k images as test data.The results show that the proposed soft neighbor voting is better than the standard neighbor voting algorithm,improving mean Average Precision from 0.764 to 0.783.Soft neighbor voting is also found to be less sensitive with respect to the number of neighbors for voting.
出处
《计算机学报》
EI
CSCD
北大核心
2014年第6期1365-1371,共7页
Chinese Journal of Computers
基金
中国人民大学科学研究基金(中央高校基本科研业务费专项资金)(13XNLF05
14XNLQ01)
国家自然科学基金(61303184
61003205)
教育部高等学校博士点专项科研基金(20130004120006)资助~~
