一种偏好于查询子流形的半监督图像检索算法  

Query Sub-manifold Biased Semi-supervised Algorithm for Image Retrieval

作　　者：胡恩良[1,2] 赵富坤[2] 尹学松[1] 

机构地区：[1]南京航空航天大学信息科学与技术学院,江苏南京210016 [2]云南师范大学数学学院,云南昆明650092

出　　处：《小型微型计算机系统》2010年第2期363-368,共6页Journal of Chinese Computer Systems

摘　　要：在基于反馈的图像检索中,由于被用户标记为相关和不相关的图像数较少,使得检索问题变成了一个典型的小样本问题.流形可表达数据在低维空间中的内在几何结构,流形正则化的目的是利用这种几何结构来约束解空间,以使最优解能反映数据本身的几何分布.为了解决反馈检索中的小样本问题,本文在流形正则化框架下提出一个新的半监督图像检索算法.在新算法中,流形正则化项只依赖于文中定义的查询子流形,而不依赖于数据集的全局结构.在两个图像集上的实验结果对比表明,本文提出的新算法在检索效果上优于现有的4种state-of-the-art算法.Under the circumstance of the relevance-feedback-based image retrieval,the quantity of relevant and irrelevant images labeled by user is very few leading to a canonical small-size sample problem.The intrinsic structure of such image data usually appears in the manifold form in lower-dimension space,thus the goal of the manifold-regularization is to utilize the structure to make the optimal solution able to reflect the distribution geometry of the data.With the aim to solving such small-size sample problem,a new semi-supervised image retrieval algorithm is proposed in this paper,which based on the manifold-regularization framework.In the proposed algorithm,the term of manifold-regularization only depends on the query submanifold rather than the global structure of the whole data set.The experimental results on two image databases indicate that such new retrieval algorithm is more effective than four current state-of-the-art algorithms.

关 键 词：图像检索 相关反馈 半监督学习 支持向量机 流形正则化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]