摘要
提出了一种基于兴趣点的图像检索新方法.在尺度空间中检测兴趣点,依据兴趣点的分布将图像划分成一系列等面积的扇形子区域并提取图像特征.该特征既反映了兴趣点的局部特性,又考虑了兴趣点的空间分布结构,同时对图像旋转、缩放和平移具有不变性.在相关反馈阶段,将图像看作是由各子区域内兴趣点局部特征构成的多示例包,根据用户选择的实例图像生成正包和反包,采用多示例学习算法获得体现图像语义的目标概念.本方法缩小了用户查询中的歧义性,在Corel图像库中进行的实验表明,与其他基于兴趣点的图像检索方法相比,平均检索准确率提高7%以上,可以更准确地查找到用户所需图像.
A novel method for image retrieval based on interest points is presented.The interest points are detected in the scale space.Then the image is divided into fan-shaped sub-regions of equal area according to the distribution of the interest points.Local features representing the spatial distribution information on the interest points are extracted to describe the image,and they are also robust to the image＇s rotation,scale and translation.In the relevant feedback,images are regarded as multiple-instance bags consisting of the local domain of the interest points in every fan-shaped sub-region.Labeled images chosen by the user are generated corresponding positive and negative bags,and the multiple-instance learning algorithm is employed to obtain the target concept reflecting the query image semantics.The method can reduce the ambiguity of the user query.Experimental results based on the Core image database show that our method improves the average retrieval precision by 7 percent or more,compared with other interest points based retrieval methods.
出处
《西安电子科技大学学报》
EI
CAS
CSCD
北大核心
2011年第2期47-53,共7页
Journal of Xidian University
基金
国家自然科学基金资助项目(61003196)
中央高校基本科研业务费专项资金资助项目(K50510040004)
关键词
图像检索
兴趣点
特征提取
局部分布特征
多示例学习
image retrieval
interest points
feature extraction
local distribution features
multiple-instance learning
