一种特征库可更新的相关反馈图像检索方法

A Relevance Feedback-Based Image Retrieval with an Updatable Feature Database

相关反馈方法是对基于内容图像检索系统的有效改进,它将人类视觉特性逐步引入检索过程,有效地减小了图像低层特征表示与图像语义理解之间的差异。但传统的相关反馈算法存在反馈次数多,且无法积累用户反馈信息等缺点。本文针对这些缺点,在相关反馈图像检索系统中引入了可更新特征库。即在原始特征索引库的基础之上引入了一个用户可修改的特征索引库,系统可以将用户多次反馈的信息逐步嵌入到这个特征索引库中。与此同时,我们提出了一种基于多分辨率分析的彩色图像纹理特征描述方法,并将其用于特征库的构建中。我们在一个含有10000幅图像的图像库上所做的测试结果表明:与Illinois大学的MARS系统相比[1],本系统可明显提高系统检索准确率和相关反馈的收敛速度。

Relevance feedback is an efficient improvement to the Content-Based Image Retrieval system. It narrows down the gap between low-level feature representation of an image and its semantic meaning, gradually introducing human vision property into the retrieval process. But there are some drawbacks in many conventional relevance feedback algorithms, such as the system needs too many feedback iterations, and cannot reuse the former user's feedback information. In this paper, we improve one conventional relevance feedback algorithm by introducing an updatable feature database into the improved algorithm. With this algorithm, a system can gradually embed the users' feedback information into the updatable database. At the same time, we propose a new texture description of a color image, based on multi-level resolution analysis, and use this description as part of our low-level feature vectors. The experiment on an image database, including 10 000 pictures, shows that the improved algorithm can greatly improve the retrieval precision and the convergent speed of feedback, compared with that in MARS of Illinois University^([1]).