基于主动相关反馈的图像检索方法  

An Image Retrieval Method Based on Active Relevance Feedback

作　　者：陈可佳[1] 姜远[1] 周志华[1] 

机构地区：[1]南京大学计算机软件新技术国家重点实验室,南京210093

出　　处：《模式识别与人工智能》2005年第4期480-485,共6页Pattern Recognition and Artificial Intelligence

基　　金：973计划(No.2002CB312002);国家杰出青年科学基金(No.60325207);全国优秀博士学位论文作者专项基金(No.200343)

摘　　要：在基于内容的图像检索中,如何找到有效方法以缩小图像的视觉描述和用户语义解释之间的差异是一个关键问题。本文在图像检索的相关反馈过程中引入主动学习的思想,提出了主动相关反馈方法,通过将用户的注意力集中在检索系统分类确定性低的图像上,使得用户反馈针对最能有效提高检索性能的图像进行。实验结果表明,该方法有助于缩小用户查询中的歧义性,有效提高检索效率。In content-based image retrieval, how to seek out an effective method which can bridge the gap between the visual feature description and the user semantic interpretation is a fundamental problem. In this paper, a novel method inspired by active learning, especially selective sampling, is introduced into the relevance feedback process, which results in the active relevance feedback method. By this method, the system actively selects the most informative images for the user so that the uncertainty of the system could be decreased to the least with the same number of feedbacks. Experiments show that this method can help to improve the effectiveness of retrieval through the disambiguation of the user query.

关 键 词：主动学习 选择性取样 相关反馈 基于内容的图像检索 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术]