中文摘要：

基于贝叶斯（Bayesian）理论的相关反馈技术是可有效提高图像检索性能的重要手段之一。然而，当前大多数的Bayesian反馈算法普遍受到小样本问题和训练样本不对称问题的制约。本文提出一种新的相关反馈算法，该算法将查询点移动（query point movement，QPM）技术嵌入Bayesian框架中，并采用不对称的学习策略处理正、负反馈信息，故而称之为不对称Bayesian学习（asymmetry Bayesian learning，ABL）．对于正例样本，该算法同时考虑用户提供的正、负反馈信息，并借助QPM技术估计相关语义类图像的概率分布．对于负例样本，采用一种半监督学习机制以应对负例样本稀缺问题．首先，通过随机采样从数据库中选取一组无标记图像，然后，利用QPM技术对其进行数据审计．最后，将审计后的无标记图像作为额外的负例样本，并与用户标记的负反馈信息一起用于估计不相关语义类图像的概率分布．仿真实验及对比结果表明，不对称Bayesian学习策略可显著提高相关反馈的效率，且本文算法的检索性能明显优于当前其它的相关反馈算法．

英文摘要：

With the explosive growth in digital image records and the rapid increase of computer power, content based image retrieval （CBIR） has become a very active research field during the past years. One of the fundamental problems in CBIR is the gap between the low-level visual features and the high-level semantic concepts. Relevance feedback （RF）, as a powerful tool for bridging the gap, is introduced into CBIR. Among many approaches, Bayesian learning （BL） plays a key role for boosting RF performance. However, most BL- based RF methods are challenged hy the small size example collection and the asymmetric distributions between the positive and the negative examples. To overcome these drawbacks indwelled in current BL-based RF methods, this paper presents a novel scheme that embeds the query point movement （QPM） technique into the Bayesian framework for improving RF performance. In particular, we use an asymmetric learning methodology to determine the parameters of the Bayesianlearner, which is termed as asymmetric Bayesian learning （ABL）. Concretely, for the positive examples, QPM is applied to estimate the distribution of the relevant semantic class by exploiting the positive feedbacks in conjunction with the negative ones; for the negative ones, a semi-supervised learning mechanism is used to tackle the scarcity of negative examples. In detail, a random subset of the unlabeled images is firstly selected as candidate negative examples, and then a QPM-based data editing method, proposed in this paper, is used to eliminate the problematic data mixed in the selected examples. Finally, the edited unlabeled images are regarded as additional negative examples which are helpful to estimate the distribution of the irrelevant class. Experimental results show that using asymmetric Bayesian learning strategy in RF is beneficial, and the proposed method achieves better performance than other existing approaches.