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 Bayesianlearning,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.
The typical model, which involves the measures： support, confidence, and interest, is often adapted to mining association rules. In the model, the related parameters are usually chosen by experience; consequently, the number of useful rules is hard to estimate. If the number is too large, we cannot effectively extract the meaningful rules. This paper analyzes the meanings of the parameters and designs a variety of equations between the number of rules and the parameters by using regression method. Finally, we experimentally obtain a preferable regression equation. This paper uses multiple correlation coeficients to test the fitting efiects of the equations and uses significance test to verify whether the coeficients of parameters are significantly zero or not. The regression equation that has a larger multiple correlation coeficient will be chosen as the optimally fitted equation. With the selected optimal equation, we can predict the number of rules under the given parameters and further optimize the choice of the three parameters and determine their ranges of values.
