图像检索中一种有效的半监督主动相关反馈方案

摘要

Active learning plays an important role for boosting interactive image retrieval Among various methods,support vector machine(SVM) based active learning approaches have been drawn substantial attention. However, most SVM-based active learning methods are challenged by small example problem,asymmetric distribution problem,and redundancy among examples. This paper proposed two mechanisms to tackle above problems: (1) designing an asymmetric semi-supervised learning(ASL) framework that exploits unlabeled data for semantic relevant and irrelevant classes in different ways. Under the influence of ASL, the efficiency of SVM is significantly improved; and(2) developing a representative measure based active selection criterion to identify the most informative images from unlabeled data while the diversity among them is augmented. Experimental results validate the superiority of our scheme over several existing methods.%在交互式图像检索中,基于支持向量机(Support Vectot Machilies,SVM)理论的主动反馈技术扮演着重要角色.然而,现有的SVM主动反馈方法普遍受到小样本问题、不对称分布问题以及样本冗余性等问题的制约.提出两种新颖策略以应对上述问题:(1)针对相关反馈的技术特点,提出了非对称半监督学习框架,该框架采用不同的学习方法为语义相关类和无关类挑选未标记图像,以有效增强SVM的泛化能力;(2)设计了基于代表性度量的主动采样方法,该方法不仅能够从未标记数据中鉴别出富有信息(most informative)图像,而且确保了待标记图像之间具有较大的差异性.实验结果及对比分析表明,所提方案明显优于其它同类算法.