图像检索中结合文本信息的多示例原型选择及主动学习策略

Multi-instance prototype selection and active learning combined with textual information in image retrieval

针对基于区域的图像检索系统检索精度不高的问题,提出结合文本信息的多示例原型选择算法和反馈标注机制。在示例原型选择时,首先使用文本信息进行正例拓展,然后通过估计负示例分布进行最初示例选择,最后通过示例更新和分类器学习的交替优化获得真的示例原型。相关反馈采用了多策略相结合的主动学习机制,通过信息值控制主动学习策略的自动切换,使系统能够自动选择当前最适合的主动学习策略。实验结果表明,该方法有效且性能优于其他方法。

For the poor precision of region-based image retrieval, Multi-Instance Learning （MIL） prototype selection algorithm and feedback mechanism with reference to textual information were proposed. In the process of instance prototype selection, textual information was used to extend the positive examples, and negative example distribution was used to select initial instances and the iterative optimization process of instance updating and classifier training were used to obtain the true instance prototypes. In the process of relevance feedback, active learning with the combined learning methods was adopted. The switch of active learning strategy was controlled by the information value in the feedback process. The experimental results show that this algorithm is feasible, and the performance is superior to other MIL algorithms.