基于多示例学习的遥感图像检索

Remote sensing image retrieval based on multiple instance learning

将基于多示例学习的相关反馈技术应用到遥感图像检索中.为了避免局部最小值和减少冗余计算量,对传统的多示例学习算法——多样密度算法进行了改进.改进的算法利用用户标注的样本学习得到的查询概念,指导下一轮检索.为了提高查准率,综合学习得到的查询概念,提出了查询概念集来取代查询概念.实验结果表明,该算法比传统方法具有更好的检索性能.

Relevance feedback （RF） based on multiple instance learning （MIL） was applied to the remote sensing image retrieval. In order to avoid the local minimum and reduce redundant computation, the traditional MIL algorithm-diverse density （DD） was improved. The improved DD algorithm learnt the query concept by employing the samples labeled by users, thereby facilitating the following query. In order to enhance the precision through sythesizing the query concepts learnt before the query concept set was proposed to substitute the query concepts. The experimental results show that this algorithm yields better performance than traditional approaches.