基于文本特征及图像实例标注的反馈策略

Relevance feedback based on textual feature labeling and image instance labeling

针对图像检索中的"语义鸿沟"问题,将主动特征学习引入相关反馈中,提出基于文本特征标注和图像实例标注的混合反馈策略。要求用户标注图像实例和文本特征,将基于图的半监督学习与基于文本特征和图像实例的双重主动学习相结合。启发式视觉特征标注算法的提出进一步提高系统性能。最佳特征标注和真实用户标注场景下的对比实验结果表明,该方法提高了系统的效率,将标注选择和检索结果返回两个过程有机统一起来。

To bridge the sematic gap of image retrieval,active feature learning was introduced to relevance feedback,and a feedback model based on textual feature and image example was proposed.Users need to label both images and textual features in this feedback model.This model was based on both graph based semi-supervised learning and dual active learning of textural feature and image example.The extension of labeling visual feature improves system’s performance.Results of comparison between the ideal labeling and real users labeling show the proposed algorithm is very effective and improves the efficiency of the system and unifies both labeling process and retrieval process.