基于最大间隔的半监督图像搜索重排序方法

A Max Margin Based Semi-Supervised Reranking Method

提出一种基于最大间隔原理的半监督图像搜索重排序学习算法。所提算法在最大间隔原理框架下,首先利用超图正则化保持标注及未标注样本在原始空间中的局部近邻关系,增强算法的稳健性;其次,利用少量的标注样本构造优先关系对,将样本间先验的相关性等级信息引入目标函数中以更好地指导重排序模型的学习。在公开数据集MSRA-MM1.0上的实验结果表明所提方法能更好地将符合用户需求的结果靠前优先呈现给用户,提高搜索的准确性。

We propose a max margin based semi-supervised reranking method for multimedia information retrieval. We use hypergraph regularization to preserve the neighborhood of the sample in the original space and introduce the labeled and unlabeled sample information to construct the objective function, so as to achieve full and efficient use of data information for ranking. By using a small amount of annotation samples to construct the priority relationship pairs, the priority information between samples is introduced into the objective function to construct a ranking learning model. This method can show users in priority the results that meet their demand better, and improve the retrieval accuracy. The experimental results on MSRA-MM 1. 0 dataset suggest the proposed method provides superior performance compared with several state-of-the-art methods.