Multi-label Supervised Manifold Ranking for Multi-instance Image Retrieval

Current manifold ranking is mainly used in single-instance image retrieval without considering the prevailing semantic ambiguity problem. This paper introduces multi-instance technique and supervised information to image retrieval based on manifold ranking, and proposes a Multi-label Supervised Manifold Ranking algorithm (MSMR) for multi-instance image retrieval. The divergence between images is modified by using the multi-label information of training samples. Our method can solve partly the 'input ambiguity problem' in the feature extraction stage and the 'output ambiguity problem' in the output stage. Compared with the traditional Expectation Maximization Diverse Density (EMDD) and Citation-kNN algorithm on Corel Image Set, the multi-instance image retrieval experimental results show that the average precision rate of our algorithm has be enhanced