Efficient manifold ranking for cross-media retrieval

With the advent of the era of big data, all kinds of data are fused together which form a new modality-multimedia data, how to find information in multimedia data become a key problem, so cross-media retrieval has drawn great attention in recent year. Precisely and effectively measuring the similarity between different modalities of data is a key step of cross-media retrieval. The traditional methods usually use Cosine distance or Euclidean distance, which is difficult to reflect the real similarity between different modalities of data. For address this problem, in this paper, we use a novel approach to replace the above two methods, called efficient manifold ranking (EMR), which is able to better reflect the similarity between different modalities of data. Experiments on Wikipedia dataset and the challenging XMedia dataset which includes 5 media types show the effectiveness of the EMR algorithm, as compared with the 4 state-of the-art methods.