Learning a hybrid similarity measure for image retrieval

Learning similarity measure from relevance feedback has become a promising way to enhance the image retrieval performance. Existing approaches mainly focus on taking short-term learning experience to identify a visual similarity measure within a single query session, or applying long-term learning methodology to infer a semantic similarity measure crossing multiple query sessions. However, there is still a big room to elevate the retrieval effectiveness, because little is known in taking the relationship between visual similarity and semantic similarity into account. In this paper, we propose a novel hybrid similarity learning scheme to preserve both visual and semantic resemblance by integrating short-term with long-term learning processes. Concretely, the proposed scheme first learns a semantic similarity from the usersʹ query log, and then, taking this as prior knowledge, learns a visual similarity from a mixture of labeled and unlabeled images. In particular, unlabeled images are exploited for the relevant and irrelevant classes differently and the visual similarity is learned incrementally. Finally, a hybrid similarity measure is produced by fusing the visual and semantic similarities in a nonlinear way for image ranking. An empirical study shows that using hybrid similarity measure for image retrieval is beneficial, and the proposed algorithm achieves better performance than some existing approaches.