Unsupervised Concept Learning in Text Subspace for Cross-Media Retrieval

Subspace (i.e. image, text or latent subspace) learning is one of the essential parts in cross-media retrieval. And most of the existing methods deal with mapping different modalities to the latent subspace pre-defined by category labels. However, the labels need a lot of manual annotation, and the label concerned subspace may not be exact enough to represent the semantic information. In this paper, we propose a novel unsupervised concept learning approach in text subspace for cross-media retrieval, which can map images and texts to a conceptual text subspace via the neural networks trained by self-learned concept labels, therefore the well-established text subspace is more reasonable and practicable than pre-defined latent subspace. Experiments demonstrate that our proposed method not only outperforms the state-of-the-art unsupervised methods but achieves better performance than several supervised methods on two benchmark datasets.