Learning a Limited Text Space for Cross-Media Retrieval

摘要

In this paper, we propose a novel model for cross-media retrieval which relies on a limited text space rather than a common space or an image space. More specifically, the model consists of three parts: A visual part that consists of a convolutional neural network and an image understanding network; A language model part that achieves sentence understanding by recurrent neural network; An embedding part that contains a fusion layer to capture both visual label information and semantic correlations between images and sentences, as well as learn the final limited text space by optimizing pairwise ranking loss. Experimental results on three benchmark datasets show that our proposed model gains promising improvement in accuracy for cross-media retrieval especially on sentence retrieval compared with the related state-of-the-art methods.