Deep Neural Architecture for Multi-Modal Retrieval based on Joint Embedding Space for Text and Images

Recent advances in deep learning and distributed representations of images and text have resulted in the emergence of several neural architectures for cross-modal retrieval tasks, such as searching collections of images in response to textual queries and assigning textual descriptions to images. However, the multi-modal retrieval scenario, when a query can be either a text or an image and the goal is to retrieve both a textual fragment and an image, which should be considered as an atomic unit, has been significantly less studied. In this paper, we propose a gated neural architecture to project image and keyword queries as well as multi-modal retrieval units into the same low-dimensional embedding space and perform semantic matching in this space. The proposed architecture is trained to minimize structured hinge loss and can be applied to both cross- and multi-modal retrieval. Experimental results for six different cross- and multi-modal retrieval tasks obtained on publicly available datasets indicate superior retrieval accuracy of the proposed architecture in comparison to the state-of-art baselines.