Fine-grained correlation learning with stacked co-attention networks for cross-modal information retrieval