Multi-label Double Layers Learning for Cross-Modal Retrieval