Multi-Level Attention Network for Cross-Modal Hashing Retrieval

Nowadays, cross-modal hash retrieval has been widely and successfully used in multi-modal similarity search applications. However, to achieve highly discriminative representations of different modalities in a common Hamming space, it is crucial to fully explore and preserve the information provided by the training data. We proposed a multi-level attention and weighted similarity cross-modal hashing retrieval method (MAWSH). MAWSH incorporates self-attention and cross-attention mechanisms into the representation learning process to improve intra-modal and inter-modal semantic correlation. We also employ label weighting to better use label information to obtain clearer semantic connections between data when learning. Our extensive experiments on three widely used cross-modal datasets demonstrate that our MAWSH method achieves remarkable accuracy improvements for image-to-text and text-to-image retrieval tasks, outperforming five state-of-the-art approaches.