Learning Semantic Structure Correlation with Deep Semantic Alignment Network for Cross-Modal Retrieval

. Cross-modal retrieval is an important research in the field of multi-modal learning, which aims to map samples from different modalities into a shared feature space for comparing and retrieving samples. However, existing methods are unable to capture the potential semantic correlations and coexistence information of different modal data, resulting in lack of effective features. To address this issue, we propose a novel approach called Learning Semantic Structure Correlation with Deep Semantic Alignment Network for Cross- Modal Retrieval, which consists of multiple sub-networks that employ deep neural networks combined with a self-attentive mechanism to enhance the semantic information of the data and explore the underlying neighborhood structure in depth through graph convolutional networks. We develop an efficient aggregation strategy to construct more reliable supervised signals using multiple metrics to reduce bias. Furthermore, three contrastive learning are designed to share the same semantic information between different modalities. Extensive experiments on two widely used benchmark datasets demonstrate that our method has better performance with state-of-the-art cross-modal methods.