Graph Convolutional Network Semantic Enhancement Hashing for Self-supervised Cross-Modal Retrieval

Cross-modal hashing has gained widespread attention for its computational efficiency and reduced storage costs while achieving great success in cross-modal retrieval. However, the data information of different modalities is asymmetric, which means that we usually think of images as being much more informative than text. Currently, most previous methods have no obvious difference in feature extraction and hash function learning modules, and they do not address the semantic gap between different modalities. We propose Graph Convolutional Network Semantic Enhancement Hashing (GCSEH) approach. Specifically, We aim to let information-rich modalities further support information-poor modalities to bridge the semantic gap between different modalities. In addition, in order to more accurately capture the semantic affinity between modalities, in order to be able to discover high-level semantic information, we choose to use a self-supervised semantic network. Extensive experiments have proven that our GCSEH method can achieve excellent performance.