Beyond pairwise alignment: Hypergraph-based semantic calibration for image-text matching

Image-text matching is a fundamental task in multimodal retrieval, aiming to establish reliable semantic correspondences between visual regions and textual tokens. Existing methods assume consistent cross-modal semantic structures and rely on pairwise alignment, which often leads to structural misalignment and limits the modeling of high-order semantic compositions. To address these challenges, we propose a H ypergraph-based S emantic C alibration M odel (HSCM), which introduces a shared semantic reference for intra-modal calibration, enabling structurally consistent representations in a unified semantic space. Furthermore, a semantically conditioned hypergraph is constructed, treating semantic units as nodes and hyperedges as high-order relationships, extending conventional point-to-point alignment to structure-level modeling and enhancing the stability and discriminative capability of cross-modal representations. Experimental results on Flickr30K (31,000 images) and MS-COCO (123,287 images) demonstrate that HSCM outperforms state-of-the-art approaches, achieving Recall@sum improvements of 2.4%, 5.9%, and 0.8% on Flickr30K, MS-COCO 1 K, and MS-COCO 5 K, respectively.