Hypergraph-Enhanced Textual-Visual Matching Network for Cross-Modal Remote Sensing Image Retrieval via Dynamic Hypergraph Learning

Cross-modal remote sensing (RS) image retrieval aims to retrieve RS images using other modalities (e.g., text) and vice versa. The relationship between objects in the is complex, i.e., distribution of multiple types uneven, which makes matching with query text inaccurate, then restricts performance retrieval. Previous methods generally focus on feature rarely model relationships features image. Hypergraph (hyperedge connecting vertices) an extended structure a regular graph has attracted extensive attention for its superiority representing high-order relationships. Inspired by advantages hypergraph, this work, hypergraph-enhanced textual-visual network (HyperMatch) proposed circumvent inaccurate text. Specifically, multiscale hypergraph designed complex forming valuable redundant into different hyperedges. In addition, construction update method designed. For constructing running as vertices cosine similarity metric measure correlation them. Vertex hyperedge mechanisms are introduced dynamic realize alternating Quantitative qualitative experiments RSICD RSITMD datasets verify effectiveness cross-modal