Optimizing Cross-Modal Image–Text Matching Algorithms Based on Graph Neural Networks: Implications for Textile Applications

The challenge of image–text matching stems from differences in representation and structural semantics, particularly in complex scenarios such as digital textile and apparel analysis, resulting in a semantic gap that degrades matching performance. This paper proposes an optimized graph neural network (GNN)-based algorithm to improve cross-modal matching accuracy. Images and text are first represented as graph structures, where image regions and textual keywords are modeled as nodes and their relationships are encoded as edges. The GNN enables progressive information propagation and feature fusion across modalities, while an adaptive semantic alignment module dynamically adjusts propagation weights according to node-level semantic similarity. A multi-scale GNN architecture is further introduced to capture hierarchical semantic features and is optimized using a cross-modal contrastive loss that maximizes the similarity of correctly matched image–text pairs. Such structured multimodal semantic modeling also provides methodological insights for intelligent information fusion and heterogeneous data interpretation in advanced electromagnetic sensing and communication systems. Experimental results demonstrate Recall@1 scores of 70.1% and 60.5% on the MSCOCO and Flickr30k datasets, respectively, with a mean rank of 3.1 and mutual information of 0.75. The proposed method effectively narrows the semantic gap and significantly enhances cross-modal matching accuracy, demonstrating substantial potential for intelligent retrieval and analysis in the textile industry.