Hierarchical Feature-Modulated Transformer With Introspective Interactions for Visual Grounding in Remote Sensing Images

Visual grounding in remote sensing images aims to localize objects referred to by a query text in complex aerial scenes. However, existing methods still suffer from limitations in multimodal interaction and alignment. Specifically, multi-scale visual feature fusion often lacks effective textual semantic filtering, introducing redundant background information that degrades cross-modal alignment. Moreover, although bidirectional cross-attention enhances inter-modal information exchange, it tends to neglect intra-modal semantic consistency, resulting in unstable and less discriminative representations. In addition, language queries obtained after cross-modal interaction are vulnerable to linguistic noise during decoding, which further impairs localization robustness. To address these challenges in a unified and progressive manner, we propose a hierarchical vision–language alignment framework, termed Hierarchical Feature-Modulated Transformer with Introspective Interaction (HMIT). Specifically, TMCFF performs early-stage text-guided modulation on multi-scale visual features to suppress text-irrelevant background information. CMICI refines cross-modal representations through contextual introspection while preserving intra-modal semantic consistency. VGTC calibrates high-level language queries using global visual priors prior to decoding, producing more robust and discriminative query embeddings. Extensive experiments on three benchmark datasets demonstrate that HMIT consistently outperforms existing methods across multiple metrics, achieving new state-of-the-art performance.