Answers are grounded in this paper and cite their sources.
To address the limitations of existing optical remote sensing image captioning (RSIC) methods in multiscale feature representation, spatial relationship modeling, and cross-modal semantic alignment, a RSIC method via hierarchical visual representation fusion and multigrained cross-modal alignment is proposed. A Swin transformer with a shifted window mechanism is employed to extract multiscale hierarchical visual features, thereby enhancing the representation of local details, global semantics, and potential relationships among objects. A hierarchical dual-attention collaborative feature fusion module is constructed to achieve effective integration of shallow texture features and deep semantic features through multibranch feature alignment as well as joint spatial and channel attention modeling. In the decoding stage, a multigrained semantic-guided visual attention mechanism is introduced to guide visual feature selection collaboratively with global textual semantics and fine-grained word-level semantics, and a two-layer long short-term memory network is used to generate semantically consistent and coherent descriptive sequences. Experimental results demonstrate that the proposed method achieves competitive performance on three public optical remote sensing datasets, namely Sydney, UCM, and NWPU, verifying its effectiveness in scene semantic understanding and caption generation for optical remote sensing images.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: