Scale–Semantic Joint Decoupling Network for Image-Text Retrieval in Remote Sensing

I. Overview

The visual evidence corresponding to a caption may appear at different scales across remote-sensing images. This project separates scale-sensitive visual structure from shared semantic content and then learns how the two should interact during retrieval.

II. Key Contributions

• Decouples multi-scale visual information to preserve objects at different resolutions.
• Separates semantic components used for image–text alignment.
• Jointly optimizes the scale and semantic representations for bidirectional retrieval.

III. Methodology

Image and text encoders generate modality-specific features. Scale-decoupling modules refine the visual hierarchy, semantic-decoupling modules identify cross-modal concepts, and a joint matching stage combines them in the retrieval embedding space.

IV. Research Focus

The method addresses retrieval settings where global scene appearance is similar across candidates but the caption refers to a structure visible at a particular scale.

@article{zheng2023scale,
  title={Scale-semantic joint decoupling network for image-text retrieval in remote sensing},
  author={Zheng, Chengyu and Song, Ning and Zhang, Ruoyu and Huang, Lei and Wei, Zhiqiang and Nie, Jie},
  journal={ACM Transactions on Multimedia Computing, Communications and Applications},
  volume={20},
  number={1},
  pages={1--20},
  year={2023},
  publisher={ACM New York, NY}
}