Cross-Modal Progressive Perspective Matching Network for Remote Sensing Image-Text Retrieval

I. Overview

Remote-sensing scenes can be described from multiple geographic perspectives. Retrieval systems that collapse those perspectives into one representation may match a query to the wrong region or overlook the relevant spatial relationship.

CPPMN progressively learns full-image perspectives, exposes perspective-specific cross-modal relationships, and then aligns image and language features semantically.

II. Key Contributions

• Uses positive text descriptions to supervise full-perspective visual feature learning.
• Transforms implicit perspective features into explicit cross-modal relationship graphs.
• Applies cascaded Transformer layers for progressive image–text semantic alignment.

III. Methodology

The network combines a compensation module for full-perspective modeling, a graph transformation module for locating individual perspectives, and a cascaded Transformer for cross-modal semantic alignment. Graph density and connectivity help identify the perspective referred to by the query.

IV. Main Findings

Quantitative and qualitative experiments across four remote-sensing image–text retrieval datasets demonstrate the value of progressive perspective matching and semantic alignment.

@article{zheng2025cross,
  title={Cross-modal progressive perspective matching network for remote sensing image-text retrieval},
  author={Zheng, Chengyu and Li, Xiu and Liang, Xinyue and Huang, Lei and Du, Shan and Nie, Jie and Dong, Junyu},
  journal={IEEE Transactions on Multimedia},
  volume={27},
  pages={3966--3978},
  year={2025},
  publisher={IEEE}
}