Cross-Scale Graph Interaction Network for Semantic Segmentation of Remote Sensing Images

I. Overview

Feature pyramids provide useful representations at several spatial resolutions, but conventional fusion may not explicitly model how structures at one scale relate to those at another. This project builds graph interactions across scales to exchange contextual information before segmentation.

II. Key Contributions

• Constructs graph representations from multi-scale visual features.
• Exchanges relational information across resolutions rather than fusing each scale independently.
• Integrates graph-refined context into a dense remote-sensing segmentation decoder.

III. Methodology

A backbone extracts features at several resolutions. Graph nodes summarize spatial or semantic regions, cross-scale interaction modules propagate information between graph levels, and the refined representations are projected back to spatial feature maps for prediction.

IV. Research Focus

The network targets classes whose recognition depends simultaneously on fine boundaries and broad geographic context.

@article{nie2023cross,
  title={Cross-scale graph interaction network for semantic segmentation of remote sensing images},
  author={Nie, Jie and Huang, Lei and Zheng, Chengyu and Lv, Xiaowei and Wang, Rui},
  journal={ACM Transactions on Multimedia Computing, Communications and Applications},
  volume={19},
  number={6},
  pages={1--18},
  year={2023},
  publisher={ACM New York, NY}
}