Scale–Relation Joint Decoupling Network for Remote Sensing Image Semantic Segmentation

I. Overview

Objects in overhead imagery vary greatly in size, and their class can depend on relationships with surrounding regions. This network addresses both challenges by separating scale-specific information and contextual relations before learning how to combine them.

II. Key Contributions

• Decouples feature responses associated with different spatial scales.
• Models contextual relations as a complementary source of segmentation evidence.
• Jointly aggregates the scale and relation branches for dense scene labeling.

III. Methodology

The architecture extracts hierarchical image features and sends them through scale-decoupling and relation-decoupling components. Their outputs are refined and fused into a joint representation used by the segmentation decoder.

IV. Research Focus

The project studies how explicit separation can prevent large scene structures from overwhelming small objects while retaining the long-range context needed for coherent predictions.

@article{nie2022scale,
  title={Scale--relation joint decoupling network for remote sensing image semantic segmentation},
  author={Nie, Jie and Zheng, Chengyu and Wang, Chenglong and Zuo, Zijie and Lv, Xiaowei and Yu, Shusong and Wei, Zhiqiang},
  journal={IEEE Transactions on Geoscience and Remote Sensing},
  volume={60},
  pages={1--12},
  year={2022},
  publisher={IEEE}
}