H-MoC: Hierarchical Mixture-of-Experts for Semantic- and Dispersion-Consistency in Remote Sensing Image-Text Retrieval

I. Overview

Remote-sensing image–text retrieval requires a shared representation that can preserve semantic agreement while accommodating substantial variation within each modality. H-MoC organizes specialized experts hierarchically so image and language features can be routed according to both their semantic structure and their dispersion.

II. Key Contributions

• Introduces semantic-aware routing that groups features around learned semantic centroids.
• Adds dispersion-aware routing to account for differences in feature variance and distribution.
• Combines shared and specialized experts to balance common cross-modal knowledge with group-specific patterns.

III. Methodology

Image and text encoders first produce modality-specific representations. The semantic-consistency branch assigns them to groups using learned centroids, while the dispersion-consistency branch models how broadly each group is distributed. Shared and specialized experts then refine the routed features before cross-modal alignment. The supplied architecture also includes alignment, grouping, and load-balancing objectives.

IV. Research Focus

The framework focuses on improving retrieval when visually related scenes and semantically overlapping descriptions exhibit different levels of intra-class variation. Its hierarchical routing is designed to keep related concepts aligned without forcing every sample through the same feature transformation.

@ARTICLE{11673296,
  author={Zheng, Chengyu and Lu, Hanzhang and Nie, Jie and Du, Shan},
  journal={IEEE Transactions on Multimedia}, 
  title={H-MoC: Hierarchical Mixture-of-Experts for Semantic- and Dispersion-Consistency in Remote Sensing Image-Text Retrieval}, 
  year={2026},
  volume={},
  number={},
  pages={1-11},
  doi={10.1109/TMM.2026.3729400}}