H-MoC: Hierarchical Mixture-of-Experts for Semantic- and Dispersion-Consistency in Remote Sensing Image-Text Retrieval
Year: 2026
Type: article
Abstract: Recent advancements in remote sensing (RS) image text retrieval have focused on addressing the cross-modal gap, improving saliency modeling, and increasing architectural efficiency. Despite these efforts, RS retrieval performance remains inferior to that of general-domain benchmarks. This disparity is primarily due to the inherent imbalance in intra-class semantic diversity within RS imagery. To address this challenge, we propose H-MoC, a Hierarchical Mixture-of-Experts for semantic- and dispersion Consistency in remote sensing image-text retrieval. Specifically, we first perform clustering and statistical variance analysis on image and text tokens to derive semantic partitions and intrinsic dispersions. Leveraging these, a hierarchical routing mechanism is introduced: the semantic-aware router assigns tokens to initial pathways based on class-specific centroids, while the dispersion-aware router dynamically routes tokens to experts according to intra-class variance. By organizing tokens at the class level and subsequently performing fine-grained, variance guided expert allocation, this approach enables both efficient and strategic token routing, which significantly enhances the model's capacity for category-discriminative processing. To ad dress the cross-modal gap, shared learnable semantic/dispersion queries are introduced for images and texts to further optimize the routers. Comprehensive experiments on three benchmark datasets demonstrate that H-MoC effectively mitigates the impact of diversity imbalance in RS image-text retrieval.
Source: IEEE Transactions on Multimedia
Authors: Chengyu Zheng, Hanzhang Lu, Jie Nie, Shan Du
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7204802105
