Multi-source unsupervised domain adaptation method for remote sensing image segmentation based on multi-expert gating fusion and divergence constraint
By combining a frequency-domain decoupled dynamic concentric Mamba encoder with a multi-source domain expert adapter, the problems of feature interference and pseudo-label noise in multi-source domain adaptive remote sensing image segmentation are solved, achieving higher accuracy and robustness in cross-domain remote sensing image segmentation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHIJIAZHUANG TIEDAO UNIV
- Filing Date
- 2026-05-29
- Publication Date
- 2026-07-17
AI Technical Summary
Existing multi-source domain adaptive remote sensing image segmentation methods suffer from problems such as insufficient fusion of multi-source knowledge, mutual interference of source domain features, easy accumulation of false label noise in the target domain, and insufficient semantic consistency of similar types across domains under cross-domain conditions, resulting in insufficient segmentation accuracy and generalization ability.
A multi-source unsupervised domain adaptive remote sensing image segmentation method based on multi-expert gating fusion and divergence constraints is adopted. By decoupling the frequency domain dynamic concentric Mamba encoder, multi-source domain expert adapter and gating fusion module, combined with multi-expert divergence constraints and category prototype memory alignment mechanism, feature decoupling, multi-source domain expert modeling, target domain dynamic fusion and pseudo-label noise suppression are achieved.
It improves the segmentation accuracy and robustness of the model under complex multi-source cross-domain conditions, reduces the impact of multi-source feature interference and pseudo-label noise, and enhances the semantic consistency and segmentation effect of the target domain.
Smart Images
Figure CN122415650A_ABST