Related Experiment Videos
MambaX: Image Super-Resolution with State Predictive Control
IEEE Transactions on Pattern Analysis and Machine Intelligence
|August 10, 2026
Summary
We introduce MambaX, a novel nonlinear state predictive control model for image super-resolution (SR). MambaX enhances SR by dynamically learning spectral-state representations, improving fine-grained image reconstruction and multimodal fusion.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Signal Processing
Background:
- Image super-resolution (SR) is crucial for overcoming sensor hardware limits.
- Existing SR methods often fail to control error propagation in intermediate stages.
- Mamba offers sequence-based reconstruction but has limitations in flexibility and receptive field for fine-grained images.
Purpose of the Study:
- To develop a more effective SR approach addressing limitations of current methods.
- To introduce MambaX, a nonlinear state predictive control model for enhanced SR.
- To generalize SR tasks by dynamically learning nonlinear state parameters.
Main Methods:
- Developed MambaX, a nonlinear state predictive control model.
- Mapped consecutive spectral bands into a latent state space for SR.
- Employed dynamic state predictive control learning for state-space models.
- Introduced state cross-control for multimodal SR fusion.
- Utilized progressive transitional learning to handle domain and modality shifts.
Main Results:
- MambaX demonstrated superior performance in single-image SR.
- MambaX achieved superior performance in multimodal fusion-based SR.
- The model effectively mitigates heterogeneity from domain and modality shifts.
Conclusions:
- MambaX significantly advances spectrally generalized modeling across dimensions and modalities.
- The dynamic spectrum-state representation model shows substantial potential for future SR research.
- MambaX offers a flexible and effective solution for fine-grained image SR and multimodal fusion.