Uncertainty-Guided Discriminative Priors Mining for Flexible Unsupervised Spectral Reconstruction

Area of Science:

• Computer Vision
• Image Processing
• Remote Sensing

Background:

• Supervised spectral reconstruction (SR) methods typically require paired RGB and hyperspectral images (HSIs).
• Acquiring paired data presents practical challenges, including specialized hardware and complex registration processes.
• Existing methods struggle with the high demands of paired data acquisition and processing.

Purpose of the Study:

• To propose a flexible, uncertainty-aware, unsupervised SR paradigm for recovering HSIs from RGB images.
• To overcome the limitations of paired data requirements in traditional SR methods.
• To develop a robust framework for hyperspectral image reconstruction without manual intervention.

Main Methods:

• Developed an unsupervised SR paradigm that dynamically establishes constraints using RGB images.
• Introduced the uncertainty-aware saliency alignment module (USAM) for uncertainty estimation via information entropy.
• Employed a progressively parallel network with learnable rank-guided structural representation (LRSR) and coarse-to-fine bandwise semantic perception (CBSP) flows.

Main Results:

• Demonstrated superior and robust performance on visual and remote sensing benchmarks through comprehensive quantitative and qualitative experiments.
• Successfully recovered HSIs using existing SR methods within the proposed unsupervised paradigm, showcasing its generality.
• The method adaptively explores scene-agnostic features and recovers dependable HSIs by leveraging structural and semantic priors from RGB images.

Conclusions:

• The proposed uncertainty-aware unsupervised SR paradigm effectively reconstructs HSIs without paired data.
• The framework offers a flexible and robust solution, adaptable to various SR methods.
• This approach significantly reduces the practical barriers associated with hyperspectral data acquisition and processing.