Deep Blind Hyperspectral Image Super-Resolution.

Area of Science:

• Remote Sensing
• Computer Vision
• Image Processing

Background:

• Hyperspectral image (HSI) super-resolution often relies on known spatial and spectral degradations.
• Real-world scenarios frequently involve unknown or complex degradation processes.

Purpose of the Study:

• To develop an unsupervised deep framework for "blind" HSI super-resolution that estimates unknown spatial and spectral degenerations.
• To improve the fusion of low spatial resolution (LR) HSI with high spatial resolution (HR) multispectral images (MSI).

Main Methods:

• A deep network models the latent HR HSI using an image-specific generator.
• Degenerations in spatial and spectral domains are modeled using a convolution layer and a fully connected layer, respectively.
• The framework is trained end-to-end using only LR HSI and HR MSI inputs via backpropagation.

Main Results:

• The proposed method demonstrates superior performance in HSI super-resolution compared to existing approaches.
• It effectively addresses unknown degenerations in spatial, spectral, or both domains.
• Experiments on natural and remote sensing datasets validate the framework's robustness.

Conclusions:

• The unsupervised deep framework successfully performs blind HSI super-resolution by estimating unknown degenerations.
• This approach offers a more practical solution for HSI super-resolution in real-world applications.
• The method significantly enhances the quality of reconstructed HR HSI data.