CS2DIPs: Unsupervised HSI Super-Resolution Using Coupled Spatial and Spectral DIPs.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Hyperspectral image super-resolution (HSI-SR) commonly fuses high spatial resolution multispectral images (HR-MSIs) with low spatial resolution hyperspectral images (LR-HSIs).
• Deep image prior (DIP) methods are popular for unsupervised HSI-SR due to no pre-training needs, but often struggle with latent information extraction.

Purpose of the Study:

• To develop an advanced unsupervised method for hyperspectral image super-resolution (HSI-SR) by improving the fusion of HR-MSIs and LR-HSIs.
• To enhance the performance of DIP-based HSI-SR methods by better capturing spatial and spectral information.

Main Methods:

• Proposed a coupled spatial and spectral deep image priors (CS2DIPs) method integrating nonnegative matrix-vector tensor factorization (NMVTF) into the DIP framework.
• Jointly learned the abundance tensor and spectral feature matrix using two coupled DIPs to capture spatial and spectral features from HR-MSI and LR-HSI.
• Guided HSI-SR fusion using mode-3 tensor product and incorporated physical constraints.

Main Results:

• The CS2DIPs method effectively captures rich spatial and spectral information without requiring training data.
• Demonstrated significantly superior performance and faster convergence compared to existing DIP-based HSI-SR methods.
• Extensive experiments confirmed state-of-the-art overall performance against benchmark methods.

Conclusions:

• CS2DIPs offers an effective unsupervised approach for HSI-SR, overcoming limitations of previous DIP-based techniques.
• The method shows strong potential for applications requiring high-resolution hyperspectral data fusion.
• The integration of NMVTF and coupled DIPs provides a robust framework for HSI-SR.