Hyperspectral image super-resolution via a multi-stage scheme without employing spatial degradation

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Hyperspectral image (HSI) super-resolution (SR) commonly fuses low spatial resolution HSI with high spatial resolution RGB images.
• Existing methods rely on known spatial degradation models, which are often inaccurate in practice, limiting performance.

Purpose of the Study:

• To develop a novel HSI super-resolution method that does not require a spatial degradation model.
• To improve the accuracy and efficiency of hyperspectral image reconstruction.

Main Methods:

• A multi-stage scheme involving initialization, modification, and refinement.
• Initialization uses angle similarity between HR-RGB pixels and LR-HSI spectra.
• Modification employs a polynomial function to match spectral RGB values to HR-RGB.
• Refinement utilizes a spectral-spatial total variation (SSTV) regularizer within an optimization model.

Main Results:

• The proposed method successfully reconstructs high spatial resolution hyperspectral images (HR-HSI) without spatial degradation models.
• Experimental results demonstrate superior performance over eight state-of-the-art methods.
• The method shows improvements in both reconstruction accuracy and computational efficiency.

Conclusions:

• The novel multi-stage scheme effectively addresses the limitations of existing HSI SR methods.
• The proposed SSTV regularizer is crucial for preserving spectral and spatial integrity.
• This approach offers a more robust and efficient solution for HSI super-resolution.