Unmixing Guided Unsupervised Network for RGB Spectral Super-Resolution

Area of Science:

• Computer Vision
• Image Processing
• Remote Sensing

Background:

• Spectral super-resolution aims to generate high-resolution hyperspectral images from low-resolution RGB images.
• Existing methods often rely on supervised learning, necessitating difficult-to-obtain paired training data.

Purpose of the Study:

• To develop an unsupervised spectral super-resolution framework that eliminates the need for paired hyperspectral-RGB training data.
• To leverage arbitrary hyperspectral imagery for guidance in reconstructing spectral information.

Main Methods:

• Proposes the Unmixing Guided Unsupervised Network (UnGUN) with two unmixing branches and a reconstruction branch.
• Utilizes spectral unmixing to extract spectral and spatial priors from guidance and RGB images.
• Employs a discriminator to ensure the generated image distribution matches real hyperspectral data.

Main Results:

• Successfully achieves unsupervised spectral super-resolution without paired data.
• Demonstrates superior performance compared to state-of-the-art (SOTA) methods in experimental evaluations.
• Reconstructed images exhibit characteristics of real hyperspectral imagery.

Conclusions:

• UnGUN provides a viable unsupervised approach for spectral super-resolution.
• The method effectively utilizes spectral unmixing for robust image reconstruction.
• Offers a significant advancement for applications requiring hyperspectral imaging without paired data.