Model-Guided Deep Hyperspectral Image Super-Resolution.

Area of Science:

• Remote Sensing
• Computer Vision
• Image Processing

Background:

• Hyperspectral images (HSI) face a fundamental trade-off between spatial and spectral resolution.
• Acquiring high-resolution HSI directly is challenging, necessitating fusion of low-resolution (LR) and high-resolution (HR) spectral and spatial data.
• Existing model-based fusion methods lack end-to-end optimization, and current deep learning approaches underutilize HSI domain knowledge.

Purpose of the Study:

• To develop an advanced hyperspectral image super-resolution (HSISR) algorithm.
• To leverage both HSI domain knowledge and deep image priors for improved fusion.
• To address the limitations of existing HSISR methods by incorporating an iterative approach and end-to-end optimization.

Main Methods:

• An iterative HSISR algorithm was developed, integrating a deep HSI denoiser.
• The algorithm was unfolded into a novel model-guided deep convolutional network (MoG-DCN) through end-to-end optimization.
• Subnetworks were used to represent the HSI observation matrix, enhancing network flexibility for various HSI scenarios.

Main Results:

• The proposed MoG-DCN demonstrated superior performance compared to leading HSISR methods.
• The method achieved better results in terms of implementation cost and visual quality.
• Experimental results validated the effectiveness of leveraging HSI domain knowledge and deep image priors.

Conclusions:

• The MoG-DCN effectively addresses the spatial-spectral resolution trade-off in HSI.
• The model-guided network architecture offers flexibility and improved performance in HSISR tasks.
• This approach represents a significant advancement in reconstructing high-resolution hyperspectral images.