Multiscale RGB-Guided Fusion for Hyperspectral Image Super-Resolution

Area of Science:

• Computer Vision
• Remote Sensing
• Image Processing

Background:

• Hyperspectral imaging (HSI) offers detailed spectral analysis but suffers from low spatial resolution due to sensor limitations.
• Existing super-resolution methods struggle to balance spatial detail recovery with spectral fidelity in HSI.

Purpose of the Study:

• To introduce CGNet, a novel color-guided hyperspectral super-resolution network.
• To enhance spatial resolution of HSI by effectively fusing information from RGB images.

Main Methods:

• CGNet employs a dual-encoder architecture: an RGB encoder for spatial features and an HSI encoder for spectral features.
• A multi-scale fusion decoder integrates features from both modalities for high-resolution HSI reconstruction.
• Training utilizes a hybrid L1 and Spectral Angle Mapper (SAM) loss function.

Main Results:

• CGNet achieved superior performance on ARAD1K and StereoMSI datasets at ×4 and ×6 upscaling factors.
• The network demonstrated significant improvements in Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and reduced SAM and ΔE00.
• Ablation studies validated the effectiveness of the hybrid loss function.

Conclusions:

• CGNet successfully reconstructs high-resolution HSI with sharp spatial structures and preserved spectral fidelity.
• The proposed method outperforms state-of-the-art baselines in hyperspectral super-resolution tasks.
• CGNet offers a promising solution for applications requiring high-resolution hyperspectral data.