Spectral Super-Resolution via Model-Guided Cross-Fusion Network.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Spectral super-resolution aims to reconstruct hyperspectral images (HSIs) from limited RGB data.
• Convolutional Neural Networks (CNNs) show promise but struggle to integrate imaging models with HSI characteristics.

Purpose of the Study:

• To develop a novel model-guided network (SSRNet) for improved spectral super-resolution.
• To effectively leverage both the imaging model and complex HSI spatial-spectral features.

Main Methods:

• Proposed a cross fusion (CF)-based model-guided network (SSRNet).
• Introduced an HSI prior learning (HPL) module with two subnetworks for spatial and spectral priors.
• Incorporated an imaging model guiding (IMG) module for adaptive feature optimization and merging.
• Alternately connected HPL and IMG modules for optimal HSI reconstruction.

Main Results:

• SSRNet demonstrated superior spectral reconstruction performance on both simulated and real data.
• The proposed method achieved high-quality results with a relatively small model size.
• Cross fusion strategy enhanced CNN learning performance by connecting subnetworks.

Conclusions:

• SSRNet effectively addresses limitations of existing CNNs in spectral super-resolution.
• The network successfully integrates imaging models with HSI characteristics for enhanced reconstruction.
• The proposed approach offers a promising solution for high-fidelity HSI reconstruction.