Alternating Direction Unfolding With a Cross Spectral Attention Prior for Dual-Camera Compressive Hyperspectral...

Area of Science:

• Computational Imaging
• Image Reconstruction
• Spectral Imaging

Background:

• Coded Aperture Snapshot Spectral Imaging (CASSI) captures dynamic spectral scenes but loses spatial information.
• Dual-Camera Compressive Hyperspectral Imaging (DCCHI) uses a Panchromatic (PAN) camera to recover spatial details, but its structure complicates model-based reconstruction.
• Existing methods struggle with the data subproblem due to disrupted matrix properties in DCCHI.

Purpose of the Study:

• To develop an efficient and accurate method for hyperspectral image reconstruction from DCCHI data.
• To overcome the limitations of traditional model-based methods and Deep Unfolding Networks (DUNs) in DCCHI systems.
• To propose a novel deep unfolding network that addresses the specific challenges of the DCCHI imaging model.

Main Methods:

• Proposed an Alternating Direction DUN (ADRNN) that decouples the DCCHI model into CASSI and PAN subproblems.
• ADRNN analytically solves data terms and a joint prior term iteratively within these subproblems.
• Introduced a Cross Spectral Transformer (XST) with cross spectral attention and Grouped-Query Attention (GQA) to leverage joint priors and reduce computational cost.

Main Results:

• The proposed ADRNN method demonstrates state-of-the-art (SOTA) performance in hyperspectral image reconstruction.
• Experiments on both simulated and real DCCHI datasets validate the effectiveness of the ADRNN and XST approach.
• A real DCCHI system was built, and large-scale datasets were captured for further research.

Conclusions:

• ADRNN effectively reconstructs hyperspectral images from DCCHI data by addressing the inherent imaging model challenges.
• The integration of XST enhances the exploitation of joint priors between compressed HSI and PAN images.
• The open-sourced code and datasets facilitate future research in compressive hyperspectral imaging.