RCUMP: Residual Completion Unrolling With Mixed Priors for Snapshot Compressive Imaging

Area of Science:

• Optics and Photonics
• Computer Vision
• Signal Processing

Background:

• Deep unrolling methods excel at reconstructing 3D hyperspectral images (HSIs) from 2D measurements using coded aperture snapshot spectral imaging (CASSI).
• Existing methods face limitations due to Taylor approximation residuals and restricted representation capabilities of single priors.

Purpose of the Study:

• To develop a novel HSI reconstruction method that overcomes the limitations of existing deep unrolling techniques.
• To enhance the representation ability and reduce memory costs in HSI reconstruction.

Main Methods:

• Proposed Residual Completion Unrolling with Mixed Priors (RCUMP), a novel HSI construction method.
• RCUMP incorporates a residual completion branch to address residual issues.
• Employed mixed priors, including a deep sparse prior and a mask prior, to improve representation.

Main Results:

• The proposed CNN-based RCUMP model demonstrated significant memory cost reduction compared to previous CNN methods.
• RCUMP achieved superior performance over state-of-the-art transformer and RNN-based methods.
• Consistent outperformance across 10 scenes compared to 9 recent baselines, with memory consumption reduced by up to 80%.

Conclusions:

• RCUMP effectively addresses residual problems and enhances representation ability in HSI reconstruction.
• The method offers a substantial improvement in both accuracy and efficiency, significantly reducing memory footprint.
• RCUMP represents a significant advancement in hyperspectral image reconstruction technology.