Combining Low-Rank and Deep Plug-and-Play Priors for Snapshot Compressive Imaging.

Area of Science:

• Optics and Photonics
• Computer Vision
• Signal Processing

Background:

• Snapshot compressive imaging (SCI) enables 3-D hyperspectral image (HSI) acquisition using 2-D detectors, but reconstruction from compressed data is challenging.
• Existing methods often overlook spectral correlations or require extensive training datasets, limiting their performance and applicability.
• A need exists for reconstruction techniques that are efficient, generalizable, and interpretable.

Purpose of the Study:

• To develop a novel approach, LR2DP, for reconstructing HSI from SCI measurements.
• To integrate model-driven low-rank priors with data-driven deep priors for improved reconstruction.
• To address the limitations of current SCI reconstruction methods regarding spectral correlation and training data dependency.

Main Methods:

• Proposed LR2DP method leverages low-rank priors to exploit spectral correlations in HSI.
• Integrated unsupervised deep image prior (DIP) and pre-trained deep denoising prior (DDP) within a plug-and-play (PnP) framework.
• Employed half-quadratic splitting (HQS) for optimizing the HSI reconstruction model.

Main Results:

• LR2DP effectively captures spectral correlations and deep spatial features of HSI.
• The method demonstrates superior performance compared to state-of-the-art approaches on simulated and real datasets.
• Achieved a balance between performance, generalizability, and interpretability without additional training datasets.

Conclusions:

• LR2DP offers a robust and efficient solution for hyperspectral image reconstruction in SCI.
• The integration of low-rank and deep priors provides a powerful framework for addressing ill-posed inverse problems.
• The proposed method advances the field of compressed sensing for hyperspectral imaging.