Prior Image Guided Snapshot High-Resolution Spectral Imaging in Near Infrared.

Area of Science:

• Optics and Photonics
• Computer Vision
• Spectroscopy

Background:

• Near-Infrared (NIR) hyperspectral imaging offers vast application potential but faces challenges in dynamic scenes.
• Compressive Spectral Imaging (CSI) is promising for 3D spatio-spectral data but is limited by costly InGaAs Focal Plane Arrays (FPAs) in the NIR spectrum, degrading high-frequency information.

Purpose of the Study:

• To develop a novel system for high-resolution NIR spectral imaging overcoming the limitations of current CSI techniques.
• To introduce a new deep learning framework for high-fidelity spectral reconstruction using prior image guidance.

Main Methods:

• Demonstration of EpiSpec, a system utilizing the tail response of low-cost silicon sensors for high spatial resolution guidance.
• Proposal of the Prior Image Guided Deep Unfolding Framework (PIUF) to handle content-aware chromatic errors and reconstruct high-fidelity spectral data.
• Introduction of the NIR Spectral Images Dataset (NISID) for evaluating spectral sparse structures and designing CSI sensing matrices.

Main Results:

• EpiSpec effectively reconstructs high-quality hyperspectral scenes from degraded measurements by leveraging prior image guidance.
• The PIUF framework implicitly models non-linear relationships for accurate spectral reconstruction.
• Evaluations on the NISID dataset and representative CSI systems confirm the effectiveness of the EpiSpec framework.

Conclusions:

• EpiSpec provides a viable solution for high-resolution NIR spectral imaging, overcoming cost and resolution limitations.
• The developed framework and dataset contribute to advancements in CSI and spectral reconstruction techniques.
• Lab prototypes validate the practical applicability of EpiSpec for real-world NIR spectral imaging.