Snapshot spectral imaging based on aberration model-driven deep learning.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Machine Learning for Imaging

Background:

• Coded aperture snapshot spectral imaging (CASSI) enables single-shot hyperspectral image (HSI) acquisition.
• Optical aberrations in CASSI systems degrade the quality of reconstructed HSIs.
• Current deep learning methods struggle with real-world CASSI data due to these aberrations.

Purpose of the Study:

• To develop a robust method for restoring high-resolution HSIs from low-resolution CASSI measurements.
• To address the performance limitations of existing deep learning techniques caused by optical aberrations.
• To improve the adaptability and generalizability of CASSI reconstruction algorithms.

Main Methods:

• Generated realistic training data simulating CASSI optical aberrations using spectral imaging simulation.
• Trained a generative network on simulated data to recover HSIs from blurred and distorted CASSI measurements.
• Developed a method that adapts to the optical system degradation model for improved robustness.

Main Results:

• Significantly enhanced image quality in reconstructed HSIs compared to existing methods.
• Demonstrated improved reconstruction robustness on both simulated and real-world CASSI data.
• Validated the method's applicability across different CASSI system configurations.

Conclusions:

• The proposed generative network approach effectively mitigates optical aberrations in CASSI.
• The method offers a robust solution for high-resolution HSI reconstruction from degraded measurements.
• This technique shows promise for enhancing the performance of various CASSI systems.