Channeled imaging spectropolarimeter reconstruction by neural networks

Area of Science:

• Optics and Photonics
• Computational Imaging
• Data Science

Background:

• Snapshot channeled imaging spectropolarimetry (SCISP) enables rapid spectral and polarization imaging in a single exposure.
• Traditional SCISP methods relying on Fourier transforms face limitations like channel crosstalk and reduced accuracy.
• Complex phase calibration in Fourier-based SCISP hinders practical applications.

Purpose of the Study:

• To introduce a novel neural network (NN)-based technique to replace Fourier reconstruction in SCISP.
• To overcome inherent limitations of Fourier transform-based SCISP, improving data quality and stability.
• To demonstrate a more robust and accurate method for spectral and polarization imaging.

Main Methods:

• Development of specialized data generators to create abundant spectral and polarization datasets.
• Training neural networks (NNs) to learn the forward conversion process by minimizing a loss function.
• Implementing the trained NNs for reconstructing spectral, polarization, and spatial information from SCISP data.

Main Results:

• The NN-based method effectively learns the complex forward conversion procedure.
• Stable output containing comprehensive spectral, polarization, and spatial information is achieved.
• Experimental validation confirms the technique's utility and reliability, maintaining high accuracy.

Conclusions:

• Neural networks offer a powerful alternative to Fourier reconstruction for SCISP.
• The proposed NN-based technique significantly improves the accuracy and stability of spectral and polarization imaging.
• This advancement holds promise for enhanced optical sensing and imaging applications.