Spectral-temporal channeled spectropolarimetry using deep-learning-based adaptive filtering.

Area of Science:

• Optics and Photonics
• Computational Imaging
• Artificial Intelligence

Background:

• Channeled spectropolarimetry (CSP) traditionally uses low-pass filters, leading to crosstalk and reduced spectral resolution.
• Empirical filter design for CSP is data-dependent and suboptimal.
• Spectrally-temporally modulated CSP presents unique challenges for traditional filtering methods.

Purpose of the Study:

• To develop an adaptive filtering framework for spectrally-temporally modulated CSP.
• To overcome limitations of traditional low-pass filters in CSP, specifically crosstalk and spectral resolution loss.
• To improve the accuracy of spectral reconstruction in CSP.

Main Methods:

• A convolutional deep neural network (DNN) framework was proposed for adaptive channel filtering.
• The DNN was trained to predict spectral magnitude filters (SMFs) in the 2D Fourier domain.
• SMFs were designed with wide bandwidths and anti-crosstalk properties, adapting to scene data.
• Mixed filters combining low-pass and SMF advantages were evaluated.

Main Results:

• The DNN-based framework adaptively predicts effective SMFs.
• The proposed SMFs exhibit wide bandwidths and anti-crosstalk characteristics.
• Mixed filters demonstrated superior performance in reconstruction accuracy compared to traditional methods.
• The framework effectively addresses spectral resolution loss and crosstalk issues.

Conclusions:

• A novel DNN-based framework offers adaptive channel filtering for CSP.
• The proposed SMFs significantly enhance spectral reconstruction accuracy and reduce artifacts.
• This approach provides a more robust and accurate method for spectrally-temporally modulated CSP analysis.