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

Channeled spectropolarimetry (CSP) employing low-pass channel extraction filters suffers from cross talk and spectral resolution loss. These are aggravated by empirically defining the shape and scope of the filters for different measured. Here, we propose a convolutional deep-neural-network-based channel filtering framework for spectrally-temporally modulated CSP. The network is trained to adaptively predict spectral magnitude filters (SMFs) that possess wide bandwidths and anti-cross-talk features that adapt to scene data in the two-dimensional Fourier domain. Mixed filters that combine the advantages of low-pass filters and SMFs demonstrate superior performance in reconstruction accuracy. (C) 2021 Optical Society of America

Automated summary

The study introduces a deep-learning-based channel filtering framework to address the cross talk and spectral resolution loss issues in channeled spectropolarimetry. Through training, the network can adaptively predict spectral magnitude filters with wide bandwidths and anti-cross-talk features, leading to improved reconstruction accuracy.