High-accuracy reconstruction of Stokes vectors via spatially modulated polarimetry using deep learning at low light field

Abstract

Polarization measurement is generally performed in scenes with a low signal-to-noise ratio (SNR) such as remote sensing and biological tissue detection. The spatially modulated polarimeter can satisfy the real-time measurement requirements in low SNR scenes by establishing the mapping between photon spatial distribution and polarization information. However, accurately measuring the polarization state under low-light illumination becomes highly challenging owing to the interference of background noise. In this paper, a deep learning method is proposed and applied to the high-accuracy reconstruction of polarization information at low light field. A reinforced two-layer deep convolutional neural network is designed to respectively extract global and local features of noise in this method. Accurate photon spatial distribution can be obtained by fusing and processing these features. Experimental results illustrate the excellent accuracy achieved by the proposed method with a maximum average value of the absolute measured error below 0.04. More importantly, the proposed method is well-performed for the reconstruction of Stokes vectors at low light fields of various levels without requiring changes to the model, enhancing its practicality and simplicity.