Physics-guided unsupervised network demodulation technique for a snapshot spatially modulated full-linear polarization imaging

Abstract

Spatially modulated snapshot imaging polarimeters (SMSIPs) exhibit substantial potential for dynamic-scene detection; however, accurately demodulating Stokes parameters from a single interferogram is an inherently underdetermined and ill-posed inverse problem. Traditional demodulation algorithms suffer from resolution loss and inter-channel crosstalk, whereas purely data-driven deep learning methods require costly labeled datasets and are susceptible to nonphysical phase-amplitude crosstalk. To address these challenges, a spatially modulated full-linear polarization imaging system incorporating modified Savart polariscopes is constructed to efficiently encode the first three Stokes parameters of the target. Furthermore, a physics-guided unsupervised cascaded neural network (PGUCN) framework is proposed for high-accuracy demodulation. The method embeds the forward interferometric physical mechanism as a prior constraint into the network, progressively narrows the feasible solution space through three-stage cascaded optimization, and suppresses nonphysical phase-amplitude crosstalk without paired ground-truth data. Simulations and optical experiments show that, compared with representative existing methods, PGUCN reduces the reliance on complex hardware parameter calibration and improves native spatial-resolution preservation, noise robustness, and orthogonal decoupling accuracy. This study provides a physics-guided paradigm for calibration-free, accurate snapshot polarimetric detection in complex environments.