Abstract
Miniaturized real-time ultraviolet (UV) full-Stokes polarimeters hold significant application prospects in deep-space exploration, satellite remote sensing, and astronautics. However, UV’s inherently short wavelength, high energy, and strong scattering/absorption properties substantially increase the difficulty of developing polarization elements with both conventional crystals and artificial materials. In this work, we demonstrate a one-snapshot UV full‑Stokes polarimeter based on large‑area disordered calcite crystal films (DCCFs) fabricated by a dry rubbing process and a residual convolutional neural network algorithm. The DCCFs exhibit significant and spatially varying polarization response and can effectively encode the spatial polarization information, while the AI algorithm can achieve high-precision decoding. The detection mean-square error (MSE) at 365 nm can reach about 0.00011 (S
1
), 0.00045 (S
2
), and 0.00092 (S
3
) with the superpixel size of 100 × 100 pixels, which still maintains an acceptable value of approximately 0.0616 even when reducing the superpixel size to as small as 2 × 2 pixels. In addition, the impact of superpixel size, dataset number, noise, and effective pixel number on the polarization detection accuracy has been systematically investigated. Our work provides new strategies for developing high-performance UV polarimeters and polarization imaging devices.