Abstract
Spatially modulated full-polarization imaging enables the acquisition of full Stokes parameters in a single snapshot. However, at low carrier frequencies or in complex noise environments, conventional frequency-domain demodulation is susceptible to spectral aliasing, channel crosstalk, and loss of high-frequency polarization information. To address these limitations, this study proposes a physics-driven robust attention generative adversarial network (PRA-GAN) for end-to-end demodulation from a single interferogram to four Stokes-parameter images (S0–S3). By incorporating robust residual modules, spatial attention, and a PatchGAN discriminator, the network enhances the recovery of high-frequency fringes and weak polarization structures; it also introduces forward interferogram reconstruction and degree-of-polarization and angle-of-polarization constraints to further improve physical consistency. Simulations and validation experiments based on real-scene polarization measurement data are conducted to compare PRA-GAN with conventional frequency-domain methods and the StandardUNet baseline, while module-wise ablation experiments evaluate the effects of the network improvements and physical constraints. PRA-GAN substantially improves full-Stokes demodulation quality, achieving an average SSIM of 0.96 and a PSNR above 38 dB for the S0–S3 components. These results demonstrate its potential for improving spatially modulated full-polarization imaging.
Funding Information
  • National Key Research and Development Program of China (2022YFB3901800)
  • National Key Research and Development Program of China (2022YFB3901803)
  • Science and Technology Department of Guangxi Zhuang Autonomous Region (AD25069073)
  • National Research Center of Intelligent Equipment for Agriculture Project (KFZN2026W001)
  • Innovation Project of GUET Graduate Education (2026YCXS250)
  • Guangxi University Young and Middle-aged Teachers’ Basic Research Ability Improvement Projects (2025KY0248)
  • Guangxi Key Laboratory of Optoelectronic Information Processing (GD24102)

This publication has 36 references indexed in Scilit: