Physics-initialized self-supervised spatio-frequency fusion for polarimetric single-pixel imaging through scattering media

Imaging through scattering media is difficult because scattering degrades measurement quality and suppresses fine spatial details. We present PSFF-Net, a self-supervised reconstruction framework for polarimetric single-pixel imaging that operates without ground truth or offline pretraining, optimizing each target using only the acquired measurements and the known sensing operator. A physics-based projection maps compressive measurements into a spatial initialization, and a spatio-frequency fusion core—combining cross-attention in the spatial domain with frequency-decoupled fusion in the Fourier domain—integrates complementary polarization channels while limiting cross-channel noise propagation. Experiments under controlled scattering conditions show that PSFF-Net improves reconstruction quality over existing methods, particularly at low sampling rates.