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

Area of Science:

• Optics and Photonics
• Image Reconstruction
• Computational Imaging

Background:

• Scattering media significantly degrade image quality and obscure fine details.
• Polarimetric imaging offers potential for enhanced imaging through scattering, but noise propagation is a challenge.

Purpose of the Study:

• To introduce PSFF-Net, a self-supervised framework for high-fidelity image reconstruction in polarimetric single-pixel imaging.
• To enable effective imaging through scattering media without requiring ground truth data or offline pretraining.

Main Methods:

• A physics-based projection maps compressive measurements to a spatial initialization.
• A spatio-frequency fusion core integrates polarization channels using cross-attention and frequency-decoupled fusion.
• The framework optimizes reconstructions using only acquired measurements and the sensing operator.

Main Results:

• PSFF-Net demonstrates improved reconstruction quality compared to existing methods.
• Performance gains are particularly notable at low sampling rates.
• The method effectively limits cross-channel noise propagation.

Conclusions:

• PSFF-Net offers a robust and efficient solution for imaging through scattering media.
• The self-supervised, physics-based approach advances the capabilities of polarimetric single-pixel imaging.
• This framework holds promise for applications requiring high-resolution imaging in challenging environments.