End-to-End Inverse Designed Metasurfaces for Snapshot RGB-Achromatic Full-Stokes Polarization Imaging

TL;DR

This paper introduces an end-to-end optical-digital system combining metasurfaces and neural networks to enable compact, high-quality snapshot full-Stokes polarization imaging across multiple wavelengths from a single measurement.

Contribution

It presents a novel integrated design of metasurfaces and neural networks for efficient, achromatic polarization imaging, advancing beyond traditional bulky systems.

Findings

Hybrid metasurface-refractive system achieves 30.00 dB PSNR in monochromatic imaging.

Pure meta-optic system achieves 26.94 dB PSNR for monochromatic imaging.

System enables high-performance polarization imaging with a 12x compression ratio.

Abstract

Snapshot full-Stokes polarimetry across multiple wavelengths remains challenging because conventional architectures rely on multiplexed measurements and bulky optics. We present an end-to-end framework that reconstructs RGB full-Stokes images from a single monochrome sensor measurement. The system combines a differentiable 4f optical frontend with a U-Net backend for joint optimization. A metasurface modeled by the multilayer perceptron (MLP) is employed to encode the full-Stokes polarization information. We implement the design in two stages: first in a hybrid metasurface-refractive 4f architecture, and then in a pure meta-optic configuration. On a real-world dataset, the hybrid metasurface-refractive system achieves 30.00 dB peak signal-to-noise ratio (PSNR) and 0.8291 structural similarity index measure (SSIM) for monochromatic imaging in the visible range, and 26.71 dB/0.7044 for RGB-achromatic imaging. The pure meta-optic system yields 26.94 dB/0.7184 for the monochromatic case and 24.10 dB/0.6015 for the RGB-achromatic case. These results show that end-to-end optical-digital co-design enables high-performance snapshot full-Stokes polarimetric imaging at a high compression ratio of 12.