Joint optimization of coded aperture metasurface and residual self-attention network for snapshot full-Stokes imaging

Abstract

Traditional full-Stokes polarization imaging typically relies on the movements or segmentation of imaging systems, often accompanied by sacrifices in temporal or spatial resolution. Therefore, simultaneous encoding of full-Stokes vectors at the pixel scale is of great significance. Benefiting from the multi-dimensional light field control capability of metasurfaces, a coded aperture metasurface for polarization imaging is proposed in this paper, which can achieve pixel-level encoding of four Stokes vectors in a single imaging session. In addition, a Stokes residual self-attention network is designed to restore the encoded image, where the introduction of a channel-wise self-attention mechanism can effectively address the impact of intensity differences between Stokes vectors. Since the control of polarization states by metasurfaces is a differentiable physical process, the front-end metasurface encoding and back-end recovery network parameters can be jointly optimized, and this work achieves high-quality polarization imaging via such co-optimization methods. The proposed work demonstrates the flexibility and designability of metasurfaces in compact computational full-Stokes imaging systems.