Non-Line-of-Sight Full-Stokes Polarimetric Imaging with Solution-Processed Metagratings and Shallow Neural Networks

There is currently significant interest in approaches that combine metaphotonics with back-end algorithms to advance imaging capabilities with less complicated hardware. Here, we combine computational imaging with a low-cost polarimetric encoder to construct a non-line-of-sight full-Stokes polarimetric camera. The polarimetric encoder is a multiscale, solution-processed metagrating composed of conducting-polymer nanofibers. We image the highly corrugated speckle patterns from the metagrating with a polarization-agnostic CCD sensor and achieve full-Stokes imaging from single-image capture with a trained shallow neural network (SNN) model. As SNNs require large amounts of training data, we present an effective method to generate numerous polarimetric scenes that span the full range of the Poincare ' sphere. To guide the paired encoder and algorithm design, we also compare the reconstruction performance of SNN to pseudoinverse algorithms with varied sensor sampling size and explore the issues of compressive sensing. Our results provide new guidelines and impressive possibilities with meso-ordered, multiscale, self-assembled materials in future hybrid computing and polarimetric imaging systems.