Learned large Field-of-View imager with a simple spherical optical module

Multiple lenses are used in most modern imaging systems to reduce deviations from the perfect optical imaging, which also results in a significant increase in prices. Computational Imaging Technology (CIT), which combines the traditional optical design and image reconstruction algorithms, has provided many methods for simplifying optical systems in recent years. However, the Field-of-View (FOV) and the relatively simple image degradation model limit the CIT approach. In this work, we present a novel and low-cost CIT approach for large-FOV imaging. Our system consists of a wide-angle optical module with two spherical lenses and a deep learning network for image reconstruction. Aiming at improving image quality, we introduce the Weighted Patch Degradation Model (WPDM) for the simple optical module with a wide range of spatial variants for large FOV and then construct a dataset. In addition, we present a DMPH-SE Network for our reconstruction task. Experiments show that our large-FOV imager could obtain excellent imaging results with a simple optical structure.