Three-dimensional light field displays are not yet widely adopted due to the bulky form-factor and limited image quality caused by optical aberrations of microlens arrays (MLAs). Conventional optimization techniques cannot approach the maximal displayed image quality as they rely on intermediary metrics such as focal spot size. In order to optimize for the full-color wide field-of-view image quality, the point spread function of the MLA should be modeled to provide more flexibility. We developed a modeling approach for both refractive and metasurface MLAs and assessed the accuracy by judicious comparisons between the numerical simulations and experimental characterization.

Three-dimensional Light Field Displays (LFDs) promise to provide realistic and comfortable viewing for one or multiple users simultaneously without any eyewear by overcoming the vergence-accommodation conflict. However, LFDs have not yet gained widespread adoption and remain a hot topic of research. Currently, LFDs are based on refractive Microlens Array (MLA) optics, which have inherent limitations including high optical aberrations and/or bulkiness. Metasurfaces are flat optics made of a distribution of subwavelength size nanopillars that can manipulate light wave properties including phase, amplitude, and polarization and be fabricated in a single lithographic step. They can be used as a more compact alternative to refractive MLAs. However, current designs cannot achieve comparable full-color and wide field-of-view imaging by multiple layers of refractive lenses. In this work, we demonstrate a deconvolution neural network model based on the U-Net architecture and Wiener non-blind deconvolution that reduces the effects of aberrations caused by a designed metasurface, enabling high image quality 3D LFDs. We employ an analytical model to determine the metasurface phase profile and point spread function for a five-by-five view LFD scenario. Our model is trained and evaluated using 52 images of 8.1 megapixels each from online databases of multiview images. To minimize the spatially varying aberration effects, a loss function is used that incorporates spatial pixel-wise error, structural quality, and angular consistency. Compared to the output images without preprocessing images using the designed PSFs, our neural network model improved PSNR by 10 dB and MS-SSIM by 2% overall for all views and reduced variations between different views by 40% and 70%, respectively, for PSNR and MS-SSIM.