Deep Learning-Based Imaging using Single-Lens and Multi-Aperture Diffractive Optical Systems

Artem V. Nikonorov, Viktoria Evdokimova, Maksim V. Petrov, Pavel Y. Yakimov, Sergey A. Bibikov, Yuriy V. Yuzifovich, Roman V. Skidanov, Nikolay Lvovich Kazanskiy

Abstract: The pressure to reduce weight and improve image quality of the imaging devices continues to push research in the area of flat optics with computational image reconstruction. This paper presents a new end-to-end framework applying two convolutional neural networks (CNNs) to reconstruct images captured with multilevel diffractive lenses (MDLs). We show that the patch-wise chromatic blur and image-wise context-aware color highlights, the distortions inherent to MDLs, can be successfully addressed with the suggested reconstruction pipeline. The generative adversarial network (GAN) is first used to remove image-wise color distortion, while a patch-wise network is then used to apply chromatic deblur. The proposed approach produces better image quality improvement than the context-independent color correction with a deconvolution-based chromatic deblur. We also show that the proposed end-to-end reconstruction is equally applicable for single-and multi-aperture MDL-based imaging systems.