Abstract
We have investigated a learning-based model to automatically and accurately reconstruct quantitative phase images from holograms captured by a digital holographic microscope (DHM) operating in non-telecentric regime. Reported automatic reconstruction methods for non-telecentric DHM systems are time consuming, and their performance is highly dependent on the sample field of view and the optical configuration of the system. In a recent work, our research group proposed a generative adversarial network to accurately reconstruct quantitative phase images with minimum phase distortions from a hologram recorded by telecentric-based DHM systems. In this contribution, we have analyzed the performance of such a network to fully compensate and reconstruct holograms recorded by non-telecentric DHM systems without the need for any manual computational processing. This learning-based model was trained and validated using simulated hologram paired with phase image of HeLa kinases.