Automated phase reconstruction and super-resolution with deep learning in digital holography

Abstract

Digital holography can provide quantitative phase images that are related to the shape and content of biological samples. In particular, high-resolution phase images contain more important details and information about the sample. However, to obtain a high-resolution phase image, various phase reconstruction processes must be performed, while the finite pixel size of the sensor needs to be overcome. We propose a deep learning model that can obtain high-resolution phase images from low-resolution holograms. The proposed model consists of image translation and super-resolution parts, and performs phase reconstruction and the super-resolution process at the same time. We successfully generated sophisticated phase values that closely resembled real images for three scaling factors of (×2, ×3, and ×4). Comparative evaluations with various deep learning models demonstrated the favorable performance of our proposed model. Multi-scale training was also possible, so it was shown that high-resolution phase images could be generated regardless of the scale factor. The proposed model can automatically generate accurate high-resolution phase images from low-resolution holograms, reducing the cost of digital holography, and providing great benefits to biological sample measurement processes.