Deep digital holographic microscopy

Abstract

This paper presents recent developments in hybrid computational imaging approaches for 3D motion analysis and quantitative phase imaging in automated microscopy. By integrating digital holographic microscopy with deep learning models, this work enhances the capabilities of machine microvision. High-tech fields such as micro-robotics and photonics require accurate, robust and rapid in situ measurements using complex multi-scale parameters. Unlike conventional data-driven methods, physics-based deep learning does not require training or requires minimal data, enabling realtime optical information processing. A prototype of digital holographic microscope, with a hybrid deep neural network for autofocusing and phase imaging is proposed. The main challenge is to achieve sub-micrometre-scale accuracy whilst ensuring reliability and interpretability in environments where the ‘hallucinations’ of standard AI models must be avoided.