Machine-learning approach for quantified resolvability enhancement of low-dose STEM data

Abstract High-resolution electron microscopy is achievable only when a high dose employed, practice that may cause damage to the specimen and, in general, affects observation. This drawback sets some limitations on range of applications high-resolution microscopy. Our work proposes strategy, based machine learning, which enables significant improvement quality Scanning Transmission Electron Microscope images generated at low dose, strongly affected by Poisson noise. In particular, we develop an autoencoder, trained large database images, thoroughly tested both synthetic and actual data. The algorithm demonstrated drastically reduce noise level approach ground-truth precision over broad beam intensities. Importantly, it does not require human data pre-processing or explicit knowledge employed can run speed compatible with live acquisition. Furthermore, quantitative unbiased benchmarking protocol proposed compare different denoising workflows.