Viability of Compressed Sensing as a Dose Reduction Strategy in STEM

Abstract

The goal of dose reduction in scanning transmission electron microscopy (STEM) is to reduce electron beam damage to the specimen as the latter is often related to the total dose [1,2]. Other damage mechanisms such a dose rate effects or the spatial distribution of the dose are not considered here. Compressed sensing (CS) [3] is a recording procedure that, through application of an accompanying reconstruction technique, is capable of retrieving a signal from a surprisingly low amount of measurements. It has thus been considered as a candidate for dose reduction in STEM. However, since the mere number of measurements does not necessarily guarantee a lower dose, it is examined in this work if a CS reconstruction can outperform a conventional Shannon scan that is recorded with the same electron budget and denoised based on the same prior knowledge applied in the CS reconstruction. An important difference with much of CS literature is the use of positive, flux-preserving sensing matrices and the presence of Poisson noise. In CS, the M measurements y are linear combinations of the unknown signal x: y = IAx. Here, I is the beam intensity, the N by 1 vector x is the image written in long vector format, A is the M by N sensing matrix, and M