Electron Microscopes have been used to investigate materials from micron to nano scale. Scanning electron microscopes (SEM) as well as scanning transmission electron microscopes (STEM) can acquire image data relatively fast, however acquiring spectroscopic data requires longer data collection times. Depending on the desired resolution or sample area, this can make a significant difference in the duration and feasibility of the experiment. Moreover, for electron beam sensitive samples, it is necessary to acquire the image data with minimal exposure time as not to further damage the sample [1]. Here, we propose an under-sampling and reconstruction method to reduce the data collection time while maintaining imaging accuracy.

A training dataset consisting of several images with features of interest is initially collected. Small patches of nxn pixels are extracted from this set of images and flattened into a library of column vectors. Each column in the library is normalized by subtracting the mean and dividing by the standard deviation of that column. Singular Value Decomposition (SVD) is then applied on the library and [US V] matrices are calculated. Left singular vectors of the matrix U becomes our dictionary D, whose columns correspond to the features for image reconstruction.