Improvement of ultra-low dose electron image of TEM by deep learning

Abstract

  We apply a convolutional neural network (CNN) to low-dose electron images obtained by a transmission electron microscope. Using our original dataset including short-exposure images and long-exposure images, the CNN model is trained. The CNN model produced images that were equivalent to images 10000 times brighter. The particles in improved images become clearer and can be counted. The waiting time for image conversion is approximately 8 ms, and the in-situ observation is also possible.