Deep learning is the application of neural networks with many layers to machine learning problems. We expect that deep learning can be a game-changer for analyzing data from Transmission Electron Microscopy (TEM). It can automatize the analysis of large amounts of data, eliminate the bias of the person analysing the images, provide quantitative and statistical data, and even help extracting information that is normally difficult to get from TEM.
In a collaboration between the CAMD theory group at DTU Physics and the electron microscopists at DTU CEN/Danchip we have developed a method for training deep convolutional neural networks on simulated TEM images and applying them on real experimental data . We have made our prototype software publicly available .
Unlike most image analysis problems, we are in the beneficial and quite unusual situation of having access to almost free training data, since high-resolution TEM images can be simulated reliably from atomic structures, including realistic noise and imperfections in the microscope optics.
The neural network architecture is inspired by the networks generally used for automated image analysis and segmentation. However, the network architecture has almost not been optimized, nor has the training procedure. We expect that the performance can be significantly enhanced by addressing this.
This project has two aspects, a BSc project will focus one one of them while a MSc will focus on both.
* To improve the neural network by optimizing the network architecture and the training algorithm, and by incorporating recent ideas from the litterature on image analysis and image segmentation.
* To turn the current collection of scripts and software snippets into a package that can realistically be used by an electron microscopist with no prior experience in machine learning.experience with
Figure: The architecture of the neural network . Information flows from left to right. The different-colored rectangles refer to the different architecture elements. Below the rectangles, the spatial and channel dimensions are given as height × width × number of channels (feature maps). The features are downsampled in an encoding path and upsampled through a decoding path in order to represent non-local information. Skip connections ensure that it is possible to retain the original spatial information. Although the size of the input images is shown as 256 × 256, this is not part of the network architecture, and the network can be used on images of any size.
 J. Madsen et al.: A Deep Learning Approach to Identify Local Structures in Atomic-Resolution Transmission Electron Microscopy Images, Adv. Theory Simul
, 1800037 (2018). Preprint available at https://arxiv.org/abs/1802.03008
Prior knowledge of Python and neural networks. Some knowledge of solid state physics and/or electron microscopy is an advantage, but not required.