Non-Local Denoising and Unsupervised Quantitative Analysis in Scanning Transmission Electron Microscopy
Images of materials at atomic scale can be obtained using Scanning Transmission Electron Microscopy (STEM). Due to the combination of the large magnification and the line-by-line and pixel-by-pixel acquisition in STEM imaging, movements of the observed material at the nanometer scale lead to local distortions in the image. On the one hand, these distortions are amplified with increased exposure time. On the other hand, an increased exposure time results in a better signal-to-noise ratio. Both the distortions and a low signal-to-noise ratio decrease the precision when determining the exact positions of the atoms, which are important for the research of material properties in STEM . High precision reconstructions have been achieved based on a recently proposed non-rigid, non-parametric registration method, which increases the signal-to-noise ratio and reduces distortion by combining the information contained in multiple frames of an entire image series [1,6]. A disadvantage of this technique is that it relies on a large series of images and an overall electron dose that only few materials can withstand, thus making it only applicable to a small range of materials.
The aim of this project is to develop a reconstruction technique that achieves a comparable precision as the previously mentioned one with a reduced overall electron dose. The main focus of this project lies on the investigation of non-local regularization techniques that exploit the self similarity of the data to reduce the dwell time required for a good signal-to-noise ratio. The starting point of the project is the classical non-local means algorithm (NLM) . We will treat the adaptation of the patch similarity measure to the characteristics of typical STEM images, as well as the efficient search for similar patches. The newly developed denoising method will then be combined with the already existing non-rigid registration based image reconstruction techniques.