Abstract
We present an unsupervised machine learning approach for segmentation of static and dynamic atomic-resolution microscopy data sets in the form of images and video sequences. In our approach, we first extract local features via symmetry operations. Subsequent dimension reduction and clustering analysis are performed in feature space to assign pattern labels to each pixel. Furthermore, we propose the stride and upsampling scheme as well as separability analysis to speed up the segmentation process of image sequences. We apply our approach to static atomic-resolution scanning transmission electron microscopy images and video sequences. Our code is released as a python module that can be used as a standalone program or as a plugin to other microscopy packages.
Highlights
The substantial advances of microscopy techniques and aberration-corrected electron optics since the second half of the last century have made it possible to probe materials at atomic resolution (Hansma et al, 1988; Haider et al, 1998; Batson et al, 2002), which opened up an era where materials can be studied down to the level of single atoms (Krivanek et al, 2010)
We present an unsupervised machine learning approach for segmentation of static and dynamic atomic-resolution microscopy data sets in the form of images and video sequences
We evaluate the symmetry score equation (3) for the selected translation shifts directly in the low-level C code that is parallelized with open multi-processing (OpenMP) and wrapped as a python module
Summary
The substantial advances of microscopy techniques and aberration-corrected electron optics since the second half of the last century have made it possible to probe materials at atomic resolution (Hansma et al, 1988; Haider et al, 1998; Batson et al, 2002), which opened up an era where materials can be studied down to the level of single atoms (Krivanek et al, 2010). Supervised algorithms use descriptor-based classifiers, but the classification nowadays employs nonlinear mappings (e.g., deep neural networks) trained on pre-labeled data Such machinelearning methods (Arganda-Carreras et al, 2017; Vu et al, 2019) have been explored in the biological community to segment tissue images, which require a large amount of labeled data to train a classification model or a neural network, which are not yet available for atomic scale images of crystalline matter. To summarize the requirements at this point, we aim at a segmentation algorithm that (1) detects crystallographic features of the imaged area without prior knowledge, (2) is robust against relatively high noise levels, (3) is computationally fast, and (4) is interpretable in each of its stages With these criteria in mind, we developed an unsupervised segmentation algorithm based on local symmetry in atomic-resolution images, which is encoded in local descriptors.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
