Image registration of low signal-to-noise cryo-STEM data

• Published in: Ultramicroscopy Vol. 191; no. C; pp. 56 - 65
• Main Authors: Savitzky, Benjamin H., El Baggari, Ismail, Clement, Colin B., Waite, Emily, Goodge, Berit H., Baek, David J., Sheckelton, John P., Pasco, Christopher, Nair, Hari, Schreiber, Nathaniel J., Hoffman, Jason, Admasu, Alemayehu S., Kim, Jaewook, Cheong, Sang-Wook, Bhattacharya, Anand, Schlom, Darrell G., McQueen, Tyrel M., Hovden, Robert, Kourkoutis, Lena F.
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 01.08.2018; Elsevier
• Subjects: Atomic Resolution; Cryogenic STEM (cryo-STEM); Image reconstruction; Low signal-to-noise ratio (SNR); OTHER INSTRUMENTATION; Rigid registration; Scanning Transmission Electron Microscopy (STEM); Low signal-to-noise ratio (SNR); Scanning Transmission Electron Microscopy (STEM); Rigid registration; Cryogenic STEM (cryo-STEM); Image reconstruction; Atomic Resolution
• ISSN: 0304-3991; 1879-2723
• DOI: 10.1016/j.ultramic.2018.04.008

•An image registration method optimized for low-SNR, cryogenic STEM data is presented.•Registration errors and subtle artifacts involving unit-cell misalignment are detailed.•Enforcing consistency between alignments of all possible image pairs yields robust image reconstructions.•The importance of Fourier space weighting in low-SNR atomic resolution applications is demonstrated.•All source code is freely available in python, including user-friendly Jupyter notebooks and a GUI. Combining multiple fast image acquisitions to mitigate scan noise and drift artifacts has proven essential for picometer precision, quantitative analysis of atomic resolution scanning transmission electron microscopy (STEM) data. For very low signal-to-noise ratio (SNR) image stacks – frequently required for undistorted imaging at liquid nitrogen temperatures – image registration is particularly delicate, and standard approaches may either fail, or produce subtly specious reconstructed lattice images. We present an approach which effectively registers and averages image stacks which are challenging due to their low-SNR and propensity for unit cell misalignments. Registering all possible image pairs in a multi-image stack leads to significant information surplus. In combination with a simple physical picture of stage drift, this enables identification of incorrect image registrations, and determination of the optimal image shifts from the complete set of relative shifts. We demonstrate the effectiveness of our approach on experimental, cryogenic STEM datasets, highlighting subtle artifacts endemic to low-SNR lattice images and how they can be avoided. High-SNR average images with information transfer out to 0.72 Å are achieved at 300 kV and with the sample cooled to near liquid nitrogen temperature.