Fig. 7: Comparison of denoising approaches for a wedge-shaped sample.
From: Unsupervised deep denoising for four-dimensional scanning transmission electron microscopy
a The blue structure inset shows a side view of the sample geometry with varying thickness from 80 Å to 120 Å along [100] direction (the electron beam is in the vertical direction). The yellow image above the structure shows the STEM image. The vertical axis, EI, gives the Frobenius norm (FN) and cross-correlation coefficient (CC) metrics for both denoising methods to compare the denoised ID and noise free I* datasets. The black arrows show the two rows (horizontal) and four columns (vertical) of where the diffraction patterns in (b, c) are sampled from. CBED patterns from locations (b) in between atoms (from the upper row in the yellow image in (a)) and (c) closest to heavy atoms (from the lower row in the yellow image in (a)), for noise-free, noisy and denoised datasets using both methods. The noisy data has integer values. The overall electron counts are less for locations on heavy atoms due to absorption (i.e. thermal scattering to high angles).
