Fig. 13: Reconstruction error with varying network size.
From: Unsupervised deep denoising for four-dimensional scanning transmission electron microscopy
The effect of changing the size of the neural network by multiples of number of kernels of the network on the reconstruction error (comparing the noise-free against the denoised, scaled by Ne to allow direct comparison, such that low error means the denoising has been effective) and cross-correlation, CC, is shown for (a) the wedge-shaped SrTiO3 sample (as considered in Fig. 7) and (b) 28 Å CeB6 (as considered in Fig. 8a–e), for five different dose levels, shown by colour and measured by the number of electrons per diffraction pattern Ne.
