Framework for denoising transmission electron microscope (TEM) data. We design a deep feedforward convolutional neural network with about 1 million parameters. The network is trained using a large dataset of simulated images, which are variations of a 3D atomic model similar to the one observed in the experimental data (top row). The variation of different structures and electron-optic parameters leads to almost 20,000 simulated images, which are corrupted by Poisson noise and fed into the network (middle row). The network parameters are calibrated to optimize the mean squared error between the denoised output of the network and the true simulated image. The network can then be applied to real noisy images, as shown in the bottom row of the diagram. In contrast to network architectures typically used to denoise natural images, we employ a network with a very large field of view known as UNet. The UNet achieves a large field of view by applying downsampling operations in the intermediate layers. Our experiments show that this architecture outperforms baseline networks by a significant margin. A linear-algebraic analysis of the trained network reveals that the large field of view enables it to exploit the periodic structure of the atoms.
Get new content delivered directly to your inbox.