TEM data compression and denoising by autoencoders and convolutional neural networks

Abstract text (incl. figure legends and references)

Introduction

Machine learning methods such as Principal Component Analysis are widely employed for data denoising and compression in Electron Microscopy and Spectroscopy [1]. Recently, neural network based methods have been successfully introduced to the field, potentially increasing robustness toward noise , while increasing the application range.

Objectives

Here, we apply suitable autoencoder neural networks [2,3,4] to denoise and compress electron microscopy data comprising magnetic imaging and spectrum imaging. Our focus is removal of ubiquitous Poissonian noise, scattering artifacts and identification of optimal compression thresholds.

Materials and Methods

Electron Microscopy datasets comprise different types of magnetic images. It can be vortex matter or electron energy loss / energy dispersive x-ray spectrum images. In particularly the latter is notoriously affected by Poissonian noise as the spectra comprise regions of very low and large intensity.

The autoencoder neural network consist of a 2-3 layers of convolutionary and pooling type on the encoder side and their analogues on the decoder side. As training data we employed several images with lowest noise level. In this work were used just 4 images. For extending of training dataset was applied scaling with different scale factors. All neural networks were implemented utilizing the Tensorflow and Keras frameworks.

Results

Autoencoder network as applied to denoising of magnetic images in Fig. 1a.

In this example the image data contains a regular hexagonal structure. Due to not perfect sample composition, not optimal exposure, image drift and other experimental challenges, the image contains a lot of sharp artifacts, contrast variations and of course the experimental noise.

The application of the autoencoder (Fig. 1b) removes a large part of the noise, rendering the dataset amenable to, e.g., vortex positions and anomalies analysis. Closer analysis reveals that the data compression is similar to a truncation of the Fourier basis in this particular example, where the periodic repetition of the contrast is the dominating feature.

Conclusion

Autoencoder neural networks represent a viable alternative to conventional denoising algorithms for treatment of electron microscopy imaging and spectrum imaging data. We demonstrate successful denoising utilizing relatively simple network designs and when applied to different types of images.

References

1. Jolliffe, I.T.: Principal component analysis, 2nd edn. Springer Verlag, Berlin (2002)

2. Lovedeep Gondara, "Medical Image Denoising Using Convolutional Denoising Autoencoders", 2016 IEEE 16th International Conference on Data Mining Workshops (ICDMW).

3.Yasenko, Y. Klyatchenko and O. Tarasenko-Klyatchenko, "Image noise reduction by denoising autoencoder", 2020 IEEE 11th International Conference on Dependable Systems Services and Technologies (DESSERT), pp. 351-355, 2020.

4.https://keras.io/examples/vision/autoencoder/