Non-local denoising and unsupervised quantitative analysis in scanning transmission electron microscopy
- Nicht-lokales Entrauschen und unüberwachte quantitative Analyse in der Rastertransmissionselektronenmikroskopie
Mevenkamp, Niklas; Berkels, Benjamin (Thesis advisor); Dahmen, Wolfgang (Thesis advisor); Binev, Peter (Thesis advisor)
Dissertation / PhD Thesis
Dissertation, RWTH Aachen University, 2017
Modern scanning transmission electron microscopes (STEM) provide atomic resolution images of inorganic materials. Atom positions and other information extracted from such images are key ingredients in the understanding of the relation between microscopic material structure and macroscopic material properties. Unfortunately, in certain applications, the ability to obtain this information is severely obstructed by low signal-to-noise ratios of the acquired images. These result from restrictions on the electron dose (exposure time) due to potential beam damage. Aiming at resolving this deficiency, we propose an effective denoising strategy that is specifically tailored to the special structure of atomic-resolution crystal images. It is based on the non-local denoising principle and uses the block-matching and 3D filtering algorithm (BM3D) by Dabov et al. as a starting point. We employ an adaptive piecewise periodic block-matching strategy that exploits the crystal geometry - accounting for typical irregularities such as dislocations, crystal interfaces and image distortion - and provides efficient and truly non-local denoising. The required crystal geometry information is extracted from the noisy raw image in an unsupervised fashion. To this end, we present novel real-space algorithms for accurate unit cell extraction and crystal segmentation. Furthermore, we analyze the noise behavior of experimental high-angle annular dark-field (HAADF)-STEM images and show that a simple additive Gaussian white noise model is not suitable for low-dose images. Instead, we propose to employ a more complex mixed Poisson-Gaussian noise model which results in a much better fit and present an unsupervised algorithm to estimate the required noise parameters from a given raw image. Then, the generalized Anscombe transform is used for variance-stabilization, which enables the use of BM3D for noise removal. Results on both artificial and real experimental single-shot HAADF-STEM images are presented which show that the proposed method significantly improves the visual quality and, more importantly, the precision of detected atom positions. We also present an extension of the method to series of images including a coupling of non-local denoising with non-rigid image alignment. An evaluation based on experimental images reveals that compared to plain averaging of an aligned image stacks the number of frames required to obtain a high SNR reconstruction can be significantly reduced. Also, we show that this way state-of-the-art precisions can be obtained using less than ten frames. Besides this, we propose an extension of our denoising method to spectral data and present very promising results on an electron energy loss spectroscopy dataset. Finally, we present a multi-modal and multi-scale similarity measure intended for joint denoising of STEM and spectral data. Using a jointly acquired dataset consisting of an HAADF-STEM image and an energy-dispersive X-ray scan, we demonstrate that extreme gains in SNR are achievable without noticeably sacrificing spatial resolution.