Abstract: This talk will introduce a framework for processing electron microscopy data obtained using energy dispersive X-ray spectroscopy (EDX). The specimen is observed from different directions in order to obtain a tomographic reconstruction. The physics of the process suggests considering the spectrum as a realization of a linear combination of a few random variables representing the chemical elements in the probe with constants reflecting their relative concentrations. The standard approach is the direct classification of the individual X-rays based on their energy levels. We propose a process involving calculation of filters that are biorthogonal (in expectation) for the system of the random variables. The application of these filters is then followed by a constrained discrete minimization using a priority queue to further improve the learning routine. After transforming the spectra, we can represent each frame (the data obtained at a specific tilt angle) as a collection of element maps. The frames are aligned using a procedure based on local centers of mass of all element maps, where each map is processed separately using tomographic reconstruction based on TV-regularized minimization. The process allows a feedback by analyzing the discrepancies in the fitting term and modifying the priority queue of the discrete minimization. This result generates a more accurate representation of nanoscale chemical structure.
- Күн бұрын
Prof. Peter Binev -Learning Based Sparsity-enhanced Processing of Nanoscale Electron Tomography Data
- Рет қаралды 85
Пікірлер