Bayesian Density Estimation via Multiple Sequential Inversions of 2-D Images with Application in Electron Microscopy

TL;DR

This paper introduces a Bayesian approach to estimate material density from 2D electron microscopy images, accounting for unknown correction functions and high contrast discontinuities, using multiple imaging parameters and adaptive inference.

Contribution

It develops a novel Bayesian methodology with adaptive priors for density and correction functions, enabling density estimation without training data and handling high contrast, discontinuous structures.

Findings

Successfully applied to real nano-structure data.

Effective in simulated alloy sample analysis.

Handles high contrast and discontinuities robustly.

Abstract

We present a new Bayesian methodology to learn the unknown material density of a given sample by inverting its two-dimensional images that are taken with a Scanning Electron Microscope. An image results from a sequence of projections of the convolution of the density function with the unknown microscopy correction function that we also learn from the data. We invoke a novel design of experiment, involving imaging at multiple values of the parameter that controls the sub-surface depth from which information about the density structure is carried, to result in the image. Real-life material density functions are characterised by high density contrasts and typically are highly discontinuous, implying that they exhibit correlation structures that do not vary smoothly. In the absence of training data, modelling such correlation structures of real material density functions is not possible. So we discretise the material sample and treat values of the density function at chosen locations inside it as independent and distribution-free parameters. Resolution of the available image dictates the discretisation length of the model; three models pertaining to distinct resolution classes are developed. We develop priors on the material density, such that these priors adapt to the sparsity inherent in the density function. The likelihood is defined in terms of the distance between the convolution of the unknown functions and the image data. The posterior probability density of the unknowns given the data is expressed using the developed priors on the density and priors on the microscopy correction function as elicitated from the Microscopy literature. We achieve posterior samples using an adaptive Metropolis-within-Gibbs inference scheme. The method is applied to learn the material density of a 3-D sample of a real nano-structure and of simulated alloy samples.