Denoising Particle Beam Micrographs with Plug-and-Play Methods

TL;DR

This paper develops plug-and-play denoising methods tailored for particle beam micrographs, effectively reducing noise by exploiting image structure and handling complex data likelihoods, with significant improvements demonstrated in simulations.

Contribution

It introduces novel denoising algorithms that incorporate the unique statistical properties of particle beam micrograph data within the plug-and-play framework.

Findings

RMSE reduced by factors of 2.24 to 4.11

Significant improvements in SSIM and visual quality

Applicable to both conventional and time-resolved measurements

Abstract

In a particle beam microscope, a raster-scanned focused beam of particles interacts with a sample to generate a secondary electron (SE) signal pixel by pixel. Conventionally formed micrographs are noisy because of limitations on acquisition time and dose. Recent work has shown that estimation methods applicable to a time-resolved measurement paradigm can greatly reduce noise, but these methods apply pixel by pixel without exploiting image structure. Raw SE count data can be modeled with a compound Poisson (Neyman Type A) likelihood, which implies data variance that is signal-dependent and greater than the variation in the underlying particle-sample interaction. These statistical properties make methods that assume additive white Gaussian noise ineffective. This paper introduces methods for particle beam micrograph denoising that use the plug-and-play framework to exploit image structure while being applicable to the unusual data likelihoods of this modality. Approximations of the data likelihood that vary in accuracy and computational complexity are combined with denoising by total variation regularization, BM3D, and DnCNN. Methods are provided for both conventional and time-resolved measurements, assuming SE counts are available. In simulations representative of helium ion microscopy and scanning electron microscopy, significant improvements in root mean-squared error (RMSE), structural similarity index measure (SSIM), and qualitative appearance are obtained. Average reductions in RMSE are by factors ranging from 2.24 to 4.11.