研究目的
To develop and validate an objective and quantitative method for detecting atom columns and single atoms from HAADF STEM images, particularly for beam-sensitive nanomaterials with low contrast-to-noise ratios, using a Bayesian framework that combines statistical parameter estimation and model-order selection.
研究成果
The MAP probability rule provides an effective and objective method for atom detection in HAADF STEM images, outperforming other model-selection criteria like AIC and BIC, especially for low CNR images. It allows for flexible incorporation of prior knowledge and introduces the integrated CNR (ICNR) as a robust image quality measure. The method is particularly useful for beam-sensitive nanomaterials, enabling reliable structure quantification without visual bias.
研究不足
The method relies on approximations, such as the normality assumption for Poisson noise, which may not hold for very low electron counts. It is most accurate for high electron doses. The performance depends on the choice of prior distributions and parameter ranges, which could introduce bias if not chosen carefully. The simulations assume specific models (e.g., Gaussian shapes), which may not capture all real-world variations.
1：Experimental Design and Method Selection:
The study uses a Bayesian approach with the maximum a posteriori (MAP) probability rule for model-order selection. It involves deriving analytical expressions for the MAP probability and comparing it with other criteria like AIC and BIC. Simulations are performed to validate the method.
2：Sample Selection and Data Sources:
Simulated HAADF STEM images of various nanomaterials (e.g., graphene, Au atoms) are generated using the MULTEM software, with parameters such as incoming electron dose, pixel size, and background varied to control image quality measures like SNR and CNR.
3：List of Experimental Equipment and Materials:
The primary tool is the MULTEM software for simulating STEM images. No physical equipment is mentioned; the focus is on computational simulations.
4：Experimental Procedures and Operational Workflow:
Images are simulated with Poisson noise. The MAP probability rule is applied by fitting parametric models (e.g., Gaussian peaks) to the images, optimizing parameters, and computing posterior probabilities for different numbers of atomic columns. The process includes adding peaks iteratively and comparing probabilities.
5：Data Analysis Methods:
Statistical analysis involves calculating detection rates, confidence intervals, and comparing performance of different model-selection criteria. Parameters are estimated using maximum likelihood methods, and results are visualized through plots and tables.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
