On an Application of Denoising to the Uncertainty Quantification of Line Edge Roughness Estimation

Abstract

Prediction intervals which describe the reliability of the predictive performance of regression models are useful to influence decision making and to build trust in machine learning. Normalized conformal prediction is a rigorous and simple guideline to construct prediction intervals which has no distributional assumptions but requires other types of modeling to assess a regression model fit to training data, and quantile regression is a widely used technique in other fields to construct prediction intervals. We propose image denoising and other image processing techniques as a foundation to prediction interval construction procedures for line edge roughness (LER) estimation from noisy scanning electron microscope (SEM) images and show that these innovations offer significant improvements in efficiency over earlier approaches used to study the deep convolutional neural network EDGENet.