This paper introduces a framework (EstimateNoiseSEM) to automate noise estimation in scanning electron microscopy (SEM) images. Within this framework, a classification network selection mechanism facilitates the choice of a more optimized classification approach. Consequently, the classification stage determines the image's noise type, while the regression model predicts the corresponding noise level. Noise estimation, which includes the noise type and level, is necessary to perform denoising in most cases. This study targeted the noise in scanning electron microscopy (SEM) images. Indeed, depending on the dwell time, the SEM produces different types of noise (Gaussian or Gamma) that can pose uncertainty problems during denoising. That is why, the multi-stage scheme based on deep learning was proposed in this study. The proposed approach performed better in Gaussian noise classification with more than 80% Accuracy, Precision, Recall, and F1-score on synthetic noisy samples and 0.98+/-0.01 root squared error in Gaussian noise classification. The classification network once achieved 97% of accuracy for Gaussian noise classification which decreased to 80% later on because of the uncertainty of Gamma noise levels. However, this study also provides detailed insights into the Gamma noise estimation process. These insights may guide us or the community in developing deep learning-based Gamma noise estimation techniques.
Keywords: Deep learning; Noise estimation; Scanning electron microscopy images.
Copyright © 2025. Published by Elsevier B.V.