Applications of deep learning-based denoising methodologies for Scanning Electron Microscope images

Abstract In this paper, we use four types of deep-learning algorithms for denoisingScanning Electron Microscope (SEM) measurement data. Denoising of SEM imagesis an important task since the images often suffer from noise, which can makeit difficult to accurately interpret the data. We also investigate realistic SEMdenoising characteristics using a variety of metrics to assess the quality of denoisedimages. Overall, we find that the trained generative models provide superiordenoising performance and that it is crucial to objectively quantify the performance,just like in the scanning process itself. It is anticipated that the deep-learningbased technique can accelerate image measurements, which can be utilized forvery fast analytical investigations. We also demonstrate that the success of agenerative model may depend on the appropriate assessment of noise characteristicsin the specific image data analysis of interest. Moreover, it is addressed thatdenoising performance can be properly evaluated when a relevant metrics thataligns well with human visual systems. Our implementation is available athttps://github.com/seoleuns/SEMdenoising.