Scanning transmission electron microscopy image atomic position detection by using deep learning

Abstract

Scanning Transmission Electron Microscopy (STEM) is a powerful imaging technique that offers high spatial resolution and detection sensitivity. However, traditional image processing techniques face limitations when dealing with complex and noisy STEM images. This paper explores the use of deep learning methods to analyze STEM images and proposes AtomID-net, an atomic finding model based on UNet. The model is trained and tested using real STEM data, addressing the limitations of simulated images. Furthermore, a new data labeling method is introduced for quick and accurate annotation of noisy images. The performance of AtomID-net is compared with traditional image processing techniques and other deep learning methods, demonstrating its superior performance and flexibility in detecting atomic positions. This research contributes to the advancement of deep learning in STEM image analysis and provides a reliable evaluation scheme.