Scanning Tunneling Microscopy (STM) Image Segmentation Using Unsupervised and Few-Shot Learning

Abstract

Scanning tunneling microscopy (STM) is a powerful technique for imaging surfaces with atomic resolution, providing invaluable insights into surface structure and physical and chemical processes occurring on surfaces. A regular task of STM image analysis is detecting and labelling features of interest against the background of the unperturbed surface. Performing this segmentation manually is a labor-intensive task, requiring significant human effort. ​In this paper, we propose an automated approach to the segmentation of STM images that leverages few-shot learning and unsupervised learning to remove the requirement for large manually annotated datasets. Our technique offers greater flexibility compared to previous supervised methods, being easier to adapt to an unseen surface while maintaining high accuracy, reaching up to 90%. We demonstrate the effectiveness of our approach on two distinct surfaces: Si(001), TiO2(110). We show that our model exhibits strong generalization capabilities, adapting well to unseen surfaces with only as little as one additional labeled data point after initial training. This work represents a significant step towards more efficient and adaptable segmentation of STM images.