Regensburg 2025 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz
AKPIK 6: AI Methods for Materials Science
AKPIK 6.2: Talk
Thursday, March 20, 2025, 16:45–17:00, H5
Synthetic Data Generation for Enhanced Segmentation of Electron Microscopy Images — •Amir Omidvarnia1, Ali Ghaznavi2, Junbeom Park1, Shibabrata Basak1, and Simone Köcher1 — 1Institute of Energy Technologies - Fundamental Electrochemistry (IET-1), Forschungszentrum Jülich GmbH, 52428 Jülich, Germany — 2Federal Institute for Materials Research and Testing (BAM), 12205 Berlin, Germany
Monitoring structural changes in materials via operando electron microscopy (EM) is crucial for understanding the material's changes under operating conditions, which directly impacts its durability and performance. However, accurate segmentation of features such as cracks or pores in EM time-series data requires extensive labeled datasets, which are challenging to produce due to the labor-intensive, subjective nature of manual annotation. Our research addresses this limitation by exploring different synthetic data generation techniques that can effectively supplement scarce annotated data and improve segmentation outcomes. We evaluate multiple synthetic image generation approaches to enhance the training dataset for U-Net models aimed at segmenting EM time-series data: (1) deep convolutional generative adversarial networks (DCGAN) to produce realistic textures; (2) time-varying image synthesis that mimics the temporal development of features and introduces artificial features to simulate structural imperfections; and (3) standard data augmentation methods. These methods are evaluated in generating training data for EM image segmentation from a limited set of EM images.
Keywords: electron microscopy; machine learning; synthetic image generation; image segmentation