Title: Enhancing TEM Defect Detection and Classification Through Mask-Conditioned Latent Diffusion Augmentation

Abstract:

The scarcity of high-quality, annotated data often hinders the analysis of microstructural defects in transmission electron microscopy (TEM) images, especially within irradiated metal alloys. To overcome this limitation, we propose a generative data augmentation strategy employing a mask-conditioned latent diffusion model (LDM). This approach synthesizes realistic TEM images paired with automatically labeled, multi-class defect masks that can be controlled during generation. By sampling from distributions derived from experimental masks, the method produces synthetic image-mask pairs without the need for manual annotation.

We utilized these generated samples to augment small experimental datasets containing 10, 50, and 100 labeled images. These augmented datasets were then used to train a Mask Regional Convolutional Neural Network (R-CNN) for the dual tasks of defect detection and classification. Our analysis reveals that while generative augmentation leads to modest overall improvements in model performance—achieving a maximum increase of 0.02 in the harmonic mean of detection and classification F1 scores—the specific benefits for detection versus classification vary depending on the train/test data split. These results underscore the capability of specialized generative models to boost deep learning outcomes in microscopy-based image quantification tasks where data is limited.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC