DOI: 10.11118/ecos2026_033

CLASSIFYING SCANNING ELECTRON MICROSCOPE DEFECTS WITH A CNN AND A CUSTOM FOURIER-TRANSFORM LAYER

Magdaléna Skřičilová¹, Oldřich Trenz¹

¹ Department of Informatics, Faculty of Business and Economics, Mendel University in Brno, Zemědělská 1, 613 00 Brno, Czech Republic

A scanning electron microscope has to be checked periodically, and two of its most common image faults, defocus and astigmatism, are tedious to judge by eye. This article presents a convolutional neural network that sorts pairs of microscope images into four focus and astigmatism classes so the check can run without an operator. Each sample is two shots of the same spot: one in the base position and one with a shifted working distance, because the shifted shot is what makes astigmatism visible. The decisive ingredient is a custom Fourier-transform preprocessing layer that converts each input to its magnitude spectrum before the convolutional part sees it. On a dataset of about three thousand image pairs, the base model variants with-out the Fourier layer reached 49–65% test accuracy, while the corresponding variants with the Fourier layer reached 89–94%; the best result was 94% with channel-axis concatenation. The same shallow model beat three pretrained ImageNet networks (MobileNetV2 87%, EfficientNetV2L 52%, ResNet50 44%), which we did not expect. The best fusion of the two images was concatenation along the channel axis together with the Fourier layer.

Keywords: convolutional neural network, scanning electron microscope, Fourier transform, image classification, astigmatism

pages: 308-315, online: 2026