Artifact segmentation using the U-Net architecture for powder X-ray diffraction images

Vong, A.; Yanxon, H.; Roberts, E.; Parraga, H.; Weng, J.; Xu, W.; Ruett, U.; Hexemer, A.; Zwart, P.H.; Schwarz, N.

doi:10.1107/S160057752600754X

Figure 5
An XRD image of Z25 dataset (a) without mask, (b) with GSAS-II algorithm mask and (c) with U-Net mask. There are preferred orientation textures (black arrow, left) that are flagged by GSAS-II, which removes them from the integrated spectrum. The U-Net aggressively masks center pixels, but avoids almost all preferred orientation patterns due to its training. The corresponding integrated 1D patterns are displayed in (d), (e) and (f), respectively. Black arrows indicate peaks that are either partially or fully suppressed by GSAS-II, but are correctly shown by the U-Net, showing how neural networks can encode human preferences.

JOURNAL OF
SYNCHROTRON
RADIATION

ISSN: 1600-5775

Volume 33| Part 5| September 2026|

https://doi.org/10.1107/S160057752600754X

Open

access