Building towards a universal neural network to segment large materials science imaging datasets

Tiberiu Stan*, Zachary T. Thompson, Peter W. Voorhees

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contribution

1 Scopus citations

Abstract

Segmentation of large images can be one of the most time-consuming steps in the analysis of materials science datasets. Convolutional neural networks (NNs) have been shown to reduce segmentation time compared to manual techniques, but training a new NN is often required for each dataset. We show that simply combining NN training datasets does not necessarily lead to a NN capable of segmenting multiple types of images. In the present study, we first show that SegNet-based neural networks (NNs) can be trained to accurately segment Al-Zn x-ray computed tomography and Pb-Sn serial sectioning images. Applying the Al-Zn NN to the Pb-Sn test image led to misclassified smudges as dendrites, and misclassified speckles as background. Applying the Pb-Sn NN to the Al-Zn test image was unsuccessful, likely because the Al-Zn dendrites had a higher luminance than the Pb-Sn dendrites. The Mix NN (trained using the combined Al-Zn and Pb-Sn datasets) was better at segmenting the Pb-Sn test image than the Al-Zn test image. This is likely because the Pb-Sn training dataset contained ∼4.5 times as many dendrite pixels as the Al-Zn training dataset, thus the Mix NN was over-tuned to identify Pb-Sn dendrites. Simply combining the training datasets was overall detrimental to NN performance, but assigning different classes to the Al-Zn and Pb-Sn dendrites may lead to enhanced performance in the future. These findings serve as guidelines in the quest to develop a universal NN for segmentation of large materials science datasets.

Original languageEnglish (US)
Title of host publicationDevelopments in X-Ray Tomography XII
EditorsBert Muller, Ge Wang
PublisherSPIE
ISBN (Electronic)9781510629196
DOIs
StatePublished - Jan 1 2019
Event12th SPIE Conference on Developments in X-Ray Tomography 2019 - San Diego, United States
Duration: Aug 13 2019Aug 15 2019

Publication series

NameProceedings of SPIE - The International Society for Optical Engineering
Volume11113
ISSN (Print)0277-786X
ISSN (Electronic)1996-756X

Conference

Conference12th SPIE Conference on Developments in X-Ray Tomography 2019
CountryUnited States
CitySan Diego
Period8/13/198/15/19

Keywords

  • Aluminum - zinc
  • Lead - tin
  • Machine learning
  • Neural network
  • Semantic segmentation
  • Serial sectioning
  • Solidification
  • X-ray computed tomography

ASJC Scopus subject areas

  • Electronic, Optical and Magnetic Materials
  • Condensed Matter Physics
  • Computer Science Applications
  • Applied Mathematics
  • Electrical and Electronic Engineering

Fingerprint Dive into the research topics of 'Building towards a universal neural network to segment large materials science imaging datasets'. Together they form a unique fingerprint.

  • Cite this

    Stan, T., Thompson, Z. T., & Voorhees, P. W. (2019). Building towards a universal neural network to segment large materials science imaging datasets. In B. Muller, & G. Wang (Eds.), Developments in X-Ray Tomography XII [111131G] (Proceedings of SPIE - The International Society for Optical Engineering; Vol. 11113). SPIE. https://doi.org/10.1117/12.2525290