Building Towards a Universal Neural Network to Segment Large Materials Science Imaging Datasets

摘要

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 notnecessarily lead to a NN capable of segmenting multiple types of images. In the present study, we first show thatSegNet-based neural networks (NNs) can be trained to accurately segment Al-Zn x-ray computed tomography and Pb-Snserial sectioning images. Applying the Al-Zn NN to the Pb-Sn test image led to misclassified smudges as dendrites, andmisclassified speckles as background. Applying the Pb-Sn NN to the Al-Zn test image was unsuccessful, likely becausethe Al-Zn dendrites had a higher luminance than the Pb-Sn dendrites. The Mix NN (trained using the combined Al-Znand Pb-Sn datasets) was better at segmenting the Pb-Sn test image than the Al-Zn test image. This is likely because thePb-Sn training dataset contained ~4.5 times as many dendrite pixels as the Al-Zn training dataset, thus the Mix NN wasover-tuned to identify Pb-Sn dendrites. Simply combining the training datasets was overall detrimental to NNperformance, but assigning different classes to the Al-Zn and Pb-Sn dendrites may lead to enhanced performance in thefuture. These findings serve as guidelines in the quest to develop a universal NN for segmentation of large materialsscience datasets.