|About this Abstract
||Materials Science & Technology 2019
||Data Science for Material Property Interpretation
||Deep Learning and MC-X ray, toward Automatic Sample Segmentation
||Samantha Rudinsky, Yu Yuan, Raynald Gauvin, Mike Marsh, Benjamin Provencher, Nicolas Piché
|On-Site Speaker (Planned)
Image segmentation, the process of labeling pixels according to their composition, is nearly always required to achieve proper quantitative analysis in materials science imaging studies. Deep Learning, the application of multi-layer neural networks, has recently emerged as the state-of-the-art method. Unfortunately, successful networks require extensive time-consuming human-curated segmentation examples that must serve as training data, and the resulting models fail to generalize beyond the highly specific images they were trained to solve. We present here a solution for simulating training data. We have integrated MC X-Ray simulation platform with our Deep Learning engine so that we can generate synthetic backscatter electron (BSE) images of artificial samples, sufficient for training neural networks that can later be used to segment real, and often noisy, BSE images. By scaling up to deeper networks with vastly more training data, our models show promise for fully automatic segmentation across a broader range of materials.
||Definite: At-meeting proceedings