Image Processing Using Machine Learning
Many modern materials characterization tools output complex imaging datasets that must be quantitatively analyzed to extract useful materials information. Segmentation (the act of partitioning images into multiple parts) can be one of the most subjective and time-consuming steps in the analysis of large datasets. We show that it is possible to quickly segment both x-ray computed tomography and serial sectioning datasets using a variety of 2D and 3D machine learning algorithms. In this project we address the challenge of obtaining enough images for convolutional neural network (CNN) training using both manually-created segmentations as well as computationally-generated phase field images. The number of phase field images needed for CNN training and the most important microstructural features required for CNNs to “understand” the contents of the datasets are discussed. The ability to quickly and quantitatively analyze complex microstructural images will accelerate the rate of materials development, design, and discovery.
Participants
Selected Publications
J. Yeom, T. Stan, S. Hong, P. Voorhees, “Segmentation of Experimental Datasets via Convolutional Neural Networks Trained on Phase Field Simulations” doi: 10.1016/j.actamat.2021.116990
T. Stan, Z. Thompson, P. Voorhees, “Optimizing Convolutional Neural Networks to Perform Semantic Segmentation on Large Materials Imaging Datasets: X-ray Tomography and Serial Sectioning” Materials Characterization, 160 (2020) doi: 10.1016/j.matchar.2020.110119
T. Stan, Z. Thompson, P. Voorhees, “Building Towards a Universal Neural Network to Segment Large Materials Science Imaging Datasets,” Proc. SPIE 11113, Developments in X-ray Tomography XII, 111131G (2019) doi: 10.1117/12.2525290
![Machine Learning Image](https://sites.northwestern.edu/voorhees/files/2021/07/Machine-Learning-Image.png)