Quantifying physical values from just an image

Three groups of physicists and mathematicians from Massachusetts Institute of Technology, National Securities Technology and Los Alamos National Laboratory recently teamed up to address the question: How to extract numerical values for various physical properties from images in a reproducible way? Led by Leora Dresselhaus-Cooper, a physical chemist from MIT, and Marylesa Howard, a mathematician from NSTec, the collaborating researchers came up with a method that would enable scientists to better quantify images, of which they report in the Journal of Applied Physics.

Although current methods can extract numerical values from images, these require images with high contrast and uniform intensities — characteristics that are hard to achieve. So Dresselhaus-Cooper and Howard led the collective effort in developing a supervised machine learning algorithm that allows physicists to extract information no matter the quality of the image. They began with a traditional segmentation algorithm, but enhanced it so that it could learn about local classes. The final algorithm, known as LADA, works by separating an image based on the intensity of each object in the scene.

As a test run, the team analyzed images of shocked systems. To their surprise, LADA was able to find boundaries between features that were difficult to see by eye. And that gave a different result than expected. While conventional knowledge suggests converging shock waves decelerate after passing through the center, the image shows that shock waves accelerated most upon divergence.

While this study analyzed shock physics, the team is hopeful that the method can be applied to a wide range of different fields.

Source: “Machine learning to analyze images of shocked materials for precise and accurate measurements,” by Leora Dresselhaus-Cooper, Marylesa Howard, Margaret C. Hock, B. T. Meehan, Kyle J. Ramos, Cindy A. Bolme, Richard L. Sandberg, and Keith A. Nelson, Journal of Applied Physics (2017). The article can be accessed at https://doi.org/10.1063/1.4998959 .