"Deep Learning Driven Image Super-Resolution for X-Ray Computed Tomography" 

Tyriek Craigs, MSE PhD Candidate 

Advisors: Prof. Rodney Trice & Prof. Arun Mannodi-Kanakkithodi

WebEx Link

ABSTRACT

In materials science, we often relate material properties to the microstructure. However, as materials become more advanced, as in the case for composites, it is not enough to image in a 2D space. To fully understand the interactions of features and grains of more advanced materials, we need to be able to obtain high-resolution images in a 3D space. X-ray computed tomography (XCT) is a viable non-destructive imaging technique that measures the x-ray attenuation as it passes through the sample and maps those measurements to grayscale values. XCT has the benefit of being able to produce 3D image data rapidly; however, obtaining higher resolutions can be challenging due to the sample's thickness, the required field of view, and the atomic density of its constituents. Techniques such as serial sectioning or synchrotron imaging can also produce 3D images with higher resolution than XCT, but come with their own drawbacks. They are usually some combination of being destructive, hard to get access to, or requiring a great deal of time to complete. With the integration and growth of artificial intelligence-based approaches in materials science over the last decade, techniques such as deep learning-driven image super-resolution have emerged, showing promise as an option to overcome some imaging challenges by taking low-resolution images and reconstructing them into higher resolution images, requiring less time and setup in the image acquisition, thereby enabling a high-throughput workflow.