Superresolution (SR) is a type of imaging technique that attempts to apply an apparent increase in resolution to some image. In nature, all imaging instruments have an inherent limit on the resolution that they can capture and image with - often referred to as the diffraction limit. However there may be other factors in an imaging situation that encumbers the resolution further (any number of weather-based factors or cosmic effects can cause this in regards to astronomy). Thus, a way to increase the quality of an image can be very useful.

SR in Medical Imaging

Recently, a team of data scientists created an open-source framework designed to make the use of Machine Learning in medical imaging much simpler and more streamlined. They called this software the Deep Learning Toolkit (DLTK). The DLTK uses Tensorflow and is effectively a library of useful functions that are typical inclusions in many ML script. The DLTK comes with a few examples and tutorials which demonstrate how the software is used. For our purposes,  we are interested in the software's ability to implement SR. Luckily, we are provided with an example that does just that, so we decided to run the example ourselves on Pleiades. We had to perform a few tweaks in order to get the code to work to our desire, which is documented here. We ran the code for 100 000 Epochs which took roughly 12 hours. The following are some examples showing the original image, followed by what the image looks like downscaled (effectively the model input), a linear upscaling of the low resolution image and finally the resulting model output.