This project was put together as a part of the McGill University Physics Hackathon in Montreal in 2018. The problem we were trying to solve was that of galaxy classification, specifically if an image of a galaxy is that of a spiral, elliptical, or an irregular galaxy.
Our approach first performed some pre-processing on the images to highlight the potential features for the model to use (in an effort to create a better performing model) as well as remove potential rotational bias from the dataset. To do this, we first detected the edge of the galaxy and applied a mask to remove all background information. We then proceeded to sharpen the remainder of the image in order to make features easier to pick up on. Additionally, inside our data we duplicated the image four times with the following transformations: - No transformation - 45o rotation - Reflection across central vertical axis - 45o rotation and Reflection across central vertical axis This was done in order to help keep the model from deciding on classification due to orientation of the galaxy. Finally, we trained a neural net on this dataset using keras.