To understand how this works, it helps to know what an image looks like to a computer. While we see a galaxy image as a picture with stars and structure, the computer only sees numbers. A color image is stored as a grid of pixels, and each pixel has three values representing red, green, and blue brightness. So, an image is really just a large table of numbers. The job of the machine learning model is to find patterns in these numbers that usually correspond to certain shapes, such as smooth round light distributions for elliptical galaxies or curved arm patterns for spiral galaxies.

An image recognition model learns in stages. In the early layers, it detects very simple features like edges, bright and dark regions, and lines. In the middle layers, it starts combining these into more complex shapes like curves, bars, and rings. In the deeper layers, it can recognize larger structures, such as spiral arms, central bulges, or smooth elliptical profiles. By stacking many layers together, the model gradually builds an understanding of what makes one type of galaxy look different from another, even though it only ever works with numbers.

In this project, you start by downloading galaxy images from the EFIGI dataset, which is a collection of real, labeled galaxy images used for research and education. These images are placed into a folder on your computer. Then, you organize them into a structure that the learning model understands, separating them into training, validation, and testing sets, and also into folders for each class: elliptical, lenticular, and spiral. The training images are used to teach the model, the validation images are used to check its progress while it is learning, and the test images are used at the end to see how well it performs on images it has never seen before.