Galaxy Morphology Classification

Project Overview

A Machine Learning model built with PyTorch, trained on several datasets to classify galaxies by their morphological shapes (Elliptical, Spiral, & Irregular). Currently the model reaches a training accuracy of 99% and a validational accuracy of around 91%.

I'm having issues with deploying the demo site, so if you'd like to test this out yourself, download the contents, open it in an IDE, and run it locally via python main.py. Make sure you have all dependancies and libraries installed.

Pre-Processing

Preprocessing the selected images to ensure a non-biased training sequence.

• Image Resizing | To ensure consistency across the entire dataset, all images were resized to a fixed 224 x 224.
• Augmentation | Increasing the datasets variability & robustness by adjusting rotation, brightness, etc.
• Normalization | Pixel sizes were normalized to ensure a more stable gradient descent, and faster convergence.

Model Architecture

Utilizing a Convolutional Neural Network model, comprising of,

• Convolutional Layers | To clear up background noise and spatial features.
• Pooling Layers | Further reducing size to only highlight the primary focus of each image (in this case, the galaxy)
• Fully Connected Layers | Classify each data piece based on the extracted features.

Training Process

Trained on around 30 epochs using a training-validation split

• Loss Function | Leveraging categorical cross-entropy, to determine how close the models predictions were.
• Adam Optimizer | An Adam optimizer was to minimize losss for more efficient convergence during training.
• Relevant Metrics | tracking Accuracy and Loss to determine the effeciveness of each Epoch.

Directory Structure

galaxy-morphology-classification/
│
├── dataset/
│   ├── train/                 # training dataset
|     ├── elliptical 
|     ├── irregular
|     └── spiral
│   └── val/                   # validation dataset
|     ├── elliptical 
|     ├── irregular
|     └── spiral
│
├── static/
│   └── background.jpg          # stuff for the demo frontend
│
├── templates/
│   └── index.html              # quick frontend demo page if you wanna test out the model          
│
├── app.py                      # training + flask route to demo
├── main.py                     # cnn model
├── README.md                   # documentation
└── requirements.txt            # packages