🌌 Galaxy Morphology Classification

A machine learning project for automatic classification of galaxy morphologies from images, developed as part of the Machine Learning course (MSc, University of Pavia, 2025).

Two classification approaches are implemented and compared: a Logistic Regression model with handcrafted features and a Convolutional Neural Network based on transfer learning with ResNet50.

---

Problem Definition

Galaxies come in four morphological categories:

Class	Description
Spiral	Disc-shaped with spiral arms
Barred Spiral	Spiral with a central bar structure
Elliptical	Smooth, featureless ellipsoidal shape
Irregular	No defined structure

Dataset: 11,000 images (224×224 px, RGB), balanced across 4 classes — 10,000 training / 1,000 test.

Special challenges addressed:

• How much does color information help classification?
• How to exploit rotation invariance (galaxies have no preferred orientation)?

---

Approaches

1. Logistic Regression with Feature Engineering (Galaxies_classificationLR_HOKJA.ipynb)

A linear classifier trained on handcrafted image features:

• Grayscale histogram (64 bins) — intensity distribution
• Edge direction histogram (Sobel, 64 bins) — structural features
• Co-occurrence matrix (64 features) — texture patterns
• Color histogram (RGB, 64 bins) — color information

Multiple feature combinations are benchmarked to isolate the contribution of each feature type.

Key findings:

• Best configuration: Shape + Texture features (Shape_Texture)
• Color adds limited benefit for morphological classification
• Dataset is balanced → accuracy is a reliable metric

2. CNN with Transfer Learning (Galaxies_classificationCNN_HOKJA.ipynb)

Fine-tuning of a ResNet50 pre-trained on ImageNet, with the final fully-connected layer replaced for 4-class classification.

Three experimental configurations:

Configuration	Input	Augmentation
RGB	3-channel color	Standard
Grayscale (baseline)	1-channel	Standard
Gray + RotInv	1-channel	Rotation-invariant augmentation

Rotation-invariant augmentation (RandomRotation(360°)) exploits the physical symmetry of galaxy images to improve generalisation.

---

Repository Structure

galaxies-classification/
│
├── Galaxies_classificationLR_HOKJA.ipynb   # Logistic Regression notebook
├── Galaxies_classificationCNN_HOKJA.ipynb  # CNN / ResNet50 notebook
│
├── galaxies/
│   ├── train/
│   │   ├── spiral/
│   │   ├── barred/
│   │   ├── elliptical/
│   │   └── irregular/
│   └── test/
│       ├── spiral/
│       ├── barred/
│       ├── elliptical/
│       └── irregular/
│
└── Galaxies_classification.pdf   # Assignment specification

---

Requirements

pip install torch torchvision numpy matplotlib opencv-python scikit-learn gdown tqdm

Notebooks are designed to run on Google Colab with GPU acceleration. The dataset is downloaded automatically via gdown from Google Drive.

---

Usage

Open the notebooks in Google Colab:

Logistic Regression:

Galaxies_classificationLR_HOKJA.ipynb

Run all cells sequentially. The notebook covers data exploration, feature extraction, model training, confusion matrix analysis, and per-class error inspection.

CNN (ResNet50):

Galaxies_classificationCNN_HOKJA.ipynb

Run all cells sequentially. Three experiments are run: RGB, Grayscale, and Grayscale + rotation-invariant augmentation. Confusion matrices are generated for each.

---

Results Summary

Model	Configuration	Test Accuracy
Logistic Regression	Shape + Texture	see notebook
ResNet50	RGB	see notebook
ResNet50	Grayscale	see notebook
ResNet50	Gray + RotInv	see notebook

Full metrics and confusion matrices are available in the notebooks.

---

References

• He et al. — Deep Residual Learning for Image Recognition (ResNet, 2016)
• Galaxy Zoo project — www.galaxyzoo.org

---

BSc Physics — Machine Learning Course, University of Pavia, 2025
Author: Kristina Hokja