1. Automated Classification of Galaxy Images
INSTITUTO NACIONAL DE ASTROFÍSICA, ÓPTICA Y ELECTRÓNICA
Automated Classification of Galaxy Images
Jorge de la Calleja and Olac Fuentes
Instituto Nacional de Astrofísica Óptica y Electrónica
Tonantzintla, Puebla, Mexico

2. Introduction A galaxy contains millions of stars.
There are several billion galaxies in the universe.
Galaxy classification can provide important clues about the origin and evolution of the universe.
Traditionally, galaxy classification is perfomed by visual inspection of images by expert astronomers
Current Sky Surveys make manual classification unfeasible.

3. Introduction

4. Hubble Classification Scheme
Introduction
Hubble Classification Scheme

5. Hubble Classification Scheme: Three Main Classes
Introduction
Hubble Classification Scheme: Three Main Classes
Spiral Galaxies
Eliptical Galaxies
Irregular Galaxies
Classes can in turn be divided into subclasses.

6. Introduction Previous Work: Semi-automated Schemes: Problems:
Combine manually-extracted features with features obtained using image-analysis techniques
Use these features to train a neural network.
Problems:
Need for human intervention
Feature extraction is not robust
Accuracy much lower than that obtained by humans.

7. Completely Automated Classification
Proposed Method
Completely Automated Classification

8. Proposed Method
Image Standardization: Rotation, Centering, Cropping and Resizing
Data Compression: Principal Component Analysis (PCA)
Machine Learning: Naive Bayes, C4.5 and Random Forest

9. Proposed Method Image Standardization
Rotation
Extract binary image from original image using a threshold
Find main axis of image using PCA
Rotate image so that the main axis is horizontal
Centering
Translate galaxy to center of image
Cropping
Eliminate rows and/or columns that don’t contain pixels that belong to the galaxy image
Resizing
Convert image to standard (128X128) size

10. Proposed Method Image Standardization

11. Proposed Method Image Standardization
Examples:

12. Proposed Method Data Compression
From set of images perform PCA to obtain “eigengalaxies” (analogous to eigenfaces [Turk and Pentland, 91]).
Project images onto the first few eigengalaxies.
Use projection as attributes for the Machine Learning algorithm.

13. Proposed Method Data Compression
Original
Images
Reconstructed images using first 25 PCs

14. Proposed Method Machine Learning
Perform experiments using the WEKA implementation of:
Naive Bayes Classifier
C4.5
Random Forest

15. Data and Experiments
We used 292 images of galaxies downloaded from the SEDS database.
8, 13 y 25 principal components were used.
We report averages of 5 runs using 10-fold cross-validation.

16. Results
Accuracies

17. Conclussions
The proposed method yields better results that any other automated method reported in the literature (still not as good as those obtained by experts).
Random Forest obtained the highest accuracy among learning algorithms.
A small number of PCs is enough to obtain good accuracy.
Image standardization results in an improvement in accuracy.

18. Future Work
Extend method to a larger variety of astronomical objects (nebulae, star clusters, peculiar objects).
Work on wide-field images and large sky surveys, such as the Sloan Digital Sky Survey (SDSS).
Experiment with more sophisticated image analysis techniques and learning algorithms to improve accuracy.

19. For more information, see ccc.inaoep.mx/~fuentes/
Thanks! Questions?
For more information, see
ccc.inaoep.mx/~fuentes/