Rotation Invariant Feature Extraction in the Classification of Galaxy Morphologies

Keywords
Loading...
Thumbnail Image
Issue Date
2014-07-08
Language
en
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The Galaxy Zoo project is a crowdsourcing platform to classify the morphology of galaxies into different categories. Recently, the project set out a challenge to automatically predict these crowdsourced classifications using machine learning techniques. In this thesis, one of these machine learning techniques is explored and modified. This technique, designed by Coates et al. [4], works by learning features in an unsupervised manner using k-means. These features are rotation sensitive, but since there is no up or down in space, the system would ideally work rotation invariantly. Therefore, the system was modified to account for rotation sensitivity in an efficient manner by changing the distance metric that is used by the k-means algorithm. Results show that this improves the performance significantly by more than 5%: the regular method achieves a root mean squared error of 0.10789 while the rotation invariant method achieves a score of 0.10256.
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen