Automatic detection of peculiar galaxies in large datasets of galaxy images

We propose an image analysis unsupervised learning algorithm that can detect peculiar galaxies in datasets of galaxy images. The algorithm first computes a large set of calculated characteristics reflecting different aspects of the visual content, and then weighs them based on the σ of the values computed from the galaxy images. The weighted Euclidean distance of each galaxy image from the median is measured, and the peculiarity of each galaxy is determined based on that distance. Experimental results using irregular galaxy images show that the method can effectively detect peculiar galaxies. Code and data used in the experiments are freely available.

► Autonomous sky surveys produce massive datasets of galaxy images.
► The method automatically detects peculiar galaxies in large galaxy image datasets.
► The method is unsupervised, and can therefore detect uncharacterized celestial objects.
► The method is tested with SDSS, but can be applied to any sky survey.