Abstract

The classification of galaxies based on their morphology is instrumental for the understanding of galaxy formation and evolution. This, in addition to the ever-growing digital astronomical datasets, has motivated the application of advanced computer vision techniques, such as Deep Learning. However, these models have not been implemented as single pipelines that replicate detection, segmentation and morphological classification of galaxies directly from images, as it would be made by experts. We present the first implementation of an automatic machine learning pipeline for detection, segmentation and morphological classification of galaxies based on the Mask R-CNN Deep Learning architecture. This state-of-the-art model of Instance Segmentation also performs image segmentation at the pixel level, which is a recurrent need in the astronomical community. We achieve Mean Average Precision (mAP) of 0.93 in the morphological classification of Spiral or Elliptical galaxies for a set of 239,639 objects from the Galaxy Zoo sample and JPEG images from the Sloan Digital Sky Survey. As a direct use of segmentation, we test the model for deriving centroids of extended sources, reaching a precision better than 1.0 arcsecond. We also test the network under additive Gaussian noise. We find that the Mask R-CNN network is able to perform with accuracy over 92% for a distribution scale of 76.5 counts. (C) 2020 Elsevier B.V. All rights reserved.