Galaxy Classification using Transfer Learning

Date:

  • The shape of a galaxy provides insights about its evolution and is an important aspect in astronomical research.
  • Novel deep learning classifiers were proposed and trained to categorize galaxies into 10 classes using the galaxy images captured by space exploration missions.
  • Transfer learning was used as opposed to the earlier works that trained networks from scratch.
  • VGG-16, ResNet-50, XceptionNet, InceptionNet, and DenseNet-121 architectures were used to create 5 separate feature extraction models.
  • The VGG-16 based model turned out to be the best one outperforming a state of the art model in 8 of the 10 classes giving only 2% lower overall accuracy.
  • The resource consumption in training the model was brought down by 60% and the number of models trained before finding the one that gives the best performance were reduced from several 100s to 10.