Presentation #301.02 in the session Machine Learning Applications.
Galaxy model subtraction is often needed to identify and measure the photometry of celestial objects projected close to the centers of nearby galaxies, such as star clusters or gravitationally lensed background sources. In a large dataset, the traditional methods of fitting and subtracting galaxy models in sky images can be laborious and time-consuming. Also, the results produced on non-elliptical galaxies are sometimes inaccurate. We present a machine-learning approach to perform galaxy modeling with an image denoising autoencoder. An autoencoder will first extract features by encoding the image into low-dimensional latent space, and then reconstruct the image based on its latent representation. We used convolutional neural networks to build the encoder and decoder in this model and trained it on model galaxies generated by GALFIT. We also implemented techniques such as binning and masking to enhance the performance of the model. We tested our model on real galaxy images from the Next Generation Virgo Cluster Survey (NGVS) and found that it can produce results similar to traditional methods on elliptical galaxies and outperform traditional methods on some non-elliptical galaxies. The model is fast and highly automated and therefore can be very useful when processing a large amount of collected image data in the future.