Deep transfer learning for blended source identification in galaxy survey data

We present B LEND H UNTER , a proof-of-concept deep-transfer-learning-based approach for the automated and robust identification of blended sources in galaxy survey data. take VGG-16 network with pre-trained convolutional layers train fully connected on parametric models COSMOS images. test efficacy transfer learning by taking weights learned using them to identify blends more realistic Canada-France Imaging Survey (CFIS)-like compare performance this method SEP (a Python implementation SE XTRACTOR ) as function noise levels separation between sources. find that outperforms ∼15% terms classification accuracy close (< 10 pixel sources) regardless level used training. Additionally, provides consistent results distant (≥10 provided is trained data has relatively standard deviation target The code have been made publicly available ensure reproducibility results.