Deep Learning Based Detection of Cosmological Diffuse Radio Sources

In this paper we introduce a reliable, fully automated and fast algorithm to detect extended extragalactic radio sources (cluster of galaxies, filaments) in existing and forthcoming surveys (like LOFAR and SKA). The proposed solution is based on the adoption of a Deep Learning approach, more specifically a Convolutional Neural Network, that proved to perform outstandingly in the processing, recognition and classification of images. The challenge, in the case of radio interferometric data, is the presence of noise and the lack of a sufficiently large number of labeled images for the training. We have specifically addressed these problems and the resulting software, COSMODEEP proved to be an accurate, efficient and effective solution for detecting very faint sources in the simulated radio images. We present the comparison with standard source finding techniques, and discuss advantages and limitations of our new approach.