Abstract

Contributed Talk - Splinter EScience   (MW-2235)

Compact Convolutional Neural Networks for Object Detection

Günther Heemann, Henri Cecatka, Dominik Bomans
Astronomisches Institut, Ruhr-Universität Bochum

Modern wide-field surveys share two main problems: classical detection algorithms are saturated with unauthentic sources, while large, deep networks are expensive to deploy at survey scale. We propose compact convolutional neural networks (cCNs) as a middle ground, architectures that are deliberately restricted in parameter count. Such networks are cheap to train, transparent enough to characterise, and efficient enough to be applied at full survey scale with low computing power. A cCN works between classical detection and parametric modelling and provides the morphological decision that neither stage is optimised for. We present a working example in which cCNs are used as a morphological validation stage for the detection of low-surface-brightness galaxies in HSC-SSP images.