Abstract
When applying the foreground removal methods to uncover the faintcosmological signal from the epoch of reionization (EoR), the foregroundspectra are assumed to be smooth. However, this assumption can be seriouslyviolated in practice since the unresolved or mis-subtracted foreground sources,which are further complicated by the frequency-dependent beam effects ofinterferometers, will generate significant fluctuations along the frequencydimension. To address this issue, we propose a novel deep-learning-based methodthat uses a 9-layer convolutional denoising autoencoder (CDAE) to separate theEoR signal. After being trained on the SKA images simulated with realistic beameffects, the CDAE achieves excellent performance as the mean correlationcoefficient ($\bar{\rho}$) between the reconstructed and input EoR signalsreaches $0.929 \pm 0.045$. In comparison, the two representative traditionalmethods, namely the polynomial fitting method and the continuous wavelettransform method, both have difficulties in modelling and removing theforeground emission complicated with the beam effects, yielding only$\bar{\rho}_{\text{poly}} = 0.296 \pm 0.121$ and $\bar{\rho}_{\text{cwt}} =0.198 \pm 0.160$, respectively. We conclude that, by hierarchically learningsophisticated features through multiple convolutional layers, the CDAE is apowerful tool that can be used to overcome the complicated beam effects andaccurately separate the EoR signal. Our results also exhibit the greatpotential of deep-learning-based methods in future EoR experiments.