The radio detection technique, with advantages like inexpensive detector hardware and full year duty cycle, can prove to be a vital player in cosmic-ray detection at the highest energies and can lead us to the discovery of high energy particle accelerators in the universe. However, radio detection has to deal with continuous, irreducible background. The Galactic and thermal backgrounds, which contaminate the radio signal from air showers, lead to a relatively high detection threshold compared to other techniques. For the purpose of reducing the background, we employ a deep learning technique namely, convolutional neural networks (CNN). This technique has already proven to be efficient for radio pulse recognition e.g., in the Tunka-Rex experiment. We train CNNs on the radio signal and background to separate both from each other. The goal is to improve the radio detection threshold on the one hand, and on the other hand, increase the accuracy of the arrival time and amplitude of the radio pulses and consequently improve the reconstruction of the primary cosmic-ray properties. Here we present two different networks: a Classifier, which can be used to distinguish the radio signals from the pure background waveforms, and a Denoiser, which allows us to mitigate the background from the noisy traces and hence recover the underlying radio signal.