Leveraging end-to-end denoisers for denoising periodic signals

Jules Rio, Olivier Alata, Fabien Momey, Christophe Ducottet

Abstract: In this paper, we propose a new framework for denoising 1D periodic signals with deep learning models by exploiting their periodic properties. Our method lies on a transformation of the raw waveform into a grid containing the different periods. Networks used with these data can be simply obtained by leveraging end-to-end fully convolutional denoisers containing only 1D convolutions, by replacing some of their layers by 2D convolutions. Our method also offers the advantage of being able to learn one model for generalizing to a large band of frequencies, including unseen ones, instead of requiring to learn one model per frequency. We also study the generalization of our method to real data.