Laboratoire Hubert Curien, Université Jean Monnet, Saint-Étienne
Date(s) : 15/10/2021 iCal
14 h 30 min - 15 h 00 min
In recent years, deep learning based methods have established a new state of the art in multiple tasks, including denoising. Additionally to their better performances compared to most of conventional algorithms, deep learning based methods are relatively fast once trained. Generally, they also do not require much a priori knowledge, which is advantageous in most cases as information such as the initial signal-to-noise ratio are rarely available before denoising.
In our case, we are dealing with periodic signals. Even though deep learning based methods already provide good results on these signals, they hardly exploit the periodicity. Here, we propose a new method allowing the consideration of periodicity in deep neural networks. Our proposed method allows generalisation to various frequencies, including the ones that were not in the training set. We show the improvements obtained with our methods, and discuss its current limitations.