Adversarial Denoising

Adversarial Learning for Denoising Scientific Director: Vasileios Belagiannis Contact Person(s): Leslie Casas

Abstract

We present the idea of adversarial learning for one-dimensional signal denoising. An adversarial encoder-decoder network is proposed for denoising signals, represented by a sequence of measurements. We rely on the network’s latent representation as input to the discriminator to classify the signal into clean or noisy. Aligning the noisy and clean signal distributions results in removing the signal noise. Unlike the standard GAN training, we propose a new formulation that suits to one-dimensional signal denoising. In the evaluation, we show better performance than the related approaches, such as autoencoders and recurrent neural networks, demonstrating the benefits of our adversarial encoder-decoder network in different signal and noise types.

Pictures

Figure 1: Noisy vs. Clean signal latent space - Before Denoising

Figure 2: Noisy vs. Clean signal latent space - After Denoising

Team

Contact Person(s)

M.Sc. Leslie Casas

Location

Technische Universität München Institut für Informatik / I16 Boltzmannstr. 3 85748 Garching bei München Tel.: +49 89 289-17058 Fax: +49 89 289-17059

Visit our lab at Garching.



internal project page

Please contact Leslie Casas for available student projects within this research project.

Related Publications

2018
	L. Casas, N. Navab, V. Belagiannis Adversarial Signal Denoising with Encoder-Decoder Networks This works can be found at: https://arxiv.org/abs/1812.08555 (bib)