Adversarial Signal Denoising with Encoder-Decoder Networks

The presence of noise is common in signal processing independent of the signal type. Deep neural networks have shown good performance in removing signal noise, especially in the image domain. In this work, we consider deep neural networks as a denoising tool where our focus is on one dimensional signals. For that purpose, we introduce an encoder-decoder network architecture to denoise signals, represented by a sequence of measurements. Instead of relying only on the standard reconstruction error to train the encoder-decoder network, we treat the task of signal denoising as distribution alignment between the clean and noisy signals. Then, we propose to train the encoder-decoder with adversarial learning, where the goal is to align the clean and noisy signal latent representation. Unlike standard adversarial learning, we do not have access to the distribution of the clean signal's latent representation in advance. For that reason, we propose a new formulation where both clean and noisy signals pass through the encoder to produce the latent representation. Afterwards, a discriminator neural network has to detect whether the latent representation comes from the clean or noisy signal. At the end of training, aligning the two signal distributions results in removing the noise. In our experiments, we study two signal types with complex noise models. First, we evaluate on electrocardiography and later on motion signal denoising. We show better performance than the related learning-based and non-learning approaches, such as autoencoders, wavenet denoiser, recurrent neural networks and wavelets, demonstrating the benefits of adversarial learning for one dimensional signal denoising.