Recovering data from high amounts of loss and corruption would be useful for a wide variety of civilian and military applications. Highly corrupted data (e.g., speech and images) has been less studied relative to the problem of light corruption, but would be advantageous for applications such as low-light imagery and weak signal reception in acoustic sensing and radio communication. Unlike milder signal corruptions, resolving strong noise interference may require a more robust approach than simply removing predictable noise, namely actively looking for the expected signal, a type of problem well suited for machine learning. In this work, we evaluate a variant of the U-net autoencoder neural network topology for accomplishing the difficult task of denoising highly corrupted images and English speech when noise floors are 2-10x stronger than the clean signal. We test our methods on corruptions including additive white Gaussian noise and channel dropout.
|