1D Piecewise-Constant Functions

This page contains code to reproduce the computational experiments in the tutorial Learn, Denoise and Discover: A Guide to Deep Denoising with an Application to Electron Microscopy , involving networks trained to denoise 1D piecewise-constant signals. The code is in the form of Jupyter notebooks that can be run in Google colab. The full code repository is available on GitHub.

Simple example

Code to train and apply a neural network that denoises 1D piecewise-constant signals corrupted by additive Gaussian noise.

Network design

This notebook illustrates the effect of two different design choices: the presence of a skip connection and the number of layers in the network. It also includes a comparison with linear shift-invariant denoising and total-variation regularization.

Supervised vs unsupervised learning

This notebook compares neural networks trained via supervised and unsupervised learning.

Interpretability

This notebook visualizes the denoising strategy learned by a neural network for specific inputs, by visualizing equivalent filters that are equal to the network gradient with respect to its input.