In this project, we implement a denoising autoencoder trained only with noisy images, inspired by the paper Noise2Noise: Learning Image Restoration without Clean Data.

• In Miniproject 1, we implement such decoder using PyTorch, reaching performances similar to the original architecture but with 20x less parameters. Our experiments and findings are summarized in Report 1.
• In Miniproject 2, we implement such decoder from scratch, writing all the building blocks on our own. Our implementation choices and experiments are summarized in Report 2.

This work is part of the course "Deep Learning" at EPFL, and is the joint effort of Francesco Salvi, Lukas Van Der Huevel and Jirka Lhotka.