Learning Image Restoration Without Clean Data

We employ machine learning techniques, specifically neural networks, to perform signal reconstruction by learning to map corrupted or degraded observations to their corresponding clean signals. The key idea is to train a neural network model to estimate the underlying clean signal from its corrupted version, leveraging the ability of deep networks to learn complex mapping functions from data. Our approach leads to a simple yet powerful conclusion: it is possible to learn how to restore images by only looking at corrupted examples, achieving equal performance and sometimes surpassing training on clean data, without requiring specific image prior or probability models of corruption. In practice, we demonstrate that a single model can learn photographic noise removal, denoising of synthetic Monte Carlo images, and reconstruction of under sampled MRI scans - all corrupted by different processes - based solely on noisy data. The model learns these diverse image restoration tasks from corrupted observations alone, without needing clean training examples or explicit corruption models.