Learning to Recorrupt: Noise Distribution Agnostic Self-Supervised Image Denoising

A team of researchers led by Brayan Monroy has introduced a breakthrough in image denoising with their paper "Learning to Recorrupt: Noise Distribution Agnostic Self-Supervised Image Denoising." The new method, called Learning to Recorrupt (L2R), solves a major limitation of previous self-supervised techniques like Noisier2Noise, which required precise prior knowledge of the statistical noise distribution corrupting the images. L2R circumvents this by using a min-max saddle-point objective to train a monotonic neural network that learns the recorruption process directly from the noisy data itself. This makes the model agnostic to the underlying noise, a critical advancement for handling real-world images where the corruption is often unknown or follows complex, heavy-tailed distributions.

The practical impact is significant. L2R has demonstrated state-of-the-art performance across a range of unconventional and difficult noise models that commonly challenge other denoisers. These include log-gamma, Laplace, and spatially correlated noise, as well as signal-dependent models like Poisson-Gaussian noise common in low-light photography and medical imaging. By removing the need for manual noise modeling, L2R automates and generalizes the denoising process, paving the way for more robust computer vision systems in fields from astronomy to smartphone photography, where clean training data is scarce but noisy images are plentiful.