Self-supervised Learning in Image Denoising with GR2R and L2R
Presenter: Brayan Monroy (UIS, Colombia)
Abstract
Learning to denoise images without clean reference data is essential in scientific imaging, where ground truth is often unavailable. This talk presents two complementary frameworks that train denoisers directly from noisy observations: Generalized Recorrupted-to-Recorrupted (GR2R) and Learning to Recorrupt (L2R).
GR2R extends recorruption-based self-supervision beyond Gaussian noise to broader additive noise models and the natural exponential family, including Poisson and Gamma distributions. By constructing training pairs from individual noisy observations, it yields objectives equivalent in expectation to supervised losses, up to model-independent constants, under appropriate conditions. Its connection to Stein’s unbiased risk estimation provides a theoretical foundation without requiring explicit divergence computation.
L2R addresses a central limitation of this approach: the need for prior knowledge of the noise distribution. It introduces a learnable recorruptor jointly optimized with the denoiser through a min–max objective, adapting auxiliary noise statistics directly from noisy data. Experiments demonstrate its effectiveness for non-Gaussian, heavy-tailed, and spatially correlated noise, alongside competitive performance on real imaging benchmarks with signal-dependent noise.
The talk will discuss the theoretical principles, practical implementation, and experimental evidence underlying this progression from known to learned recorruption.
About the Speaker
Brayan Monroy is a doctoral researcher in Computer Science at Universidad Industrial de Santander (UIS), Colombia, and a visiting researcher at GIPSA-lab.
His research combines mathematical optimization and deep learning to address imaging inverse problems, with a particular focus on self-supervised methods for linear and nonlinear forward operators. His work includes self-supervised image denoising beyond Gaussian noise, compressive spectral imaging, and deep-learning approaches for chronic wound assessment. He holds B.Sc. and M.Sc. degrees in Systems Engineering from UIS and contributes to DeepInverse, an open-source Python library for solving imaging inverse problems using deep learning.