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

arXiv:2603.25869 · eess.IV, cs.CV, stat.ML · Submitted 2026-03-26 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Learning to Recorrupt".

Tom: Learning to Recorrupt (L2R) is a noise distribution-agnostic self-supervised image denoising framework designed to eliminate the need for prior knowledge of noise statistics,

Jane: First, who's behind it and why it matters.

Title and authors: Tom: So we’re diving into "Learning to Recorrupt: Noise Distribution Agnostic Self-Supervised Image Denoising" today. It sounds like they’re tackling a really tough problem in image denoising by trying to get around the need to know the exact noise statistics beforehand.

Jane: That's right, Tom; it’s about moving away from methods that demand specific knowledge of the noise distribution, which is a huge hurdle in real-world sensing where we often don't have perfect information.

Lu: The authors are Monroy, Bacca, and Tachella, and they’re proposing a novel framework that uses a learnable monotonic neural network to model the corruption process through a specific optimization setup.

Meng: A learnable model for corruption sounds interesting from an engineering standpoint; it suggests we can handle noise that we haven't even characterized yet.

Lalam: I think this paper has the potential to fundamentally change how we approach AI systems in fields like medical imaging, because it removes one of the biggest bottlenecks—the need for perfect statistical characterization of noise.

The paper's summary: Tom: What they are saying is that instead of fixing the noise model, L2R lets a neural network learn how to simulate that corruption process using a min-max saddle-point objective. It’s essentially treating the recorruption mechanism as something the system learns directly.

Jane: So, in simple terms, they're not telling the denoiser what the noise is; instead, they let it figure out how to "recorrupt" an image in a way that helps it learn to clean it up.

Lu: They introduce a learnable noise mapping, which they constrain to be monotonic. This constraint is really important because it forces the learned corruption process to have certain mathematical properties that make the model more stable and robust across different noise types.

Meng: The objective function itself involves minimizing the reconstruction error while simultaneously trying to suppress any correlation between the image and the noise without needing explicit likelihood modeling. That’s a very clever way to frame it.

Lalam: From an AI culture perspective, this moves us toward systems that are inherently more adaptable; instead of being brittle when faced with a new type of noise, these models can adapt because they learn the corruption behavior itself.

The paper's improvements: Tom: The core improvement is this learning mechanism guided by the min-max saddle-point objective, which allows it to work across unconventional and heavy-tailed noise distributions that other methods struggle with.

Jane: They showed it works really well on things like log-gamma and Laplace noise, achieving PSNR scores up to thirty-two point four zero dB on DIV2K for the log-gamma setting, which is quite high.

Lu: They’ve established a theoretical equivalence showing that when the true noise map fits within their admissible class, the self-supervised problem actually tends toward its supervised counterpart, which connects it to other related methods like UNSURE.

Meng: The constraint on the recorruptor network being monotonic is a key design choice that ensures they get better expressivity and robustness when dealing with more complex noise structures.

Lalam: This monotonicity constraint is powerful because it gives the AI model a structural backbone that prevents it from learning overly chaotic or nonsensical corruption patterns when facing unknown data.

Conclusion: Tom: So, to wrap up, "Learning to Recorrupt: Noise Distribution Agnostic Self-Supervised Image Denoising" shows us a framework where we don't need prior knowledge of the noise distribution; we just need a learnable mechanism for simulating that corruption via a min-max objective.

Jane: The major implication is that we can achieve high performance across very different noise regimes, from log-gamma to correlated noise, without manually tuning parameters for each one.

Lu: The theoretical link to supervised learning when the noise map is in their admissible class provides a solid foundation for understanding how this self-supervised objective relates to more established estimation techniques like SURE.

Meng: Practically speaking, this means we can deploy these denoising systems in areas like remote sensing or medical imaging where the noise characteristics are constantly changing or entirely unknown.

Lalam: I think the biggest impact is on building more resilient AI; by learning to model corruption dynamically, we create models that are far less sensitive to distribution shifts in deployment environments.

Brayan Monroy, Jorge Bacca, Julian Tachella

Universidad Industrial de Santander · GIPSA-Lab, Universit´e Grenoble Alpes · CNRS, ENS de Lyon

eess.IV, cs.CV, stat.ML

Submitted: 2026-03-26

Updated: 2026-09-28

Code: https://github.com/bemc22/Learning2Recorrupt

Importance score: 78/100

The gist: Learning to Recorrupt (L2R) is a noise distribution-agnostic self-supervised image denoising framework designed to eliminate the need for prior knowledge of noise statistics, addressing a critical

Key concepts

Noise Distribution Agnostic: This refers to a framework that does not require prior knowledge of the exact statistics of the noise. Instead of fixing a specific noise model beforehand, the system learns how to simulate corruption directly through its learning process.
Learnable Monotonic Neural Network: The authors use a learnable neural network constrained to be monotonic to model the corruption process. This constraint is important because it ensures the learned corruption process has stable mathematical properties, improving robustness against different noise types.
Min-Max Saddle-Point Objective: This is the optimization setup used by L2R. It involves minimizing reconstruction error while simultaneously trying to suppress any correlation between the image and noise without needing explicit likelihood modeling. This objective allows the network to learn how to 'recorrupt' an image.
Monotonic Constraint: This constraint on the recorruptor network is a key design choice. It forces the learned corruption process to have certain mathematical properties, which gives the AI model a structural backbone that prevents it from learning overly chaotic or nonsensical corruption patterns.

Terminology

Robustness — 11× in this paper · explained in 5 episodes
Ablation — 5× in this paper · explained in 1 episode
Self-Supervised Learning — 5× in this paper · explained in 7 episodes
Generalization — 4× in this paper · explained in 13 episodes
Core — 3× in this paper · explained in 2 episodes
Inverse Problem — 3× in this paper · explained in 2 episodes
Loss Function — 3× in this paper · explained in 2 episodes
Distribution Shift — 2× in this paper · explained in 5 episodes
Generative Model — 2× in this paper · explained in 2 episodes
Generative Models — 2× in this paper · explained in 2 episodes
Objective Function — 2× in this paper · explained in 2 episodes
Regularization — 2× in this paper · explained in 3 episodes
Benchmarking — 1× in this paper · explained in 3 episodes
Computational overhead — 1× in this paper · explained in 2 episodes
Consistency — 1× in this paper · explained in 3 episodes
Fidelity — 1× in this paper · explained in 2 episodes
Normalizing Flow — 1× in this paper · explained in 3 episodes
Normalizing Flows — 1× in this paper · explained in 4 episodes
Reconstruction Error — 1× in this paper · explained in 2 episodes

Summary

Learning to Recorrupt (L2R) is a noise distribution-agnostic self-supervised image denoising framework designed to eliminate the need for prior knowledge of noise statistics, addressing a critical limitation in existing methods that rely on precise knowledge of the noise distribution. This method introduces a learnable monotonic neural network that learns the recorruption process through a min–max saddle-point objective, achieving state-of-the-art performance across unconventional and heavy-tailed noise distributions.

The Core Objective: Min–Max Saddle Point

L2R casts denoising as a min–max saddle-point problem in which the recorruption mechanism is treated as a learnable model: a recorruptor designed to simulate the synthetic recorruption process. This optimization ensures that the denoiser learns to suppress the correlation between the image and the noise without requiring explicit likelihood modeling or strong prior knowledge of the noise family. The objective function is formalized as:

min f max h Ey[LL2R(y; f, h)], where LL2R(y; f, h):= Ew′∥f(y + τh(w′)) − y∥2 + 2τf(y + τh(w′))⊤h(w').

The Learnable Recorruptor and Monotonicity Constraint

The method introduces a learnable noise mapping, denoted as the recorruptor, which is constrained to be monotonic. This mapping generates a recorrupted observation as:

y1 = y + τh(w′), where w′ ∼ N (0, In) is sampled from a fixed normal distribution and τ > 0 controls the strength of the recorruption.

The constraint on the learned map is crucial for robustness and expressivity:

h: R n → R n constrained to a class H of admissible recorruption maps that are component-wise and (a.e.) differentiable, and which output zero-mean distributions, i.e., Ew′∼N(0,In)[h(w′)] = 0.

The paper notes that To ensure g, h ∈ H, we design h to be a monotonic neural network [30].

Theoretical Equivalence to Supervised Learning and Related Methods

The theoretical analysis shows how the self-supervised objective relates to supervised learning. When the true noise map belongs to the admissible class H (i.e., if g ∈ H), the self-supervised problem tends toward its supervised counterpart. Specifically:

min f max h Ey1,y∥f(y1) − y∥2 + 2τf(y1)⊤h(w′) i = min f max h Ey1,x∥f(y1) − x∥2 s.t. Ey,w′f(y + τh(w′))⊤h(w′) = 0 ∀h ∈ H.

Furthermore, L2R is related to other frameworks:

It can be seen as a generalization of the UNSURE loss in Eq. (6), which uses a more flexible set of constraints parametrized by a (monotonic) neural network.

The method can also be linked to GR2R in Eq. (10), as the loss in Eq. (18) can be rewritten as LGR2R(y; f) = Ey1,y2y∥f(y1) − y2∥2 where the paired recorruption process is learned instead of being fixed.

Performance Across Diverse Noise Regimes

The framework demonstrates strong performance under various challenging noise models by learning the appropriate recorruptor structure:

Log-gamma noise (heavy-tailed): L2R attains the highest PSNR in both noise settings, reaching up to 32.40 dB on DIV2K.
Laplace noise: L2R provides the best PSNR in the first setting, achieving 31.59 dB on BSDS500 and 31.79 dB on DIV2K in the first setting, and remaining competitive in the second setting.
Correlated noise: L2R is the best self-supervised alternative without access to pε, reaching 29.75 dB (DIV2K) in the second correlated noise setting, substantially outperforming other distribution-agnostic baselines and narrowing the gap to SURE.
Poisson–Gaussian noise: L2R achieves the best PSNR among methods that do not assume knowledge of the noise distribution, improving over PG-UNSURE by avoiding divergence estimation.

Improvements for AI systems

As a fastidious researcher, I have analyzed Learning to Recorrupt: Noise Distribution Agnostic Self-Supervised Image Denoising (L2R). The core innovation is replacing explicit knowledge of the noise distribution with a learnable, monotonic neural network that models the corruption process via a min-max saddle-point objective.

Here are the specific improvements you can make to AI systems and what those improved systems can do:

) Specific Improvements for AI Systems:

Do not rely on ground truth clean images for training denoising models; instead, leverage single noisy observations of real-world data (e.g., medical scans, satellite imagery).
Implement a self-supervised training paradigm where the model learns to denoise by simulating noise corruption (recorruption) rather than relying on fixed architectural priors or loss functions tied to specific noise types.
Utilize a Learning to Recorrupt framework where an auxiliary, learnable neural network (the recorruptor, denoted as 'h') is trained adversarially against the denoiser ('f'). This forces the system to learn what it means for an image to be corrupted by various unknown noise distributions.
Constrain the learned recorruptor network ('h') to be monotonic. This constraint ensures that the learned corruption process respects certain ordering properties, leading to more robust and interpretable models, especially under heavy-tailed noise (like log-gamma or Laplace).
Employ a min-max saddle-point optimization objective that simultaneously optimizes the denoiser ('f') to minimize reconstruction error and the recorruptor ('h') to satisfy a zero-correlation constraint with respect to the true unknown noise distribution.
Dynamically adapt the complexity of the learned recorruptor 'h' based on the corruption structure (e.g., use element-wise for independent noise, or spatial kernels for correlated noise).
In specific regimes (like Poisson-Gaussian noise), incorporate signal-dependent scaling mechanisms into the recorruption process to better model multiplicative and additive corruption effects.

) What the Improved AI System Can Do:

Do not require pre-labeled datasets of clean images, drastically reducing data acquisition costs in domains like remote sensing or medical imaging where ground truth is scarce or expensive.
Achieve state-of-the-art denoising performance across a wide, diverse range of noise distributions—including heavy-tailed noise (log-gamma), sharp outliers (Laplace), and spatially correlated noise—without needing prior statistical knowledge of the corruption process.
Generate high-fidelity, robust reconstructions from single noisy inputs in real-world scenarios where the exact nature of the sensor or environment's noise is unknown or shifts over time.
Provide a practical, data-driven proxy for characterizing unknown noise statistics; by examining the learned recorruptor 'h', researchers can gain an interpretable understanding of the underlying corruption (e.g., identifying if spatial correlation or heavy tails are dominant).
Be significantly more robust to distribution shift and novel noise types compared to methods like GR2R or UNSURE, which require manual parameter tuning for new noise families.
Perform superior denoising in structured environments (like spatially correlated data) by learning appropriate spatial kernels, outperforming simple element-wise approaches that fail when noise has spatial structure.