⚡ electrical engineering

Trainable Nonexpansive Denoisers for Contractive Image Reconstruction

This paper introduces a trainable denoiser architecture that exploits image lattice permutations to guarantee global nonexpansiveness, enabling provably convergent image reconstruction for inverse problems like superresolution and deblurring while maintaining competitive performance.

Original authors: Arghya Sinha, Aditya Banerjee, Trishit Mukherjee, Kunal N. Chaudhury

Published 2026-07-28✓ Author reviewed ⓘ

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Arghya Sinha, Aditya Banerjee, Trishit Mukherjee, Kunal N. Chaudhury

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a giant, messy puzzle where some pieces are missing, and the ones you have are covered in static. This is the daily life of "computational imaging," a field of science dedicated to cleaning up blurry photos, fixing MRI scans, or zooming in on tiny details without making them look like pixelated blobs. The secret weapon for solving these puzzles is a special kind of mathematical "eraser" called a denoiser. Think of a denoiser as a super-smart editor that looks at a noisy picture and guesses what the clean version should look like.

For years, scientists have been teaching computers to be these editors using deep learning, which is like training a digital brain with millions of examples. But there's a catch: sometimes, when you ask this digital brain to work repeatedly to fix a picture, it gets confused, starts hallucinating weird patterns, or just spins its wheels forever without finding the answer. To stop this, researchers need to build a "safety cage" around the editor, ensuring it never makes a move that pushes the picture further away from the truth. This paper tackles the tricky problem of building a denoiser that is both a genius editor and a perfectly obedient robot, guaranteed to never go rogue.

The Problem: The Wild Editor

In the world of image reconstruction, we often use a method called "Plug-and-Play" (PnP). Imagine you are trying to restore an old, scratched photo. You have a rulebook (the math of how the photo got damaged) and a magic editor (the denoiser). You take a guess, apply the rulebook, then ask the magic editor to clean it up. You repeat this over and over.

The problem is that most magic editors are "wild." They are trained to be amazing at cleaning up photos, but they don't have a strict rule that says, "You must never make the picture more chaotic than it was before." If you ask a wild editor to work in a loop, it might accidentally invent new noise or distort the image, causing the whole process to fail.

Some researchers tried to fix this by putting "soft" constraints on the editor, like telling it, "Try not to be too crazy," during training. But this is like telling a dog to sit only when you are watching; once you look away, the dog might jump on the table. These methods work well on the photos they were trained on, but they offer no guarantee that the editor won't go wild on a new, unseen picture.

The Solution: The Permutation Party

The authors of this paper, Arghya Sinha and his team, came up with a clever way to build an editor that is nonexpansive. In plain English, this means the editor is guaranteed to never make the distance between two different images larger. If you have two slightly different versions of a photo, the editor will process them in a way that keeps them just as close (or closer) together. It's like a strict bouncer who ensures no two people ever drift further apart in a crowded room.

To achieve this, they used a concept called permutations. Imagine you have a photo of a cat. Now, imagine you have a deck of cards that can shuffle the pixels of that photo in specific, mathematical ways—rotating it, flipping it, or sliding blocks of pixels around. The team created a system where the denoiser doesn't just look at the photo once. Instead, it looks at the photo and all its shuffled "twins" at the same time.

Here is the magic trick: The neural network (the brain of the editor) is only allowed to decide how much weight to give to each shuffled version. It acts like a conductor, saying, "This shuffled version looks very similar to the original, so I'll listen to it closely. That one looks weird, so I'll ignore it." Crucially, the actual act of mixing the pixels together is done by a simple, linear math operation. By separating the "thinking" (deciding weights) from the "doing" (mixing pixels), they built a system that is mathematically proven to be stable.

How It Works: The Symmetry Rule

The paper introduces a few rules to make sure this system works perfectly:

The Mirror Rule: If the system decides to shuffle the photo one way, it must also be able to un-shuffle it in a perfectly symmetrical way. This ensures the math stays balanced.
The Positive Rule: The system is forced to only give positive numbers as weights, preventing it from canceling things out in weird ways.
The Warm-Up: When starting to fix a picture, the system uses a "warm-up" phase. It starts with a rough guess, cleans it up, and then uses that cleaner version as a reference to guide the rest of the process. This is like a musician tuning their instrument before the concert to ensure the rest of the performance is in harmony.

The Results: Safe and Sharp

The team tested their new "Nonexpansive Denoiser" on classic image problems like removing blur and making low-resolution images look sharp (super-resolution).

The Proof: They mathematically proved that their method is contractive. This is a fancy way of saying that every time the system runs a step, it gets closer to the final, correct answer. It cannot get stuck in a loop or drift away.
The Performance: In tests, their method performed just as well as the best "wild" editors that don't have these safety guarantees. For example, on a standard test set called CBSD10, their method achieved a score of 28.23 dB for image deblurring, which was competitive with top methods like ADMM+CoCo-DRUNet (28.74 dB) and GSPnP+GSDRUNet (29.17 dB).
The Stability: Unlike other methods that sometimes fail or produce weird artifacts (like the "IHQS+SPC-DRUNet" method which showed signs of instability in their tests), this new method stayed stable in every single scenario they tried.

Why It Matters

The paper argues that we don't have to choose between a powerful editor and a safe one. By using the clever trick of shuffling pixels (permutations) and letting the neural network just decide the weights, they built a system that is both smart and safe.

The authors showed that while other methods rely on "soft" rules that might fail on new data, their approach provides a global guarantee. It's like building a roller coaster that is mathematically proven to never leave the tracks, no matter how fast it goes. This is a big deal for medical imaging and scientific photography, where you can't afford for the computer to guess wrong or invent fake details. The method is available as open-source code, ready for others to use and build upon.

Technical Summary: Trainable Nonexpansive Denoisers for Contractive Image Reconstruction

Problem Statement
Image reconstruction from noisy and incomplete linear measurements is a fundamental ill-posed inverse problem in computational imaging, encompassing tasks such as deblurring, superresolution, and MRI. Classical variational approaches solve $\min_x f(x) + g(x)$ , where $f$ is a data-fidelity term and $g$ is a regularizer. However, designing effective explicit regularizers is difficult. Plug-and-Play (PnP) methods address this by replacing the proximal operator of $g$ with a general denoiser $D$ . While PnP decouples the denoiser from the forward model, allowing the reuse of pretrained networks, it often lacks convergence guarantees when $D$ is treated as a black box.

Existing approaches to ensure convergence typically enforce Lipschitz constraints (specifically, nonexpansivity, where the Lipschitz constant is $\le 1$ ) only empirically on training data via penalty terms. These "soft" constraints provide no guarantees outside the training distribution, potentially leading to unstable behavior or divergence during reconstruction. Conversely, enforcing global Lipschitz constraints on high-performance nonlinear architectures (like DRUNet or DnCNN) often significantly degrades denoising performance. The core challenge is to construct a denoiser that is globally nonexpansive by design while retaining sufficient expressiveness for high-quality reconstruction.

Methodology
The authors propose a principled framework for constructing a parametric family of denoisers that are globally nonexpansive by design. The method relies on exploiting the action of permutations on the image lattice.

Permutation-Based Aggregation:
The denoiser $D$ is constructed by aggregating permuted versions of the input image. Let $G$ be a group of permutations acting on the image lattice $\Omega$ . For a noisy image $x$ and a reference image $\xi$ , the denoiser computes a weighted aggregation:
$D(x; \xi) = \left( \sum_{\pi \in G} N_\theta(\xi, \pi \cdot \xi) \odot (\pi \cdot x) \right) \oslash \left( \sum_{\pi \in G} N_\theta(\xi, \pi \cdot \xi) \right)$
Here, $N_\theta$ is a lightweight neural network that predicts positive weights based on the reference image $\xi$ and its permuted variants. The nonlinearity is confined to the weight prediction (dependent on $\xi$ ), while the aggregation over $x$ remains linear.
Structural Conditions for Nonexpansivity:
To guarantee that the resulting operator is nonexpansive (Lipschitz constant $\le 1$ ), the authors impose specific structural conditions:
- Positivity: The network output $N_\theta$ must be strictly positive.
- Inverse-Consistency Condition (ICC): The weights must satisfy $N_\theta(\xi, \pi \cdot \xi) = \pi \cdot N_\theta(\xi, \pi^{-1} \cdot \xi)$ . This ensures the underlying linear aggregation operator is symmetric.
- Symmetrization: Since the normalization step in the definition of $D$ can break symmetry, the authors apply a symmetrization protocol $\phi(D)$ (based on Sreehari et al., 2016) to produce a symmetric, stochastic operator $D_{sym}$ .
- Permutation Set: The set $G$ is chosen to be closed under inversion and to avoid nontrivial order-2 elements (involutions) that would make ICC overly restrictive. The authors utilize a set of local translations $T[R]$ which satisfies these requirements.
Convergent Reconstruction:
The symmetrized denoiser $D_{sym}$ is integrated into the PnP-HQS (Half-Quadratic Splitting) framework. The authors prove that under mild assumptions (including the non-annihilating property of the forward operator), the resulting fixed-point iteration operator $T = D_{sym} \circ \text{prox}_{\rho f}$ is contractive. This guarantees global convergence to a unique fixed point.
Training and Implementation:
The network $N_\theta$ is trained end-to-end to minimize the Mean Squared Error (MSE) of the denoising task. During reconstruction, a "warm-up" phase is employed where the reference image $\xi$ is updated iteratively to improve weight quality; once the reference stabilizes, it is frozen to satisfy the theoretical conditions for contraction.

Key Contributions

Design of Nonexpansive Denoisers: The paper proposes a framework to construct denoisers that are globally nonexpansive by design, avoiding the need to estimate or constrain the Lipschitz constant of the network during training.
Theoretical Guarantees: The authors prove that the reconstruction operator formed by integrating their denoiser with PnP-HQS is contractive under mild assumptions, ensuring global convergence.
Empirical Performance: The proposed method achieves competitive reconstruction quality on standard inverse problems (deblurring and superresolution) compared to state-of-the-art baselines, while providing certified stability.

Experimental Results
Experiments were conducted on standard datasets (CBSD10) for image deblurring and superresolution tasks under various noise levels ( $\nu \in \{0.03, 0.05\}$ ) and blur kernels.

Denoising: At high noise levels ( $\sigma = 50/255$ ), the proposed denoiser significantly outperforms BM3D and remains competitive with the unconstrained DRUNet, preserving fine details effectively.
Reconstruction: In deblurring and superresolution tasks, the proposed "Ours" method achieves PSNR scores comparable to or better than recent convergent baselines (e.g., ADMM+CoCo-DRUNet, IHQS+SPC-DRUNet, GSPnP+GSDRUNet).
Stability: Unlike soft-constrained baselines (e.g., SPC-DRUNet, CoCo-DRUNet) which rely on empirical constraints and may exhibit divergence in practice, the proposed method remains stable across all tested scenarios due to its global nonexpansivity.
Efficiency: The use of local translations and the ICC strategy (pairing $\pi$ and $\pi^{-1}$ ) reduces computational overhead. The symmetrization step adds runtime, but caching network outputs mitigates this.

Significance and Claims
The paper claims to resolve the trade-off between expressiveness and convergence guarantees in PnP methods. By leveraging permutation group actions, the authors demonstrate that it is possible to construct a highly expressive, trainable denoiser that is provably nonexpansive without relying on empirical constraints that fail outside the training distribution. The resulting reconstruction framework offers a "strong form of convergence guarantee" (contraction) that rules out divergence, providing a more reliable and trustworthy approach for computational imaging applications where stability is essential. The work advances the development of interpretable, reliable machine learning methods for inverse problems.