---
title: 'Noisier2Inverse: Noise-Aware Inverse Problems'
url: https://www.emergentmind.com/topics/noisier2inverse
type: topic
---

# Noisier2Inverse: Noise-Aware Inverse Problems

Searching arXiv for recent papers on "Noisier2Inverse" and closely related variants to ground the article in current literature.
Noisier2Inverse denotes a self-supervised learning framework for image reconstruction in inverse problems with noisy measurements, especially when the measurement noise is statistically correlated and clean ground-truth images are unavailable. In its primary sense, the term refers to the correction-free method introduced in "Noisier2Inverse: Self-Supervised Learning for Image Reconstruction with Correlated Noise" [2503.19468], which generalizes the “learn from noisier data” idea of Noisier2Noise from denoising to general linear inverse problems. More broadly, subsequent work has used the same framework or closely related constructions for angular deblurring in photoacoustic tomography, log-domain reconstruction under multiplicative noise, and self-supervised learned iterative CT reconstruction [2604.15681] [2606.27635] [2606.26991]. In a looser conceptual sense, the label has also been invoked for noise-aware inverse design and editable diffusion inversion, although those usages are best understood as analogies rather than the canonical meaning of the term [2112.03912] [2509.25776].

## 1. Definition and scope

In the 2025 formulation, Noisier2Inverse is a self-supervised, one-step deep learning method for solving linear inverse problems with correlated additive noise, without access to clean ground-truth images [2503.19468]. The basic setting is
$$
y = A x + \xi,
$$
where $x \in \mathbb{R}^n$ is the unknown clean image, $A : \mathbb{R}^n \to \mathbb{R}^m$ is a known linear forward operator, and $y \in \mathbb{R}^m$ is the noisy measurement. The method aims to learn a reconstruction map
$$
B : \mathbb{R}^m \to \mathbb{R}^n
$$
that approximates the ideal Bayes estimator minimizing
$$
\mathbb{E}\bigl[ \|B(Y) - X\|_2^2 \bigr].
$$

A defining step is the generation of a noisier measurement by adding additional noise drawn from the same distribution:
$$
z = y + n.
$$
Using a fixed initial reconstructor $B^\sharp$ and an image-space network $f_\theta$, Noisier2Inverse trains the composite map $f_\theta \circ B^\sharp$ with a loss in measurement space, not image space [2503.19468]. The target is the extrapolated measurement $2Y-Z$, rather than the original noisy datum $Y$. This eliminates the need for an extrapolation step during inference, which would otherwise suffer from ill-posedness [2503.19468].

This scope distinguishes Noisier2Inverse from methods that rely on clean targets, paired noisy realizations, or pixelwise independence in image space. The framework is explicitly intended for general inverse problems, including computed tomography, microscopy, and seismic imaging, and for settings in which detector imperfections, photon scattering, or other physical effects induce dependencies in the noise structure [2503.19468].

## 2. Mathematical formulation

The core practical construction in Noisier2Inverse begins with noisy measurements
$$
y_t = A x_t + \xi_t, \quad t=1,\dots,T,
$$
and noisier measurements
$$
z_t = y_t + n_t = A x_t + \xi_t + n_t.
$$
Here $n_t$ is sampled from the same correlated noise distribution as $\xi_t$ [2503.19468]. The overall reconstructor is built from a fixed initial reconstruction map $B^\sharp$ and a learned image-space network $f_\theta$.

The self-supervised loss is defined by
$$
L_W(\theta) = \sum_{t=1}^T \bigl\| W \bigl[ A( f_\theta \circ B^\sharp)(z_t) \bigr]
          - W (2 y_t - z_t) \bigr\|_2^2.
$$
In expectation form,
$$
\mathcal{L}_W(\theta) = \mathbb{E} \bigl[ \| W A [ f_\theta \circ B^\sharp(Z) ] - W (2Y - Z) \|_2^2 \bigr].
$$
The operator $W$ defines the norm in measurement space; the paper considers $W = Id$ and $W = \nabla$, the latter yielding a Sobolev loss [2503.19468].

The central theorem states that for any linear operator $W$,
$$
\arg\min_f \mathbb{E} \bigl\{ \| W A[ f \circ B^\sharp(Z) ] - W A[X] \|_2^2 \bigr\}
=
\arg\min_f \mathbb{E} \bigl\{ \| W A[ f \circ B^\sharp(Z) ] - W (2Y - Z) \|_2^2 \bigr\}.
$$
Accordingly, minimizing the self-supervised risk with target $2Y-Z$ is equivalent, in terms of minimizers, to minimizing the supervised risk with respect to the clean data $AX$ [2503.19468].

This formulation depends on additive, zero-mean noise and on the ability to sample additional noise from the same distribution as the original noise. The framework does not require independence across pixels or measurements in the observed data themselves. Rather, it requires accurate modeling of the noise generation process so that the added noise $N$ matches the law of $\Xi$ [2503.19468]. This is the technical basis for its applicability to correlated noise.

## 3. Relation to predecessor methods

Noisier2Inverse is most naturally situated between Noise2Inverse and Noisier2Noise. Noise2Inverse, introduced for tomography, is a self-supervised deep CNN-based denoising method for linear image reconstruction algorithms that does not require any additional clean or noisy data, but it assumes the measured noise is element-wise independent and zero-mean [2001.11801]. Its key device is to compute multiple statistically independent reconstructions by splitting the measurement domain into disjoint subsets and reconstructing each subset separately. In tomography, this is accomplished by splitting projection angles, reconstructing sub-sinograms with linear FBP operators, and training on pairs of complementary sub-reconstructions [2001.11801].

By contrast, Noisier2Inverse inherits from Noisier2Noise the idea of creating noisier data by adding additional synthetic noise, but it modifies the training target and the domain of the loss [2503.19468]. In Noisier2Noise, one learns from $(Z,Y)$ and typically requires an extrapolation step at inference. The 2025 Noisier2Inverse paper states that such extrapolation tends to become unstable in inverse problems, especially when the forward operator is ill-conditioned [2503.19468]. The measurement-space loss against $2Y-Z$ is designed specifically to avoid that failure mode.

A related but earlier line is "Image Deconvolution via Noise-Tolerant Self-Supervised Inversion" [2006.06156], which extends Noise2Self from denoising to general inverse problems with a known forward operator by training the composite $g \circ f$ as a self-supervised denoiser. That framework relies on $\mathcal{J}$-invariant masking and independence across measurement dimensions. A plausible implication is that Noisier2Inverse and self-supervised inversion share a common objective—ground-truth-free learning of inverse maps—but they operationalize self-supervision differently: the former via noisier measurements and measurement-space extrapolated targets, the latter via masking and $\mathcal{J}$-invariance [2006.06156].

A common misconception is that Noisier2Inverse is merely Noise2Inverse with more corruption. The 2025 formulation is more specific. It is a correction-free self-supervised approach whose loss is constructed so that the network is trained to recover an extrapolated image in measurement space instead of the original noisy one [2503.19468]. That distinction is what removes the inference-time extrapolation step.

## 4. Noise models and implementation patterns

The 2025 paper is explicitly designed for correlated additive noise. In its CT experiments, the noise is modeled as
$$
\Xi = G_\sigma * \epsilon,
$$
where $\epsilon \sim \mathcal{N}(0,\delta^2 I)$ and $G_\sigma$ is a Gaussian convolution kernel with bandwidth $\sigma$ [2503.19468]. The additional noise $N$ used to construct $Z=Y+N$ is drawn from the same correlated noise distribution. The method therefore requires an approximate noise model, and its performance depends on the fidelity of that model [2503.19468].

The implementation uses the same U-Net architecture as in Noise2Inverse, with a fixed initial reconstructor $B^\sharp$, typically filtered backprojection in CT [2503.19468]. The training pipeline is: sample noisy measurements $y_t$, generate additional correlated noise $n_t$, form $z_t=y_t+n_t$, compute $x_t^{(0)} = B^\sharp(z_t)$, predict $\hat{x}_t = f_\theta(x_t^{(0)})$, re-project via $A\hat{x}_t$, and compare in measurement space with $2y_t-z_t$ [2503.19468]. Two inference modes are considered:
$$
x_W^{(y)} = (f_\theta \circ B^\sharp)(y), \qquad
x_W^{(z)} = (f_\theta \circ B^\sharp)(z).
$$
Experimentally, prediction with noisy data slightly outperforms prediction with noisier data [2503.19468].

Later work has adapted the same logic to other noise structures. "Denoising ICF Images with Multiplicative Uniform Noise: A Self-Supervised Study Based on the Log-Domain Noisier2Inverse Framework" [2606.27635] considers
$$
y = (A x) \odot n, \quad n \sim \mathrm{Uniform}(\ell_i, h_i),
$$
and applies the logarithmic transform
$$
\log y = \log(Ax) + \log n
$$
to port the problem to an additive-noise setting. The corresponding self-supervised loss is
$$
\varphi^* = \arg\min_\varphi\, \mathbb{E}\!\left[ \bigl\|\log\!\bigl(A R_\varphi(\tilde{z})\bigr) - (2\tilde{y} - \tilde{z}) \bigr\|_2^2 \right],
$$
with a theorem showing equivalence to supervised learning against $\log(Ax)$ in the transformed domain [2606.27635]. This demonstrates that Noisier2Inverse can be ported to non-additive noise models by working in a suitable transform domain.

## 5. Applications and domain-specific extensions

The most direct application in the originating paper is 2D parallel-beam CT, with experiments on the CBCT Walnut dataset and a heart CT dataset under full-angle and sparse-angle settings [2503.19468]. The framework is intended for general linear inverse problems, and the paper explicitly cites computed tomography, microscopy, and seismic imaging as target domains [2503.19468].

A substantial domain-specific extension appears in "Self-Supervised Angular Deblurring in Photoacoustic Reconstruction via Noisier2Inverse" [2604.15681]. There, finite-size detector effects in photoacoustic tomography are recast as angular convolution in polar coordinates:
$$
g_P = A_w f_P + \eta_P,
$$
where $A_w$ is angular convolution with a known angular point-spread function [2604.15681]. The Noisier2Inverse loss becomes
$$
\mathcal{L}(\theta) = \mathbb{E}\bigl[ \bigl\| A\,R_\theta(y_P + \varepsilon_P') - A(y_P - \varepsilon_P') \bigr\|_2^2 \bigr],
$$
or, in the paper’s compact notation,
$$
\mathcal{L}(\theta) = \mathbb{E}\,\Big[ \big\| \mathcal{A} \bigl(R_\theta(y_P + \varepsilon_P')\bigr) - \mathcal{A}(y_P - \varepsilon_P') \big\|_2^2 \Big].
$$
This application is notable because the transforms used before learning are non-unitary, so the resulting noise is correlated; the method is explicitly chosen because it allows correlated noise [2604.15681].

A second extension is "Enabling self-supervised learned primal dual with Noise2Inverse" [2606.26991], which is not a Noisier2Inverse paper in name but belongs to the same family of self-supervised inverse-problem methods. It extends Noise2Inverse to the Learned Primal-Dual algorithm by exploiting independence across angular subsets in CT sinograms [2606.26991]. A plausible implication is that the broader Noisier2Inverse program includes not only post-processing networks but also learned iterative reconstruction operators trained without ground truth.

The term has also been used more loosely in other areas. The RID-Noise paper on robust inverse design under noisy environments is described as something that “can be naturally understood as ‘inverse design that explicitly accounts for noise’,” and as a concrete version of what a “Noisier2Inverse” approach might look like [2112.03912]. Similarly, Editable Noise Map Inversion in diffusion editing is framed as a “Noisier2Inverse-style method” because it intentionally avoids the tightest reconstruction inverse in favor of a more editable noise-space representation [2509.25776]. These usages are conceptually related but should be distinguished from the canonical self-supervised inverse-problem framework introduced in 2025.

## 6. Empirical behavior, strengths, and limitations

The original Noisier2Inverse paper reports that the method clearly outperforms previous self-supervised approaches that account for correlated noise [2503.19468]. On full-angle CT, the reported PSNR and SSIM values show large gains over one-step Noisier2Noise and Noise2Inverse, especially as the correlation parameter increases. On sparse-angle CT, the method not only denoises but also clearly reduces sparse-data artifacts, whereas Noisier2Noise and Noise2Inverse lag behind [2503.19468]. The paper also reports that the Sobolev-loss variant better suppresses noise in homogeneous regions and improves global structural quality, while the plain measurement-space MSE variant preserves fine details slightly better [2503.19468].

In the log-domain ICF study, the best result is achieved by the log-domain approach with per-image JSON Uniform noise loading, with a mean PSNR of $21.41\db$ and SSIM of $0.8358$, a $+19.46\db$ improvement over the noisy input baseline of $1.95\db$ [2606.27635]. The paper compares against BM3D and Noise2Self, noting that Noise2Self fails because the pixel-independence assumption is violated by strong spatial correlation, while Log-Domain Noisier2Inverse remains fully self-supervised during training [2606.27635]. This reinforces the claim that correlated noise is not a marginal detail but a defining use case.

The PAT paper reports that the proposed method consistently outperforms alternative approaches that do not use supervised data and achieves performance close to supervised benchmarks, while remaining practical for real acquisitions with finite-size detectors [2604.15681]. It also introduces a statistically grounded early-stopping rule based on Earth Mover’s Distance between predicted residuals and the known noise distribution, because data-space losses in ill-posed inverse problems can otherwise overfit [2604.15681].

Several limitations recur across the literature. The original Noisier2Inverse paper requires an approximate noise model to generate the additional noise correctly, a reasonable initial reconstructor $B^\sharp$, and early stopping because training too long can reintroduce noise [2503.19468]. The PAT variant assumes a known angular PSF and a known additive noise law [2604.15681]. The log-domain variant emphasizes numerical stabilization around near-zero pixels and dependence on accurate per-image noise parameters [2606.27635]. More generally, the framework is most principled for known forward models and known or well-estimated noise distributions.

A second misconception is that Noisier2Inverse removes all artifacts of an inverse problem. The source papers do not support that interpretation. The framework is designed to learn a reconstruction function under noisy measurements without ground truth, and its guarantees are tied to the corresponding clean quantity in measurement space or transformed measurement space [2503.19468] [2606.27635]. This suggests that deterministic artifacts caused by model mismatch, limited-angle geometry, or unknown physics are not automatically resolved unless they are explicitly built into the forward model and training objective.

## 7. Conceptual significance and broader usage

In its strongest and most precise meaning, Noisier2Inverse is a self-supervised framework for general inverse problems in which one learns from noisier data and a measurement-space target chosen so that the self-supervised loss has the same minimizers as a supervised loss against the clean measurement quantity [2503.19468]. This is the sense that connects the 2025 paper to its later variants in PAT and multiplicative-noise ICF imaging [2604.15681] [2606.27635].

The broader significance of the framework is that it provides a route to self-supervised reconstruction without clean targets, without paired repeated acquisitions, and without the pixelwise independence assumptions that underwrite Noise2Self-style methods [2503.19468] [2001.11801]. It is therefore particularly relevant in modalities where the forward operator is known, the inverse problem is ill-posed, and the noise structure is correlated after acquisition or preprocessing.

A looser literature usage extends the label to any inverse method that is intentionally “noise-aware,” “learns from noisier data,” or “relaxes strict inversion in favor of robustness or editability.” The RID-Noise inverse design framework and ENM inversion for diffusion editing are explicit examples of that rhetorical extension [2112.03912] [2509.25776]. This suggests a family resemblance across fields: rather than seeking the tightest possible inverse of a clean forward map, these methods exploit controlled noise, added noise, or prompt-agnostic noise representations to improve robustness, stability, or editability. Still, the encyclopedia sense of Noisier2Inverse is anchored by the 2025 self-supervised reconstruction framework for correlated-noise inverse problems [2503.19468].

Source: https://www.emergentmind.com/topics/noisier2inverse