Fast Equivariant Imaging (FEI)

• Fast Equivariant Imaging (FEI) is an unsupervised framework that combines measurement consistency with known equivariance constraints to train deep imaging networks without ground-truth data.
• It decouples the network output from the measurement term using an auxiliary variable and augmented Lagrangian formulation, with a plug-and-play extension employing denoisers like BM3D or DnCNN to accelerate convergence.
• The framework achieves faster convergence and improved image quality in sparse-view CT, offering a modular design that can be adapted to various inverse imaging modalities and operator constraints.

Fast Equivariant Imaging (FEI) is an unsupervised learning framework for training deep imaging networks without ground-truth data by combining measurement consistency with equivariance constraints and solving the resulting problem through an augmented Lagrangian formulation. In the linear inverse-imaging setting

$y = A x^\dagger + \epsilon,$

with unknown image $x^\dagger \in \mathbb{R}^n$, measurements $y \in \mathbb{R}^m$, noise $\epsilon$, and forward operator $A \in \mathbb{R}^{m \times n}$, FEI starts from the observation that measurement consistency alone is insufficient because the reconstruction network cannot learn components in the null space of $A$. It therefore adopts a group of known unitary transforms $\mathcal{G}=\{T_g\}_g$ and enforces equivariance of $G_\theta \circ A$, then accelerates training by decoupling the optimization with an auxiliary variable and, in its plug-and-play form, auxiliary denoisers such as BM3D or DnCNN (Xu et al., 9 Jul 2025).

1. Problem formulation and relation to classical Equivariant Imaging

Classical inverse imaging is often posed through a maximum a posteriori objective,

$x^* = \arg\min_x \{ f_{mc}(Ax,y) + R(x) \},$

where $f_{mc}(\cdot,\cdot)$ is a measurement-consistency term, such as an $x^\dagger \in \mathbb{R}^n$0 or SURE loss, and $x^\dagger \in \mathbb{R}^n$1 is a regularizer. In unsupervised deep learning, one instead seeks a reconstruction network $x^\dagger \in \mathbb{R}^n$2 by minimizing

$x^\dagger \in \mathbb{R}^n$3

This objective enforces consistency with the measurements, but does not recover image components that lie in the null space of $x^\dagger \in \mathbb{R}^n$4 (Xu et al., 9 Jul 2025).

Equivariant Imaging (EI) addresses that limitation by assuming a known transformation group $x^\dagger \in \mathbb{R}^n$5, for example rotations or shifts, under which natural images are invariant. The key constraint is

$x^\dagger \in \mathbb{R}^n$6

This produces the EI objective

$x^\dagger \in \mathbb{R}^n$7

where $x^\dagger \in \mathbb{R}^n$8 balances data consistency and the equivariance penalty. FEI does not discard this EI structure; rather, it reformulates it so that the optimization becomes substantially more efficient while retaining the same unsupervised character (Xu et al., 9 Jul 2025).

A common misunderstanding is to identify FEI with a new equivariance penalty. The defining change is instead algorithmic: FEI recasts EI as an augmented Lagrangian problem with alternating updates of a latent variable, network parameters, and multipliers. The equivariance term remains central.

2. Augmented Lagrangian reformulation

FEI introduces an auxiliary latent variable $x^\dagger \in \mathbb{R}^n$9 to decouple the network output from the measurement-consistency term:

$y \in \mathbb{R}^m$0

With multiplier $y \in \mathbb{R}^m$1 and penalty $y \in \mathbb{R}^m$2, the scaled augmented Lagrangian is

$y \in \mathbb{R}^m$3

The paper also gives the classical augmented-Lagrangian form for each $y \in \mathbb{R}^m$4,

$y \in \mathbb{R}^m$5

with $y \in \mathbb{R}^m$6 and $y \in \mathbb{R}^m$7 (Xu et al., 9 Jul 2025).

This reformulation has a specific computational role. The measurement-consistency term depends on $y \in \mathbb{R}^m$8, while the network-dependent equivariance term remains in $y \in \mathbb{R}^m$9. A plausible implication is that FEI makes the optimization more modular: the inverse problem enters primarily through $\epsilon$0, whereas equivariance remains attached to the learned map $\epsilon$1. That modularity is consistent with the paper’s statement that the framework can incorporate robust EI, multi-operator EI, or sketched EI variants (Xu et al., 9 Jul 2025).

3. Alternating updates and the plug-and-play extension

FEI alternates between updating $\epsilon$2 and $\epsilon$3. For fixed $\epsilon$4 and $\epsilon$5, the $\epsilon$6-subproblem is

$\epsilon$7

The paper notes that even when $\epsilon$8, this is a large quadratic. One option is Nesterov-accelerated gradient. The second option, which defines PnP-FEI, is plug-and-play: the quadratic term is interpreted as a Gaussian MAP prior and its proximal is replaced by a denoiser $\epsilon$9. A single PnP iteration is

$A \in \mathbb{R}^{m \times n}$0

where $A \in \mathbb{R}^{m \times n}$1 is a small step size. The denoiser can be BM3D, DnCNN, and related models; empirically, this implicit prior speeds convergence and often improves final image quality (Xu et al., 9 Jul 2025).

The outer workflow is specified explicitly. For a single sample $A \in \mathbb{R}^{m \times n}$2 and random $A \in \mathbb{R}^{m \times n}$3, Step 0 computes $A \in \mathbb{R}^{m \times n}$4 and initializes $A \in \mathbb{R}^{m \times n}$5. Step 1 updates $A \in \mathbb{R}^{m \times n}$6 either by NAG on

$A \in \mathbb{R}^{m \times n}$7

or by the PnP rule above. Step 2 performs the equivariance pathway: $A \in \mathbb{R}^{m \times n}$8, $A \in \mathbb{R}^{m \times n}$9, $A$0. Step 3 takes one SGD step on

$A$1

Step 4 updates the multiplier as in ADMM: $A$2. Training repeats over mini-batches or individual $A$3 for $A$4 epochs (Xu et al., 9 Jul 2025).

PnP-FEI is therefore not merely FEI with a denoising post-processing stage. The denoiser appears inside the $A$5-update and acts as an implicit prior during optimization.

4. Reported performance on sparse-view CT

The reported evaluation is on 50-view sparse-angle CT using the CT100 dataset, with 10 training and 10 test images at $A$6. In that setting, vanilla EI requires approximately 10,000 epochs and approximately $A$7 more wall-clock time to reach PSNR $A$8 dB. PnP-FEI with DnCNN converges to PSNR $A$9 dB in the same number of epochs but at one-tenth the runtime. Pure FEI, without a denoiser, already outperforms EI in both PSNR and time to converge. Empirical PSNR-versus-iteration curves show that PnP-FEI reaches EI’s final PSNR in one-tenth the iterations, and hence about $\mathcal{G}=\{T_g\}_g$0 speedup in training time (Xu et al., 9 Jul 2025).

Method	Training PSNR (dB)	Test PSNR (dB)
EI	$\mathcal{G}=\{T_g\}_g$1	$\mathcal{G}=\{T_g\}_g$2
FEI	$\mathcal{G}=\{T_g\}_g$3	$\mathcal{G}=\{T_g\}_g$4
PnP-FEI (DnCNN)	$\mathcal{G}=\{T_g\}_g$5	$\mathcal{G}=\{T_g\}_g$6

The same test table reports a supervised upper bound of $\mathcal{G}=\{T_g\}_g$7 dB. On these numbers, both FEI and PnP-FEI improve over vanilla EI on unseen CT100 images and narrow the gap to supervised training. The paper also states that they exhibit stable OOD performance, for example in a 40-view test setting (Xu et al., 9 Jul 2025).

The reported acceleration is specific. The abstract states that PnP-FEI achieves an order-of-magnitude $\mathcal{G}=\{T_g\}_g$8 acceleration over standard EI on training U-Net with the CT100 dataset for X-ray CT reconstruction, with improved generalization performance. The evidence presented in the paper is therefore both optimization-oriented and reconstruction-oriented: faster convergence is accompanied by higher final PSNR on training and test data.

5. Generalization, strengths, and limitations

The paper characterizes FEI and PnP-FEI as improving generalization relative to vanilla EI. On unseen CT100 images, both methods exceed the EI baseline, and PnP-FEI comes closest to the supervised upper bound. The same discussion attributes stable OOD performance to both FEI variants, including the 40-view test reported in Figure 1 (Xu et al., 9 Jul 2025).

Three strengths are stated directly. First, decoupling via Lagrangian and ADMM accelerates convergence. Second, plug-and-play denoisers inject powerful image priors without explicit penalty tuning. Third, the framework is modular and can incorporate robust EI, multi-operator EI, or sketched EI variants. These properties place FEI at the intersection of self-supervised inverse problems, operator splitting, and plug-and-play reconstruction (Xu et al., 9 Jul 2025).

The limitations are equally explicit. The $\mathcal{G}=\{T_g\}_g$9-update is nonconvex, so global convergence guarantees are precluded. Hyperparameters, including $G_\theta \circ A$0, $G_\theta \circ A$1, and the number of inner $G_\theta \circ A$2-iterations, must be tuned. In PnP-FEI, the denoising step relies on an external denoiser trained on natural images, and mismatch may degrade performance on very different domains. A common misconception is that the plug-and-play prior eliminates modeling assumptions; in the paper’s formulation it replaces an explicit proximal with an external denoiser, which shifts rather than removes the dependence on prior knowledge (Xu et al., 9 Jul 2025).

These limitations also define the boundary of the current claims. The paper reports empirical acceleration and improved image quality on sparse-view CT, but it does not claim a general global-convergence theorem for the end-to-end nonconvex training problem.

6. FEI as a broader equivariant imaging principle

Although FEI is introduced in the 2025 CT reconstruction framework, related work in the supplied literature presents several methods as realizing the same principle under different forward models and symmetry groups. In sparse inverse synthetic aperture radar imaging, "Self-Supervised-ISAR-Net Enables Fast Sparse ISAR Imaging" describes a self-supervised unfolded ADMM network that exploits rotation equivariance to suppress grating lobes caused by sparse radar echo; the paper reports inference time on a modern GPU on the order of 5–10 ms per $G_\theta \circ A$3 image, over an order of magnitude faster than 100–400 ms for conventional ADMM or other iterative solvers, while using only sparse radar echo data for training (Wang et al., 1 Jun 2025).

In accelerated MRI, "Scale-Equivariant Unrolled Neural Networks for Data-Efficient Accelerated MRI Reconstruction" embeds scale equivariance into the proximal operators of an unrolled network. Under the same memory constraints, it reports improvements both with and without data augmentations on in-distribution and out-of-distribution scaled images, with inference time essentially unchanged at $G_\theta \circ A$4 s/slice for the vanilla unrolled network and $G_\theta \circ A$5 s/slice for the scale-equivariant version (Gunel et al., 2022). In compressive quantitative MRI, "Nonlinear Equivariant Imaging: Learning Multi-Parametric Tissue Mapping without Ground Truth for Compressive Quantitative MRI" extends Equivariant Imaging to a nonlinear inverse problem by combining a spatial equivariance prior with a temporal Bloch-model prior. The reported reconstruction times are under 50 ms per slice for all deep methods, and the self-supervised method closely approaches a supervised baseline despite not using ground truth during training (Fatania et al., 2022).

A different line of work treats FEI as a continuous-group architectural principle rather than a training objective. "Efficient 3D affinely equivariant CNNs with adaptive fusion of augmented spherical Fourier-Bessel bases" realizes FEI under the continuous 3D affine group by using Monte Carlo augmentation and adaptive fusion of spherical Fourier-Bessel bases. The paper states that inference remains $G_\theta \circ A$6, identical to a standard CNN with the same kernel size, while improving equivariance and data efficiency on volumetric segmentation benchmarks (Zhao et al., 2024). "Equivariant Wavelets: Fast Rotation and Translation Invariant Wavelet Scattering Transforms" realizes FEI through fixed triglet filter banks and sparse Fourier-domain implementation, with Appendix benchmarks showing EqWS.jl is 20–50× faster per coefficient than Kymatio on CPU (Saydjari et al., 2021).

Taken together, these works suggest that FEI is both a named framework and a broader design pattern: encode known symmetries of the image formation process, preserve measurement consistency, and use algorithmic structure to reduce training or inference cost. The exact form varies—augmented Lagrangian splitting, unfolded ADMM, scale-equivariant proximals, affine group convolutions, or equivariant scattering—but the recurring objective is fast reconstruction or representation under explicit geometric or physical equivariance (Xu et al., 9 Jul 2025).