Deep Pseudo-Proximal Map: A Self-Supervised Data-Fitting Agent for Iterative Reconstruction

arXiv:2609.30727 · eess.IV · Submitted 2026-09-25 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Deep Pseudo-Proximal Map".

Dev: Here's the gist of the paper. Iterative algorithms for inverse imaging problems split reconstruction into alternating subproblems, one of which requires evaluating the data-fitting proximal map <ref:2609.30727#pg2> To address this, we propose the pseudo-proximal map (PPM), a reformulation of the data-fitting proximal map as the minimum mean square error estimate of a synthetic probabilistic model The deep PPM can be trained for any operator for which the forward model A and its transpose AT can be evaluated, with provable equivalence to the proximal map when A is linear <ref:2609.30727#pg2> We validate the deep PPM on Gaussian deblurring and 4× super-resolution, where the proximal map has an analytic solution, and on X-ray computed tomography (XCT), where no analytic solution exists <ref:2609.30727#pg2> Used as the data-fitting agent within an iterative reconstruction method, the deep PPM reproduces the reference reconstruction to within 1% NRMSE for all three operators, and for XCT, it replaces the inner conjugategradient loop with a single network evaluation that is 18× faster <ref:2609.30727#pg2>.

Rosa: Core Formulation of the PPM The data-fitting proximal map is defined as Fσ,γ(v; y) = arg min x 1/2σ squared y − A(x)2 + 1/2γ squared x − v2 <ref:2609.30727#pg5> The key observation motivating our proposed PPM is that Fσ,γ(v; y) can be interpreted as the maximum a posteriori (MAP) estimate under a Gaussian prior centered at v We define the PPM F˜σ,γ(v; y) as the corresponding conditional expectation of X˜ given Y˜ = y, i.e., F˜σ,γ(v; y) = Ev,γ

X˜ Y˜ = y: When A is linear, the posterior is Gaussian, so Fσ,γ(v; y) and F˜σ,γ(v; y) are equivalent Self-Supervised Training Mechanism We can implement the PPM as a network trained to minimize the mean square error (MSE) loss, since the MMSE estimator is equivalent to the conditional expectation We define the deep PPM F˜θσ,γ(v; y) in residual form as F˜θσ,γ(v; y) = v + H˜ θσ,γ(z), where z = AT (y − Av) is the back-projected residual and H˜ θσ,γ is a network with weights θ The weights are trained to minimize the MSE loss given by L(θ) = 1/K sum k=0 γϵk − H˜ θσ,γ(zk)2 Importantly, this residual depends only on sampled Gaussian noise and the operator A Validation and Performance Metrics We compare the deep PPM against CG and GD, each run for at most 50 iterations with relative convergence tolerance 10−6 <ref:2609.30727#pg5> For deblurring and SR, the Res-NRMSE with respect to the reference proximal map is approximately 6.0% for both operators <ref:2609.30727#pg2> For XCT, the Res-NRMSE with respect to CG is 11.8%, about twice that of the other operators <ref:2609.30727#pg2> Within MACE, the reconstruction obtained with the deep PPM differs from the reconstruction with the reference proximal map by less than 1% on average for all three tasks <ref:2609.30727#pg3> Computational Efficiency Gains For XCT, however, the deep PPM replaces a 710 ms CG solve (6.0 s for GD) with a 38.7 ms evaluation, an 18× reduction in the cost of each data-fitting step <ref:2609.30727#pg5> Over a full MACE run the speedup is 2.3× over CG and 14.5× over GD <ref:2609.30727#pg5> For deblurring and SR this offers no advantage, since the analytic solution and CG each take only a few milliseconds; these tasks serve as verification benchmarks rather than speed benchmarks <ref:2609.30727#pg2> Since the formulation is well-defined for any A, the PPM extends naturally to nonlinear forward models, where data-fitting proximal maps are most expensive Training and verifying the deep PPM for non-linear operators is the subject of future work The deep PPM reconstruction matches the reference reconstruction to within 1% NRMSE on average across all three operators, and for XCT it replaces the inner CG loop with a network evaluation that is 18× faster <ref:2609.30727#pg2>.

Dev: Page 1 --- Deep Pseudo-Proximal Map: A Self-Supervised Data-Fitting Agent for Iterative Reconstruction Haley Duba-Sullivan Purdue University West Lafayette, IN, USA hduba@purdue.edu Emma J.

Rosa: Reid Oak Ridge National Laboratory Oak Ridge, TN, USA reidej@ornl.gov Charles A.

Title and authors: Rosa: So we're looking at this paper today, "Deep Pseudo-Proximal Map: A Self-Supervised Data-Fitting Agent for Iterative Reconstruction," and it's about how to handle that tricky step in iterative reconstruction called the data-fitting proximal map.

Dev: Yeah, it tackles the problem where that proximal map doesn't have an easy analytical solution, so you usually have to run some inner loop solver at every single iteration of your main reconstruction loop.

Taro: And if you do that, it really compounds the cost of the whole outer reconstruction process.

Rosa: Exactly. This paper proposes something called the pseudo-proximal map, or PPM, which they reformulate as an estimate from a synthetic probabilistic model instead of just solving that hard minimization problem directly.

Dev: It says this deep PPM can be trained as a self-supervised neural network using only sampled Gaussian noise, which is a big deal because it means you don't need any real training images for the network to learn how to approximate that map.

Taro: That sounds promising for real-world deployment where getting perfect ground truth data is hard.

Rosa: The authors show this deep PPM works for any forward operator as long as they can evaluate the forward model and its transpose, and when the operator is linear, it’s provably equivalent to that data-fitting proximal map.

Dev: That equivalence when the operator is linear is important because it gives us a theoretical backbone for how this neural network actually relates to the math of reconstruction.

Taro: So, if we look at what they tested, they used Gaussian deblurring and four times super-resolution as examples where they already have an analytic solution for comparison.

Rosa: And then they tackled X-ray computed tomography, which is a tough case because there’s no easy analytical solution for that specific operator.

Dev: For XCT, instead of running a Conjugate Gradient loop every time, the deep PPM replaces that entire inner loop with just one evaluation of the network.

Taro: That sounds like it could be a massive speed boost when you're dealing with very complex data acquisition systems like X-ray machines.

Rosa: They report that using this deep PPM as the data-fitting agent in their iterative method, they reproduce the reference reconstruction to within one percent relative mean squared error for all three types of operators.

Dev: That one percent error across deblurring, super-resolution, and XCT is a pretty strong number for a learned component.

Taro: But I have to push back on the efficiency gains they show for XCT because they're comparing it to something quite slow.

Rosa: They do compare it against a Conjugate Gradient solver that runs up to fifty iterations with relative convergence tolerance of one point ten to the minus six, and they find that the deep PPM replaces that whole thing with just thirty-eight point seven milliseconds of evaluation time.

Dev: That is an eighteen times faster evaluation for XCT reconstruction, which really cuts down on latency in real-time scenarios.

Title and authors: Taro: For deblurring and super-resolution, they say there's no speed advantage because those tasks have analytic solutions that are already very fast to compute in the Fourier domain.

Rosa: So it seems the real payoff here is specifically for those problems where you don't have an easy formula, like XCT or maybe other non-linear operators.

Dev: They also show that for deblurring and super-resolution, the deep PPM tracks the convergence of standard solvers like Conjugate Gradient and Gradient Descent very closely across all their iterations.

Taro: If it tracks the trajectory well, that means the overall iterative reconstruction process isn't messed up by using this learned approximation.

Rosa: And for XCT specifically, they show that when you look at wall-clock time plots, the deep PPM tracks the standard Conjugate Gradient path to its equilibrium in about four point three seconds instead of nine point seven seconds or sixty two seconds for Gradient Descent.

Dev: That difference between forty point three and sixty two seconds is substantial when you're talking about processing large datasets.

Taro: What I wonder is how robust this training mechanism holds up if the world misbehaves in a way that changes the noise characteristics, since they only trained it on sampled Gaussian noise.

Rosa: That’s a fair concern, Taro. They state that by formulating the deep PPM in residual form, which depends only on sampled Gaussian noise and the operator A, they remove any dependence of the trained network on the input image v itself.

Dev: That dependency removal is key for making it plug-and-play across different inputs without needing to retrain for every single image.

Taro: So this moves us closer to having a generalized data-fitting agent, rather than one that’s tuned specifically for a certain input image space.

Rosa: Right, and the paper's limitation is pretty clear: they say training and verifying this deep PPM for non-linear operators is still future work.

Dev: They also highlight that while it works well on linear models, extending it reliably to more complex non-linear systems is the next big challenge for this approach.

Taro: So we've got a powerful tool that handles the inner optimization step much faster and with high accuracy when you have a linear operator, but we still need to figure out how to make it robust for anything that isn't linear yet.

Rosa: So, in summary, this paper introduces the deep PPM as a self-supervised network trained on noise residuals that can replace expensive inner solvers with a fast neural network evaluation for inverse problems where analytical solutions are unavailable.

Dev: It achieves one percent relative error across deblurring, super-resolution, and XCT, and for XCT it provides an eighteen times speedup over the Conjugate Gradient loop.

Taro: The implication is that we can run much faster iterative reconstruction methods in practice when the cost of evaluating the proximal map is your biggest bottleneck.

Rosa: It’s a very practical piece of research because it shows how to inject deep learning into those hard, computationally expensive steps of traditional inverse imaging algorithms.

The paper's summary: Rosa: So, to wrap up what we just talked about, this paper is about taking that really hard part of iterative reconstruction—the data-fitting proximal map—and turning it into a neural network.

Dev: Right, so instead of having to solve a massive optimization problem at every single step of the main reconstruction loop, you train an AI model to give you that answer almost instantly.

Rosa: Exactly. The core idea is this deep pseudo-proximal map, or PPM, which they call it because it’s basically the best guess for the image given your measurements and a prior image.

Dev: They frame it as finding the minimum mean square error estimate of a synthetic probabilistic model instead of just doing that hard minimization directly.

Rosa: And what makes this special is how they train it, Dev, because they don't need any real training images to teach the network how to do this estimation.

Dev: They use self-supervised training, meaning the network only learns by minimizing a mean square error loss using just sampled Gaussian noise and the operator itself.

Rosa: That’s a big deal for real applications because getting perfect ground truth images for every possible scenario is just not feasible in practice.

Dev: It extends this to any forward operator, as long as you can evaluate that forward model and its transpose, and when the operator is linear, it's mathematically proven to be equivalent to the original map.

Rosa: That equivalence when the operator is linear gives us a solid theoretical foundation for using this neural network in reconstruction algorithms.

Dev: And we see it working on three different kinds of problems—Gaussian deblurring, four times super-resolution, and X-ray computed tomography.

Rosa: They show that for deblurring and super-resolution, the deep PPM gets the final reconstruction within one percent relative mean squared error compared to the original reference map.

Dev: But then they move to XCT, where there is no easy analytical solution at all, and they say it still achieves that one percent error.

Rosa: And for XCT specifically, this is where the real efficiency gain comes in. They replace a slow inner conjugate gradient loop with just one network evaluation.

Dev: That single evaluation takes about thirty-eight point seven milliseconds for XCT, which is eighteen times faster than running that conjugate gradient solver at all.

Rosa: It also shows that across all three tasks, when you look at the reconstruction quality in terms of the mean absolute error comparison, the deep PPM is within one percent of the reference result.

Dev: So what this really means for us in real-time control or complex imaging systems is that we can inject deep learning into those computationally expensive steps without sacrificing accuracy.

Rosa: It shows how you can make iterative reconstruction methods much faster when evaluating the data-fitting map is your biggest bottleneck.

Dev: The limitation they mention is that while it works well for linear operators, training and verifying this for more complex non-linear models is still future work.

Rosa: So the path forward is figuring out how to make this agent robust enough to handle those messy non-linear systems we see in the real world.

Dev: It’s a powerful tool that handles the inner optimization step much faster and with high accuracy for linear cases, but we still need to figure out how to make it reliable for anything that isn't linear yet.

Rosa: We'll keep an eye on that non-linear extension, because if we can get this agent working there, it changes how fast we can process complex physical data.

The paper's improvements: Rosa: So we're talking now about what the authors suggest they should do next with this deep pseudo-proximal map, right? They are looking at where this research goes from here and what problems they still have to tackle.

Dev: Exactly. They admit that while the method works really well for linear operators, extending it reliably to non-linear forward models is still future work.

Rosa: That makes sense because those non-linear systems are exactly where we see most of the complex physical reality out there in things like medical imaging or atmospheric modeling.

Dev: And they acknowledge that training and verifying this deep PPM for those non-linear operators is the next big challenge they need to solve.

Rosa: It’s a fair point because if you can't handle non-linearity, you're stuck dealing with systems that are way more complicated than just a simple linear transformation.

Dev: They focus on training and verifying this deep PPM for non-linear operators as the subject of future work, which means they have a clear roadmap for the next phase of development.

Rosa: It sounds like they’ve set the stage there, showing that their current success is tied to linear models first.

Dev: That's what it is. The paper establishes that this method is very effective as a plug-and-play agent when you have a linear operator, which gives us confidence in its design.

Rosa: So even though they can't do non-linear right now, the structure of the deep PPM seems robust enough to handle those harder problems eventually.

Dev: They want to see if this approach can be adapted to those more complex scenarios where we don't have simple analytical solutions for the data-fitting map anymore.

Rosa: If they can solve that, it could mean we can use this kind of learned agent across a much wider range of physical measurements.

Dev: It would open up possibilities for reconstruction in areas like advanced materials science or complex fluid dynamics where the forward model is inherently non-linear.

Conclusion: Rosa: So we're wrapping up this look at "Deep Pseudo-Proximal Map: A Self-Supervised Data-Fitting Agent for Iterative Reconstruction." Basically, this paper shows how to use a self-supervised neural network to replace a slow inner loop solver with a fast network evaluation.

Dev: Right, so the main takeaway is that you can get one percent error across deblurring, super-resolution, and even X-ray computed tomography reconstruction.

Rosa: That’s pretty impressive accuracy when you consider the computational savings for those iterative methods.

Dev: It really changes things for us in real-time loops because it cuts down on latency significantly during the data-fitting step.

Taro: I just want to push on what this means for real-world failure modes, Rosa, does this AI agent handle when the world misbehaves in a way that isn't just Gaussian noise?

Rosa: That’s a good question, Taro. They did flag that the method is most reliable when the residual depends only on sampled Gaussian noise and the operator itself.

Dev: They stated that by formulating it in residual form, they remove any dependence of the trained network on the input image itself, which makes it much more versatile for different inputs.

Taro: So even if you throw something entirely new at this system, like a signal with completely different noise characteristics or distortions from a physical process not covered in training, does it just break?

Rosa: Well, they did say that training and verifying this deep PPM for non-linear operators is still future work.

Dev: Exactly. So the paper proves the concept beautifully for linear systems but points to non-linearity as the next major hurdle for this approach.

Taro: I'm interested in seeing those non-linear extensions, because that’s where most of the complex physical measurements we deal with actually live.

Rosa: Yeah, so this paper is a solid step forward by proving that you can inject deep learning into those expensive parts of traditional inverse imaging algorithms.

Dev: It gives us a way to potentially run much faster iterative reconstruction methods when evaluating the data-fitting map is your biggest bottleneck for latency.

Purdue University · Oak Ridge National Laboratory

eess.IV

Submitted: 2026-09-25

Updated: 2026-11-04

Comments: To appear in 2026 Asilomar Conference on Signals, Systems, and Computers

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 89/100

The gist: Iterative algorithms for inverse imaging problems split reconstruction into alternating subproblems, one of which requires evaluating the data-fitting proximal map <ref:2609.30727#pg2> To address…

Key concepts

Data-fitting Proximal Map
This is a mathematical tool used in iterative reconstruction to find the best image estimate by balancing two competing goals: fitting the measured data (like an MRI scan) and keeping the image close to a desired prior (like smoothness). It is central to defining how much an image should change at each step of the reconstruction.
Pseudo-Proximal Map (PPM)
The PPM reformulates the complex data-fitting map as a minimum mean square error estimate of a synthetic probabilistic model. It is implemented as a deep neural network trained only on sampled noise, allowing it to act as a fast, learned surrogate for the original map without needing real training images.
Residual Form
The PPM is defined in residual form: the estimated correction is calculated by adding a network output to the current image estimate. This formulation is powerful because the residual only depends on sampled noise and the operator itself, meaning it doesn't need ground-truth image data during training.

Terminology

Summary

Iterative algorithms for inverse imaging problems split reconstruction into alternating subproblems, one of which requires evaluating the data-fitting proximal map <ref:2609. p. 030727#pg2> To address this, we propose the pseudo-proximal map (PPM), a reformulation of the data-fitting proximal map as the minimum mean square error estimate of a synthetic probabilistic model The deep PPM can be trained for any operator for which the forward model A and its transpose AT can be evaluated, with provable equivalence to the proximal map when A is linear <ref:2609. p. 030727#pg2> We validate the deep PPM on Gaussian deblurring and 4× super-resolution, where the proximal map has an analytic solution, and on X-ray computed tomography (XCT), where no analytic solution exists <ref:2609. p. 030727#pg2> Used as the data-fitting agent within an iterative reconstruction method, the deep PPM reproduces the reference reconstruction to within 1% NRMSE for all three operators, and for XCT, it replaces the inner conjugategradient loop with a single network evaluation that is 18× faster <ref:2609. p. 030727#pg2>.

Core Formulation of the PPM

The data-fitting proximal map is defined as Fσ,γ(v; y) = arg min x 1/2σ squared y − A(x)2 + 1/2γ squared x − v2 <ref:2609.30727#pg5> The key observation motivating our proposed PPM is that Fσ,γ(v; y) can be interpreted as the maximum a posteriori (MAP) estimate under a Gaussian prior centered at v We define the PPM F˜σ,γ(v; y) as the corresponding conditional expectation of X˜ given Y˜ = y, i. p. 1e., F˜σ,γ(v; y) = Ev,γ[X˜ Y˜ = y] When A is linear, the posterior is Gaussian, so Fσ,γ(v; y) and F˜σ,γ(v; y) are equivalent

Self-Supervised Training Mechanism

We can implement the PPM as a network trained to minimize the mean square error (MSE) loss, since the MMSE estimator is equivalent to the conditional expectation We define the deep PPM F˜θσ,γ(v; y) in residual form as F˜θσ,γ(v; y) = v + H˜ θσ,γ(z), where z = AT (y − Av) is the back-projected residual and H˜ θσ,γ is a network with weights θ The weights are trained to minimize the MSE loss given by L(θ) = 1/K sum k=0 γϵk − H˜ θσ,γ(zk)2 Importantly, this residual depends only on sampled Gaussian noise and the operator A

Validation and Performance Metrics

We compare the deep PPM against CG and GD, each run for at most 50 iterations with relative convergence tolerance 10−6 <ref:2609.30727#pg5> For deblurring and SR, the Res-NRMSE with respect to the reference proximal map is approximately 6. p. 20% for both operators <ref:2609. p. 030727#pg2> For XCT, the Res-NRMSE with respect to CG is 11. p. 28%, about twice that of the other operators <ref:2609. p. 030727#pg2> Within MACE, the reconstruction obtained with the deep PPM differs from the reconstruction with the reference proximal map by less than 1% on average for all three tasks <ref:2609. p. 230727#pg3>

Computational Efficiency Gains

For XCT, however, the deep PPM replaces a 710 ms CG solve (6.0 s for GD) with a 38.7 ms evaluation, an 18× reduction in the cost of each data-fitting step <ref:2609. p. 230727#pg5> Over a full MACE run the speedup is 2. p. 23× over CG and 14.5× over GD <ref:2609. p. 030727#pg5> For deblurring and SR this offers no advantage, since the analytic solution and CG each take only a few milliseconds; these tasks serve as verification benchmarks rather than speed benchmarks <ref:2609. p. 230727#pg2>

Conclusion on Applicability

Since the formulation is well-defined for any A, the PPM extends naturally to nonlinear forward models, where data-fitting proximal maps are most expensive Training and verifying the deep PPM for non-linear operators is the subject of future work The deep PPM reconstruction matches the reference reconstruction to within 1% NRMSE on average across all three operators, and for XCT it replaces the inner CG loop with a network evaluation that is 18× faster <ref:2609.30727#pg2>.

--- Page 1 ---

Deep Pseudo-Proximal Map: A Self-Supervised Data-Fitting Agent for Iterative Reconstruction Haley Duba-Sullivan Purdue University West Lafayette, IN, USA hduba@purdue.edu Emma J. p. 0 Reid Oak Ridge National Laboratory Oak Ridge, TN, USA reidej@ornl. p. 0gov Charles A. p. 0 Bouman Purdue University West Lafayette, IN, USA bouman@purdue. p. 0edu Gregery T. p. 0 Buzzard Purdue University West Lafayette, IN, USA buzzard@purdue. p. 0edu Abstract—Iterative algorithms for inverse imaging problems split reconstruction into alternating subproblems, one of which requires evaluating the data-fitting proximal map <ref:2609. p. 030727#pg2> In many practical cases, this proximal map has no analytic solution and so must be approximated by an inner loop solver at every iteration, compounding the cost of the outer reconstruction loop <ref:2609. p. 030727#pg3> To address this, we propose the pseudo-proximal map (PPM), a reformulation of the data-fitting proximal map as the minimum mean square error estimate of a synthetic probabilistic model <ref:2609. p. 030727#pg4> We implement the deep PPM as a self-supervised neural network trained only on sampled Gaussian noise, requiring no groundtruth training images <ref:2609. p. 030727#pg5> The deep PPM can be trained for any operator for which the forward model A and its transpose A T can be evaluated, with provable equivalence to the proximal map when A is linear We validate the deep PPM on Gaussian deblurring and 4× super-resolution, where the proximal map has an analytic solution, and on X-ray computed tomography (XCT), where no analytic solution exists Used as the data-fitting agent within an iterative reconstruction method, the deep PPM reproduces the reference reconstruction to within 1% NRMSE for all three operators, and for XCT, it replaces the inner conjugategradient loop with a single network evaluation that is 18× faster Index Terms—Proximal map, inverse imaging problems, plugand-play, consensus equilibrium, self-supervised learning

--- Page 2 ---

Inverse imaging problems require inverting a physical measurement system modeled as y = A(x) + σW, (1) where A: R N → RM is the forward model, σ > 0 is the measurement noise standard deviation, and W ∼ N (0, I) is additive white Gaussian noise (AWGN) <ref:2609.30727#pg2> The data-fitting proximal map for this forward model is defined for γ > 0 as Fσ,γ(v; y) = arg min x 1/2σ squared y − A(x)2 + 1/2γ squared x − v2 <ref:2609. p. 130727#pg5> where v ∈ R N is any input image <ref:2609. p. 130727#pg5> The key observation motivating our proposed PPM is that Fσ,γ(v; y) can be interpreted as the maximum a posteriori (MAP) estimate under a Gaussian prior centered at v We define the synthetic probabilistic model given by X˜ = v + γE˜; Y˜ = A(X˜) + σW, ˜ (3) where Ẽ, W̃ ∼ N (0, I) are independent AWGN vectors Under this synthetic model, Fσ,γ(v; y) is the MAP estimate of X˜ given Y˜ = y, i. p. 1e., Fσ,γ(v; y) = arg min x

2 − log p(x Y˜ = y). (4) We define the PPM F˜σ,γ(v; y) as the corresponding conditional expectation of X˜ given Y˜ = y, i. p. 1e., F˜σ,γ(v; y) = Ev,γ[X˜ Y˜ = y] When A is linear, the posterior is Gaussian, so Fσ,γ(v; y) and F˜σ,γ(v; y) are equivalent

--- Page 3 ---

We can implement the PPM as a network trained to minimize the mean square error (MSE) loss, since the MMSE estimator is equivalent to the conditional expectation [14] We define the deep PPM F˜θσ,γ(v; y) in residual form as F˜θσ,γ(v; y) = v + H˜ θσ,γ(z), where z = AT (y − Av) is the back-projected residual and H˜ θσ,γ is a network with weights θ The weights are trained to minimize the MSE loss given by L(θ) = 1/K sum k=0 γϵk − H˜ θσ,γ(zk)2 When A is linear, the residual satisfies y − Av = γAϵ + σw Importantly, this residual depends only on sampled Gaussian noise and the operator A. Thus, by formulating the deep PPM in residual form, we remove any dependence of the trained network on the input image v

--- Page 4 ---

A. Forward Operators We evaluate the deep PPM on Gaussian deblurring and 4× SR, both of which admit an analytic solution to the data-fitting proximal map that can be efficiently computed in the Fourier domain [15] <ref:2609. p. 130727#pg5> We call this the analytic proximal map, and use it as the reference proximal map <ref:2609. p. 130727#pg5> For Gaussian deblurring, A is convolution with a 7 × 7 Gaussian kernel of standard deviation 1. p. 10 <ref:2609. p. 030727#pg5> For 4× SR, A is convolution with a 13 × 13 Gaussian kernel of standard deviation 2. p. 10 followed by 4×spatial subsampling <ref:2609. p. 130727#pg5> We also evaluate the deep PPM on XCT reconstruction, which does not admit an analytic solution to the data-fitting proximal map <ref:2609. p. 130727#pg5> In this case, we use a CG solution run to relative tolerance 10−6 as the reference proximal map <ref:2609. p. 130727#pg5> For XCT reconstruction, A is a parallel-beam projection operator with uniformly spaced angles between 0 and 180 degrees, implemented with the ASTRA toolbox [16] <ref:2609. p. 130727#pg5>

--- Page 5 ---

B. Training Details We train one deep PPM network per operator, fixing σ = 0. p. 101 and γ = 0.05 for all tasks The network H˜ θσ,γ is built on a DRUNet architecture [17] that takes the back-projected residual z as input and outputs the estimated correction γϵ Training pairs (zk, γϵk) are generated on the fly, with ϵk and wk drawn fresh at every step, so no image data of any kind enters training Each epoch consists of 10,000 training samples at the operator’s native size (256×256 for deblurring and SR, 128 × 128 for XCT) We minimize the MSE loss given in (10) with Adam [18] for 100 epochs at batch size 96, with an initial learning rate of 10−4, five epochs of linear warmup, and cosine annealing to 10−6

--- Page 6 ---

C. Testing Details For deblurring and SR, we use 50 crops of size 256 × 256 from the DIV2K validation set [19] We compare the deep PPM against CG and GD, each run for at most 50 iterations with relative convergence tolerance 10−6 <ref:2609. p. 230727#pg5> All three proximal map variants are evaluated first as standalone proximal maps and then as the data-fitting agent within MACE [3], [4] with 10 iterations using a BM3D denoiser [21] as the prior agent <ref:2609. p. 230727#pg5> For a proximal map variant P, we report the residual NRMSE with respect to the reference proximal map, i. p. 2e. Res-NRMSE = (P(v; y) − v) − (Fσ,γ(v; y) − v) / Fσ,γ(v; y) − v <ref:2609. p. 230727#pg2>

--- Page 7 ---

Table I: Quantitative comparison of each proximal map variant against the reference proximal map, both as a standalone map and as a data-fitting agent in MACE. The reference is the analytic proximal map for deblurring and SR and the CG solution for XCT <ref:2609. p. 230727#pg2> For Res-NRMSE / NRMSE, “ref” and “GT” denote error relative to the reference proximal map and to ground truth, respectively <ref:2609. p. 230727#pg2> Measured errors are averaged over 50 DIV2K crops (deblurring, SR) or 28 LoDoPaB-CT slices (XCT) <ref:2609. p. 230727#pg2>

--- Page 8 ---

Figure 1 compares outputs of the deep PPM with the reference proximal map for each operator. In all three cases, the deep PPM output is visually indistinguishable from the reference proximal map, supported by the small absolute error maps <ref:2609. p. 230727#pg2> For deblurring, the error is spatially unstructured noise <ref:2609. p. 230727#pg2> For SR, the error concentrates on edges and periodic texture, which is exactly the high-frequency content that the downsampling operator suppresses and that the correction must restore <ref:2609. p. 230727#pg2> For XCT the error is larger and spatially structured, concentrating along streak-like features and near the boundary of the reconstruction region, consistent with the higher ResNRMSE in Table I <ref:2609. p. 230727#pg2>

--- Page 9 ---

Figure 2 compares MACE reconstructions using the reference proximal map and the deep PPM as the data-fitting agent, with a BM3D prior after 10 iterations The two reconstructions are visually indistinguishable in all three cases

--- Page 10 ---

Figure 3 plots NRMSE of MACE iterates using each proximal map variant, plotted against wall-clock time. For deblurring and SR the deep PPM curve is indistinguishable from the analytic, CG, and GD curves at every iteration, and all four finish at the same time since BM3D accounts for nearly all of the per-iteration cost For XCT the deep PPM tracks the CG trajectory to the same equilibrium in 4. p. 23 s, compared with 9. p. 27 s for CG and 62 s for GD The deep PPM is provably equal to the datafitting proximal map when A is linear, and the deep PPM closely matches the analytic proximal maps for deblurring and SR as well as the CG proximal map for XCT <ref:2609. p. 330727#pg2>.

--- Page 3 ---

We can implement the PPM as a network trained to minimize the mean square error (MSE) loss, since the MMSE estimator is equivalent to the conditional expectation [14] We define the deep PPM F˜θσ,γ(v; y) in residual form as F˜θσ,γ(v; y) = v + H˜ θσ,γ(z), where z = AT (y − Av) is the back-projected residual and H˜ θσ,γ is a network with weights θ The weights are trained to minimize the MSE loss given by L(θ) = 1/K sum k=0 γϵk − H˜ θσ,γ(zk)2 When A is linear, the residual satisfies y − Av = γAϵ + σw Importantly, this residual depends only on sampled Gaussian noise and the operator A.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Deep PPM can replace inner conjugategradient loop with a single network evaluation that is 18 times faster for XCT reconstruction, significantly reducing computational cost during iterative reconstruction.

  2. The deep PPM can reproduce the reference reconstruction to within 1% NRMSE for all three operators, ensuring high accuracy in image restoration tasks like deblurring, super-resolution, and XCT.

  3. MACE reconstructions using the deep PPM closely match the reference reconstruction to within 1% NRMSE on average across all three operators, improving overall reconstruction quality compared to standard iterative methods.

  4. The deep PPM formulation can be incorporated into reconstruction algorithms with arbitrary image priors because it is provably equivalent to the data-fitting proximal map when A is linear, and its structure allows for residual correction training without ground-truth images for linear operators.

Sources

Related papers