Unrolling Nonconvex Graph Total Variation for Image Denoising

Fei Chen; Gene Cheung; Ivan Selesnick; Songlin Wei

REVIEW 4 major objections 5 minor 35 references

Unrolling Nonconvex Graph Total Variation for Image Denoising

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read By replacing total variation with a graph-domain Huber penalty and choosing its parameter via Gershgorin circles, this paper constructs a convex denoising objective that an unrolled ADMM network can solve with only 0.10M parameters.

desk verdict A useful unrolled denoiser with a clean GCT-based parameter selection, but the graph-Huber Moreau identity is wrong for weighted edges and the convexity claim only holds for a fixed x'. read the letter →

arxiv 2506.02381 v1 pith:PICJTA4X submitted 2025-06-03 eess.IV cs.CV

Songlin Wei , Gene Cheung , Fei Chen , Ivan Selesnick This is my paper

classification eess.IVcs.CV

keywords imagedenoisingnon-convexgraphtotalvariationconvexregularizationHuberfunctionMoreauenvelopeGershgorincircletheoremADMMalgorithmunrolling

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes a non-convex graph total variation (NC-GTV) term for image denoising and argues that, paired with an $\ell_2$ fidelity term, it keeps the whole objective convex rather than introducing spurious local minima. The regularization is built from a new graph variant of the Huber function, $S_a(x',x)=x^{\top}L^p_a(x')x$, which is interpretable as a Moreau envelope. The load-bearing quantity is the graph-Huber parameter $a$: the objective in Eq. (13) is convex whenever $I-\mu L^p_a(x')$ is positive semidefinite, and the paper computes the largest such $a$ efficiently from Gershgorin disc radii. Solving the convex objective with a linear-time ADMM scheme and unrolling it into a two-layer network yields 0.10M trainable parameters while matching a 0.56M-parameter deep CNN denoiser at $\sigma=30$ and beating it under noise-variance mismatch. The point is a compact, interpretable bridge between model-based convexity guarantees and data-driven denoising performance.

What carries the argument

The central object is the graph Huber function $S_a(x',x)=x^{\top}L^p_a(x')x$, a quadratic form in a signal-dependent penalty graph Laplacian $L^p_a$. For each edge, the penalty weight interpolates between an $\ell_2$ regime for small differences and an $\ell_1$-like regime for large differences, so the whole term approximates a graph sum of Huber functions under the proximity assumption $x'\approx x$. Because $S_a$ is a Moreau envelope, subtracting it from $\|Cx\|_1$ creates a minimax-concave type penalty whose curvature is controlled by $a$. The argument works through one matrix condition: $I-\mu L^p_a(x')\succeq 0$ makes the overall objective convex, and Gershgorin's Circle Theorem converts that condition into per-row disc-radius inequalities from which the largest admissible $a$ is computed.

What would settle it

At the network's output, compare the per-edge Huber sum on the left side of Eq. (10) with the quadratic form $x^{\top}L^p_a(x')x$ on the right using the last layer's $x'$; a large difference would mean the objective actually minimized is not the claimed graph-Huber objective. A second check is to increase $a$ above the computed $a^*$ and see whether PSNR improves, which would suggest the positive-semidefiniteness bound is not the binding constraint in practice.

Watch

Extended reading notes

Core claim

The central claim is that the denoising objective in Eq. (13), $\min_x \|y-x\|_2^2 + \mu\|Cx\|_1 - \mu x^{\top}L^p_a(x')x$, is convex — no extraneous local minima — even though the regularization is non-convex, provided the matrix $I-\mu L^p_a(x')$ is positive semidefinite. The paper defines the graph Huber function $S_a(x',x)=x^{\top}L^p_a(x')x$ with a signal-dependent penalty graph Laplacian $L^p_a$ and shows that it can be written as the Moreau envelope $a(\frac{1}{a}\|\cdot\|_1)^M(Cx)$, the natural graph analogue of the scalar Huber function. To pick the largest minimax-concave (MC) parameter $a$ that keeps the objective convex, it sorts the switching thresholds $1/|x'_i-x'_j|$, binary-searches the sorted list with the Gershgorin lower bound on the smallest eigenvalue of $I-\mu L^p_a$, and solves a per-row quadratic for the final $a^*$. The optimization is performed by ADMM, whose $x$-update is a sparse positive-definite linear system solved by conjugate gradient in linear time and whose $z$-update is soft-thresholding. Unrolling this ADMM for $T=2$ layers gives a network of about 0.10M parameters that reaches 29.60 dB PSNR on the RENOIR dataset at $\sigma=30$ and 25.43 dB when trained at $\sigma=30$ and tested at $\sigma=50$, the best among the compared methods.

Load-bearing premise

The graph Huber sum in Eq. (10) is valid only under the assumption that the current estimate $x'$ is close to the true solution $x$ (the paper writes $x'_i\approx x_i$); the convexity guarantee and the Moreau-envelope interpretation of $S_a$ rest on that proxy, and the unrolled network never verifies it.

Editorial extensions

If this is right

Because the per-layer objective is convex, the unrolled network is not chasing spurious local minima during training; the model-based objective has a well-defined global minimization problem.
The $a^*$ formula is cheap enough to be recomputed at every layer from the learned graph, allowing the network to adapt its convexity margin without extra learned parameters.
The per-iteration cost of the underlying ADMM is linear, via conjugate gradient and soft-thresholding, so the unrolled network inherits light per-layer computation.
At $\sigma=50$ noise mismatch, the 25.43 dB PSNR reported for NC-GTV exceeds the unrolled graph-TV baseline and the deep priors, indicating the non-convex regularization buys robustness to noise-level error.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper's convexity guarantee is per-layer and fixed-point in nature: it holds for the $x'$ used in that layer, but the network never verifies that $x'$ stays close to the final solution, so the objective actually minimized may drift from the claimed one if the proximity assumption in Eq. (10) fails.
Because $S_a$ is a Moreau envelope, the construction is naturally related to proximal methods, and one could connect these unrolled layers to plug-and-play or diffusion-cascade denoisers; the paper does not develop that relation.
One testable extension is to let $a$ be a per-edge or per-layer trainable parameter rather than the Gershgorin-derived scalar; since the GCT bound is sufficient but not obviously necessary for positive semidefiniteness, training might find larger admissible values that preserve edges even better.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Worth a look, but with a real mathematical slip in a supporting claim. The core idea is solid: define a graph Huber regularizer via a signal-dependent Laplacian, pick the nonconvexity parameter with Gershgorin bounds so the one-step objective is convex, and unroll the ADMM solver into a 0.10M-parameter network. The GCT computation is the most original piece and it is parameter-free, not fitted to the denoising target. Results are modest but honest: comparable to CDnCNN at matched noise, best under noise mismatch, and with far fewer parameters.

The soft spots are real but not fatal. Eq. (12) is incorrect for weighted edges. The Moreau envelope of (1/a)||·||_1 at Cx gives, for an edge with |w_ij(x_i-x_j)| <= 1/a, a quadratic term (a/2) w_ij^2 (x_i-x_j)^2, not (a/2) w_ij (x_i-x_j)^2, and the switch happens at |x_i-x_j| = 1/(a w_ij), not 1/a. So the claimed equivalence between Sa(x',x) and the Moreau envelope only holds when all edge weights are 1. The stress-test also flags a misplaced factor a in Eq. (9); I don't see that — Eq. (9) is consistent with Eq. (10) as written. The real problem is Eq. (12). Since the GCT analysis is derived from Eq. (9), the convexity guarantee for the actual objective may survive, but the paper must correct Eq. (12) and stop claiming the Moreau interpretation in its current form.

Second, the convexity claim is explicitly for a fixed x'. The unrolled network updates x' between layers, so the overall procedure is a fixed-point iteration, not the minimization of a single convex functional. Calling it an objective with no extraneous local minima overstates what is proven. This should be reframed honestly.

Reproducibility is weaker than I would like: graph topology is not specified, code and data splits are not released, and there are no error bars or multiple noise levels. These are easy to fix and matter for a paper whose selling point is lightweight deployment.

Bottom line: the empirical result is plausible, the GCT trick is genuinely new, and the flaws are correctable. This deserves a serious referee, not a desk reject, but the authors need to fix the Moreau identity, clarify the fixed-point status, and release code. I would not cite it in its current form, but I would read the revision.

Referee Report

4 major / 5 minor

Summary. The paper proposes a non-convex graph total variation (NC-GTV) regularizer for image denoising. The regularizer is defined through a signal-dependent penalty graph Laplacian L^p_a(x') and a graph variant of the Huber function. The central theoretical claim is that, with a suitable parameter a computed via Gershgorin Circle Theorem, the objective in Eq. (13) is convex and therefore free of extraneous local minima. The authors solve the objective with an ADMM algorithm and unroll the iterations into a lightweight feed-forward network. Experiments on the RENOIR dataset show that the unrolled network achieves PSNR/SSIM competitive with CDnCNN and better than DeepGTV under noise variance mismatch, using only 0.10M parameters.

Significance. If the mathematical claims were correct, the paper would provide an interpretable and lightweight denoising network that combines nonconvex regularization with a convex objective, and the GCT-based parameter selection would be an interesting algorithmic contribution. The experimental part is a strength: the unrolled network is an order of magnitude smaller than CDnCNN and shows robustness to noise-level mismatch, and the comparison against DeepGTV is directly relevant. However, the graph-Huber and Moreau-envelope identities in Eqs. (9)-(12) are internally inconsistent, so the theoretical foundation of the approach is not currently established. The empirical results may still be useful as a heuristic network, but they do not support the paper's headline claims about convexity and the absence of extraneous local minima.

major comments (4)

[Section 3.1.2, Eqs. (9)-(10)] The large-difference branch of Eq. (9) contains an extra factor a. For an edge with d = x'_i - x'_j and x_i - x_j ≈ d, the contribution of the penalty Laplacian is w^p_{i,j}(x_i-x_j)^2 ≈ a w_{i,j}|d| - w_{i,j}/(2a), whereas Eq. (10) and the following paragraph claim w_{i,j}|x_i-x_j| - w_{i,j}/(2a). The extra factor a makes the edge contribution discontinuous at |d| = 1/a and breaks the intended match with the scalar Huber function in Eq. (6). Since the GCT computation in Section 3.3 is built on Eq. (9), the computed a* may be a bound for a quadratic proxy that is not the graph Huber function used in the experiments.
[Section 3.1.2, Eq. (12)] The claimed Moreau-envelope identity S_a(x',x) = min_v { ||v||_1 + (a/2)||Cx-v||_2^2 } is not correct for the incidence matrix C defined in Eq. (3) unless all edge weights are unity. For a single edge with weight w_{i,j}, the scalar z_m = w_{i,j}(x_i-x_j) enters the Moreau envelope of the l1 norm, giving a/2 w_{i,j}^2 d^2 for |d| <= 1/(a w_{i,j}) and w_{i,j}|d| - 1/(2a) otherwise. This differs from Eq. (10) both in the quadratic coefficient and in the threshold. Moreover, for fixed x' the left-hand side is a global quadratic in x, while the right-hand side is piecewise quadratic; the equality can hold at best as a local approximation under x' ≈ x, not as an identity.
[Sections 3.2, 3.4, and 4.2] The convexity guarantee is stated for a fixed x': the condition that I - mu L^p_a(x') is PSD makes the quadratic part of Eq. (13) convex for that particular x'. In the ADMM iterations and in the unrolled network, x' is updated at every layer using the previous estimate x^{(t-1)}. The paper does not prove that this sequence of changing quadratic proxies converges to the minimizer of a single convex objective, nor that the final output is a stationary point of the original objective. Consequently, the abstract's claim of 'no extraneous local minima' is not established for the actual algorithm that is implemented and evaluated.
[Section 3.3, Eqs. (16)-(19)] The GCT-based derivation of a* uses row-sum formulas that are explicitly derived from the 'otherwise' branch of Eq. (9). Because that branch contains the extra factor a identified above, the value a* computed by solving the quadratic in Eq. (19) does not necessarily ensure PSDness of I - mu L^p_a(x') for the graph Huber function as described in Eq. (10). The parameter selection step is load-bearing for the convexity claim, so it needs to be re-derived from a corrected edge-weight definition.

minor comments (5)

[Section 4.2] The initialization is incomplete: the text states z^{(0)} = C y and xi^{(0)} = 0, but does not state the initial value x^{(0)} used to compute the signal-dependent weights in the first layer; presumably x^{(0)} = y, but this should be written explicitly.
[Section 3.4.1] The claim that the linear system in Eq. (23) is solved 'in linear time' via conjugate gradient is too strong; CG complexity depends on the number of iterations, which in turn depends on the conditioning of the coefficient matrix, so a linear-time guarantee is not established.
[Section 1, contribution 3] There is a typo: 'feed-foward' should be 'feed-forward'.
[Section 2.1] The graph is called 'positive undirected' but edge weights are only required to be non-negative; if 'positive' is intended, the text should say w_{i,j} > 0 rather than >= 0.
[Section 5.1] The feature dimension K for the Mahalanobis metric M and the exact input patch size for the CNN are not specified; without these details the parameter count in Table 1 is difficult to reproduce.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the convexity guarantee is a computed sufficient condition, not a fitted prediction; the Moreau-envelope claim is a mathematical consistency issue, not a circular step.

full rationale

The paper's derivation chain is self-contained. The graph Huber function is explicitly defined in Eq. (9)-(11), and the convexity condition is derived as the PSD-ness of I - mu L^p_a(x'), with the parameter a* computed via Gershgorin Circle Theorem in Eqs. (14)-(19). This a* is a sufficient condition computed from the current estimate x', not a parameter fitted to denoising targets or to the reported PSNR/SSIM results. The unrolled network parameters (mu, rho, gamma, lambda, CNN weights) are trained by MSE on data, which is conventional supervised learning and does not make the experimental claims circular. The paper cites prior work by overlapping authors, notably [10] (Selesnick et al.) for the scalar MC/Huber definitions and [25] (Vu, Cheung, Eldar) for baseline/setup, but these citations supply standard definitions, numerical stabilization, and experimental baselines; they do not carry the load of the convexity derivation or the performance comparison. The comparison against external methods (CBM3D, CDnCNN, DGLR, DGTV) provides independent benchmarks. One caveat outside circularity: the Moreau-envelope equality in Eq. (12) is not consistent with the weighted-edge definition in Eq. (9)/(10) unless all edge weights equal 1, since Cx has entries w_ij(x_i-x_j), so the scalar Huber threshold and quadratic scaling involve w_ij^2 rather than w_ij. This is a mathematical correctness or consistency concern, not a circular reduction of the paper's claims to its inputs.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The method relies on standard tools (Gershgorin theorem, graph Laplacian PSD, Moreau envelopes) and on two domain assumptions: the current estimate approximates the solution during ADMM or network layers, and the learned exponential kernel captures image structure. No physical entities are invented; the graph Huber function is a new mathematical construct with no external validation.

free parameters (6)

a = computed per layer via GCT, not trained
Convexity parameter in the graph Huber function; chosen as the largest value satisfying the Gershgorin lower bound on the smallest eigenvalue of I - mu L^p_a(x').
mu = learned, value not reported
Regularizer weight in Eq. (13); trained end-to-end in the unrolled network.
rho = learned, value not reported
ADMM penalty parameter in Eq. (21); trained end-to-end.
gamma = learned, value not reported
Proximal gradient step size in Eq. (25); trained end-to-end.
lambda = learned, value not reported
Proximal mapping parameter in Eq. (27); trained end-to-end.
CNN features and metric M = learned
Feature extractor and Mahalanobis metric define edge weights in Eqs. (30)-(31); trained end-to-end.

assumptions (5)

standard math Gershgorin Circle Theorem gives a lower bound on the smallest eigenvalue of a symmetric matrix.
Used in Section 3.3 to select a* ensuring I - mu L^p_a is PSD.
standard math A combinatorial graph Laplacian with non-negative edge weights is positive semi-definite.
Invoked in Section 2.1 to justify the PSD property of graph Laplacians.
standard math The scalar Huber function is the Moreau envelope of the absolute value function.
Background used in Eqs. (7)-(8) to motivate the graph construction.
domain assumption The current estimate x' is approximately equal to the true signal x, so the signal-dependent quadratic form approximates the intended graph Huber function.
Stated in Eq. (10) as assuming x'_i approximately equals x_i; not verified for the unrolled network where x' changes between layers.
domain assumption The learned graph with exponential Mahalanobis-distance edge weights captures the relevant image structure for denoising.
Edge weights are defined in Eqs. (30)-(31); the choice of graph topology is unspecified and its adequacy is not validated.

invented entities (1)

Graph Huber function Sa(x',x)
purpose: Define the non-convex penalty by subtracting a graph variant of the Huber function from GTV in Eq. (13).
A new mathematical construction in Eq. (11); its claimed equivalence to a Moreau envelope in Eq. (12) appears inconsistent with the weight formula in Eq. (9), and no external falsifiable handle is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Unrolling Nonconvex Graph Total Variation for Image Denoising." pith.science (2026). https://pith.science/paper/PICJTA4X

@misc{pith2026250602381,
  author       = {Pith},
  title        = {Pith review of: Unrolling Nonconvex Graph Total Variation for Image Denoising},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PICJTA4X}},
  note         = {Machine review of arXiv:2506.02381}
}

abstract

Conventional model-based image denoising optimizations employ convex regularization terms, such as total variation (TV) that convexifies the $\ell_0$-norm to promote sparse signal representation. Instead, we propose a new non-convex total variation term in a graph setting (NC-GTV), such that when combined with an $\ell_2$-norm fidelity term for denoising, leads to a convex objective with no extraneous local minima. We define NC-GTV using a new graph variant of the Huber function, interpretable as a Moreau envelope. The crux is the selection of a parameter $a$ characterizing the graph Huber function that ensures overall objective convexity; we efficiently compute $a$ via an adaptation of Gershgorin Circle Theorem (GCT). To minimize the convex objective, we design a linear-time algorithm based on Alternating Direction Method of Multipliers (ADMM) and unroll it into a lightweight feed-forward network for data-driven parameter learning. Experiments show that our method outperforms unrolled GTV and other representative image denoising schemes, while employing far fewer network parameters.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 31 canonical work pages

[1]

black boxes

INTRODUCTION Image denoising [1] is a well-studied basic restoration prob- lem that remains important today; other image restoration tasks like deblurring [2] can be addressed using general frameworks like thePlug-and-Play Alternating Direction Method of Multipliers(PnP-ADMM) [3] that employ an image denoiser as a building block, and popular genera- tive ...

work page 2025
[2]

Unrolling Nonconvex Graph Total Variation for Image Denoising

We design a graph Huber functionSa(x′,x)—extension of the scalar Huber functionsa(x)to the graph domain— as a quadratic term using penalty graph LaplacianL p a that is amenable to fast optimization. Likes a(x), we showS a(x′,x)is interpretable as a Moreau envelope. arXiv:2506.02381v1 [eess.IV] 3 Jun 2025

work page Pith review arXiv 2025
[3]

tightest

Leveraging GCT, we efficiently compute the “tightest” parameteracharacterizing the graph Huber function that guarantees PSDness of coefficient matrixI−µL p a, and hence convexity of the optimization objective

work page
[4]

We unroll an ADMM algorithm into a lightweight in- terpretable feed-foward network for data-driven param- eter learning, with competitive denoising performance

work page
[5]

GSP Definitions A positive undirected graphG= (V,E,W)is defined by a node setV={1,

PRELIMINARIES 2.1. GSP Definitions A positive undirected graphG= (V,E,W)is defined by a node setV={1, . . . , N}and an edge setE={(i, j)} withM=|E|edges, where edge(i, j)∈ Ehas non-negative weightw i,j =W i,j, foradjacency matrixW∈R N×N . Undirected edges mean thatW i,j =W j,i, andWis sym- metric.Degree matrixD∈R N×N is a diagonal matrix with diagonal ent...

work page
[6]

Defining Graph Huber Function 3.1.1

PROBLEM FORMULATION 3.1. Defining Graph Huber Function 3.1.1. MC penalty and Huber Functions From [10], the scalarminimax-concave(MC) penalty func- tionϕ a(x) :R7→R + is defined as ϕa(x) = |x| −a 2 x2 if|x| ≤1 a1 2a o.w. (5) wherea >0is a non-negativeMC parameter. Note that for a= 0,ϕ a(x) =|x|,i.e., theℓ 1-norm. The MC penalty function can be written in ...

work page
[7]

In addition, we learn asimilarity graphGfrom data, so that incidence matrixCspecifying a positive graphGin (13) is properly defined

ALGORITHM UNROLLING We next unroll each iteration of our ADMM algorithm into a neural layer to compose a feed-forward network for data- driven parameter learning. In addition, we learn asimilarity graphGfrom data, so that incidence matrixCspecifying a positive graphGin (13) is properly defined. 4.1. Similarity Graph Learning We define a positive graphGwit...

work page
[8]

Experimental Setup We set the number of NC-GTV layers toT= 2.CNN f consists of4convolution layers

EXPERIMENTS 5.1. Experimental Setup We set the number of NC-GTV layers toT= 2.CNN f consists of4convolution layers. The first layer has3input channels and32output channels, while the last layer has32 and3. Except for the last layer, we usedReLU(·)after every convolutional layer. As done in [25], we evaluated our model on the RENOIR dataset and added Gauss...

work page

Show all 35 references

[9]

We show that our graph Huber function can be understood as a Moreau envelope

CONCLUSION Instead of convex total variation (TV) regularization term that underestimates signal discontinuities, we propose non-convex graph total variation (NC-GTV): combine a non-convex reg- ularization term defined by a graph variant of the Huber func- tion with anℓ 2-norm...
[10]

A tour of modern image filtering: New insights and methods, both practical and theoretical,

P. Milanfar, “A tour of modern image filtering: New insights and methods, both practical and theoretical,”IEEE Signal Pro- cessing Magazine, vol. 30, no. 1, pp. 106–128, 2013

2013
[11]

Graph-based blind image deblurring from a single photograph,

Y . Bai, G. Cheung, X. Liu, and W. Gao, “Graph-based blind image deblurring from a single photograph,”IEEE Transac- tions on Image Processing, vol. 28, no.3, pp. 1404–1418, 2019

2019
[12]

Plug-and-play ADMM for image restoration: Fixed-point convergence and applications,

S. H. Chan, X. Wang, and O. A. Elgendy, “Plug-and-play ADMM for image restoration: Fixed-point convergence and applications,”IEEE Transactions on Computational Imaging, vol. 3, no. 1, pp. 84–98, 2017

2017
[13]

Tutorial on diffusion models for imaging and vision,

Stanley H. Chan, “Tutorial on diffusion models for imaging and vision,” arXiv preprint arXiv:2403.18103, March 2024, Last revised January 2025

2024 arXiv
[14]

Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising,

K. Zhang, W. Zuo, Y . Chen, D. Meng, and L. Zhang, “Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising,”IEEE Transactions on Image Processing, vol. 26, no. 7, pp. 3142–3155, 2017

2017
[15]

Algorithm unrolling: Inter- pretable, efficient deep learning for signal and image process- ing,

V . Monga, Y . Li, and Y . C. Eldar, “Algorithm unrolling: Inter- pretable, efficient deep learning for signal and image process- ing,”IEEE Signal Processing Magazine, vol. 38, no. 2, pp. 18–44, 2021

2021
[16]

Graph signal processing: Overview, chal- lenges, and applications,

A. Ortega, P. Frossard, J. Kovacevic, J. M. F. Moura, and P. Vandergheynst, “Graph signal processing: Overview, chal- lenges, and applications,” inProceedings of the IEEE, May 2018, vol. 106, no.5, pp. 808–828

2018
[17]

Graph spectral image processing,

G. Cheung, E. Magli, Y . Tanaka, and M. Ng, “Graph spectral image processing,” inProceedings of the IEEE, 2018, vol. 106, no.5, pp. 907–930

2018
[18]

Edge-preserving and scale-dependent properties of total variation regularization,

D. Strong and T. Chan, “Edge-preserving and scale-dependent properties of total variation regularization,”Inverse Problems, vol. 19, no. 6, pp. S165, nov 2003

2003
[19]

Non-convex total variation regularization for convex denoising of signals,

I. Selesnick, A. Lanza, S. Morigi, and F. Sgallari, “Non-convex total variation regularization for convex denoising of signals,” Journal of Mathematical Imaging and Vision, vol. 62, no. 6, pp. 825–841, Jul 2020

2020
[20]

Convex 1-D total vari- ation denoising with non-convex regularization,

I. Selesnick, A. Parekh, and ˙I. Bayram, “Convex 1-D total vari- ation denoising with non-convex regularization,”IEEE Signal Processing Letters, vol. 22, no. 2, pp. 141–144, 2015

2015
[21]

Sparsity-inducing nonconvex nonseparable regularization for convex image processing,

Alessandro Lanza, Serena Morigi, Ivan W. Selesnick, and Fiorella Sgallari, “Sparsity-inducing nonconvex nonseparable regularization for convex image processing,”SIAM Journal on Imaging Sciences, vol. 12, no. 2, pp. 1099–1134, 2019

2019
[22]

Robust estimation of a location parameter,

P. J. Huber, “Robust estimation of a location parameter,”Ann. Math. Statist., vol. 35, no. 1, pp. 73–101, 1964

1964
[23]

H. B. Heinz and P. L. Combettes,Convex Analysis and Mono- tone Operator Theory in Hilbert Spaces, Springer, 2017

2017
[24]

R. S. Varga,Gershgorin and his circles, Springer, 2004

2004
[25]

Proximal algorithms,

N. Parikh and S. Boyd, “Proximal algorithms,” inFoundations and Trends in Optimization, 2013, vol. 1, no.3, pp. 123–231

2013
[26]

Image denoising by sparse 3-D transform-domain collaborative filter- ing,

K. Dabov, A. Foi, V . Katkovnik, and K. Egiazarian, “Image denoising by sparse 3-D transform-domain collaborative filter- ing,”IEEE Transactions on Image Processing, vol. 16, no. 8, pp. 2080–2095, 2007

2007
[27]

Graph Laplacian regularization for inverse imaging: Analysis in the continuous domain,

J. Pang and G. Cheung, “Graph Laplacian regularization for inverse imaging: Analysis in the continuous domain,” inIEEE Transactions on Image Processing, April 2017, vol. 26, no.4, pp. 1770–1785

2017
[28]

Random walk graph Laplacian based smoothness prior for soft decoding of JPEG images,

X. Liu, G. Cheung, X. Wu, and D. Zhao, “Random walk graph Laplacian based smoothness prior for soft decoding of JPEG images,”IEEE Transactions on Image Processing, vol. 26, no.2, pp. 509–524, February 2017

2017
[29]

Nonlocal discrete regularization on weighted graphs: A framework for image and manifold processing,

Abderrahim Elmoataz, Olivier Lezoray, and S ´Ebastien Bougleux, “Nonlocal discrete regularization on weighted graphs: A framework for image and manifold processing,” IEEE Transactions on Image Processing, vol. 17, no. 7, pp. 1047–1060, 2008

2008
[30]

Dual constrained TV-based regularization on graphs,

Camille Couprie, Leo Grady, Laurent Najman, Jean- Christophe Pesquet, and Hugues Talbot, “Dual constrained TV-based regularization on graphs,”SIAM Journal on Imaging Sciences, vol. 6, no. 3, pp. 1246–1273, 2013

2013
[31]

T. H. Cormen, C. E. Leiserson, and R. L. Rivest,Introduction to Algorithms, McGraw Hill, 1986

1986
[32]

Boyd and L

S. Boyd and L. Vandenberghe,Convex Optimization, Cam- bridge, 2004

2004
[33]

An introduction to the conjugate gradient method without the agonizing pain,

J. R Shewchuk, “An introduction to the conjugate gradient method without the agonizing pain,” Tech. Rep., USA, 1994

1994
[34]

Unrolling of deep graph total variation for image denoising,

Huy Vu, Gene Cheung, and Yonina C. Eldar, “Unrolling of deep graph total variation for image denoising,” inICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021, pp. 2050– 2054

2021
[35]

Deep graph Laplacian regularization for robust denoising of real im- ages,

Jin Zeng, Jiahao Pang, Wenxiu Sun, and Gene Cheung, “Deep graph Laplacian regularization for robust denoising of real im- ages,” in2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019, pp. 1759– 1768

2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

black boxes

INTRODUCTION Image denoising [1] is a well-studied basic restoration prob- lem that remains important today; other image restoration tasks like deblurring [2] can be addressed using general frameworks like thePlug-and-Play Alternating Direction Method of Multipliers(PnP-ADMM) [3] that employ an image denoiser as a building block, and popular genera- tive ...

work page 2025

[2] [2]

Unrolling Nonconvex Graph Total Variation for Image Denoising

We design a graph Huber functionSa(x′,x)—extension of the scalar Huber functionsa(x)to the graph domain— as a quadratic term using penalty graph LaplacianL p a that is amenable to fast optimization. Likes a(x), we showS a(x′,x)is interpretable as a Moreau envelope. arXiv:2506.02381v1 [eess.IV] 3 Jun 2025

work page Pith review arXiv 2025

[3] [3]

tightest

Leveraging GCT, we efficiently compute the “tightest” parameteracharacterizing the graph Huber function that guarantees PSDness of coefficient matrixI−µL p a, and hence convexity of the optimization objective

work page

[4] [4]

We unroll an ADMM algorithm into a lightweight in- terpretable feed-foward network for data-driven param- eter learning, with competitive denoising performance

work page

[5] [5]

GSP Definitions A positive undirected graphG= (V,E,W)is defined by a node setV={1,

PRELIMINARIES 2.1. GSP Definitions A positive undirected graphG= (V,E,W)is defined by a node setV={1, . . . , N}and an edge setE={(i, j)} withM=|E|edges, where edge(i, j)∈ Ehas non-negative weightw i,j =W i,j, foradjacency matrixW∈R N×N . Undirected edges mean thatW i,j =W j,i, andWis sym- metric.Degree matrixD∈R N×N is a diagonal matrix with diagonal ent...

work page

[6] [6]

Defining Graph Huber Function 3.1.1

PROBLEM FORMULATION 3.1. Defining Graph Huber Function 3.1.1. MC penalty and Huber Functions From [10], the scalarminimax-concave(MC) penalty func- tionϕ a(x) :R7→R + is defined as ϕa(x) = |x| −a 2 x2 if|x| ≤1 a1 2a o.w. (5) wherea >0is a non-negativeMC parameter. Note that for a= 0,ϕ a(x) =|x|,i.e., theℓ 1-norm. The MC penalty function can be written in ...

work page

[7] [7]

In addition, we learn asimilarity graphGfrom data, so that incidence matrixCspecifying a positive graphGin (13) is properly defined

ALGORITHM UNROLLING We next unroll each iteration of our ADMM algorithm into a neural layer to compose a feed-forward network for data- driven parameter learning. In addition, we learn asimilarity graphGfrom data, so that incidence matrixCspecifying a positive graphGin (13) is properly defined. 4.1. Similarity Graph Learning We define a positive graphGwit...

work page

[8] [8]

Experimental Setup We set the number of NC-GTV layers toT= 2.CNN f consists of4convolution layers

EXPERIMENTS 5.1. Experimental Setup We set the number of NC-GTV layers toT= 2.CNN f consists of4convolution layers. The first layer has3input channels and32output channels, while the last layer has32 and3. Except for the last layer, we usedReLU(·)after every convolutional layer. As done in [25], we evaluated our model on the RENOIR dataset and added Gauss...

work page