pith. sign in

arxiv: 2604.19314 · v1 · submitted 2026-04-21 · 💻 cs.CV · cs.NA· math.NA

Framelet-Based Blind Image Restoration with Minimax Concave Regularization

Heng Zhang , Reza Parvaz , Rui Yang This is my paper

Pith reviewed 2026-05-10 03:14 UTC · model grok-4.3

classification 💻 cs.CV cs.NAmath.NA
keywords blind image deblurringminimax concave penaltyframelet transformreweighted l1 regularizationimage restorationsparsity promotionnonconvex optimizationtotal variation
0
0 comments X

The pith

A framelet-based model with minimax concave penalty and reweighted l1 regularization approximates the l0-norm to restore blurred images while preserving edges and details.

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

The paper addresses blind image deblurring, where both the blur kernel and sharp image must be recovered from a corrupted observation. It replaces the intractable l0-norm with the minimax concave penalty to promote sparsity in framelet coefficients of image gradients, then adds a reweighted l1 term to cut estimation bias. A numerical algorithm solves the resulting nonconvex problem. Experiments on test images show the approach recovers fine structures more effectively than standard total variation methods.

Core claim

The proposed framelet-based model employs the minimax concave penalty to promote sparsity closer to the l0-norm within the total variation framework and incorporates a reweighted l1-norm to reduce estimation bias, allowing simultaneous estimation of the point spread function and the latent sharp image through a solvable optimization problem that better preserves edges and fine structures.

What carries the argument

Minimax concave penalty (MCP) applied to framelet coefficients, paired with a reweighted l1-norm term, which together approximate l0-norm sparsity promotion while remaining tractable for optimization.

If this is right

  • The nonconvex optimization problem for blind deblurring becomes solvable with a practical numerical algorithm.
  • Restored images exhibit improved edge preservation and recovery of fine details due to stronger sparsity promotion.
  • Estimation bias is reduced, resulting in better texture and detail fidelity in the output.
  • The framework demonstrates practical effectiveness across multiple test images.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same MCP-plus-reweighted-l1 combination could be tested on related inverse problems such as image denoising or inpainting.
  • Framelet bases may capture multi-directional features more effectively than wavelets in other sparsity-driven restoration tasks.
  • Analogous penalty designs might lower bias in sparsity-based methods outside imaging, such as in signal compression.

Load-bearing premise

The minimax concave penalty approximates the l0-norm more closely than alternatives and the reweighted l1 term reduces bias sufficiently to improve quality without introducing new artifacts or optimization instability.

What would settle it

Run the method on standard synthetic blurred images with known ground-truth sharp versions and compare PSNR or visual edge fidelity against l0-norm or standard TV blind deblurring baselines; failure to show improvement or presence of instability would refute the advantage.

Figures

Figures reproduced from arXiv: 2604.19314 by Heng Zhang, Reza Parvaz, Rui Yang.

Figure 1
Figure 1. Figure 1: Visual results on the Levien dataset: (a,e,i,m) clear images 1-4, respec [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visual results on the Text images: (a,c) blurred images; (b,d) restored latent [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visual results on the Köhler Dataset: (a) blurred image (image 1, kernel 5); [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual results on the Köhler Dataset: (a) blurred image (image 2, kernel 1); [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visual results on the boat and face images: (a,d) blurred images; (b,e) in [PITH_FULL_IMAGE:figures/full_fig_p018_5.png] view at source ↗
read the original abstract

Recovering corrupted images is one of the most challenging problems in image processing. Among various restoration tasks, blind image deblurring has been extensively studied due to its practical importance and inherent difficulty. In this problem, both the point spread function (PSF) and the underlying latent sharp image must be estimated simultaneously. This problem cannot be solved directly due to its ill-posed nature. One powerful tool for solving such problems is total variation (TV) regularization. The $\ell_0$-norm regularization within the TV framework has been widely adopted to promote sparsity in image gradients or transform domains, leading to improved preservation of edges and fine structures. However, the use of the $\ell_0$-norm results in a highly nonconvex and computationally intractable optimization problem, which limits its practical applicability. To overcome these difficulties, we employ the minimax concave penalty (MCP), which promotes enhanced sparsity and provides a closer approximation to the $\ell_0$-norm. In addition, a reweighted $\ell_1$-norm regularization is incorporated to further reduce estimation bias and improve the preservation of fine image details and textures. After introducing the proposed model, a numerical algorithm is developed to solve the resulting optimization problem. The effectiveness of the proposed approach is then demonstrated through experimental evaluations on several test images.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes a framelet-based variational model for blind image deblurring that replaces the ℓ0-norm with the minimax concave penalty (MCP) to promote sparsity in the framelet domain while incorporating a reweighted ℓ1 term to reduce bias. A numerical algorithm is derived to solve the resulting non-convex optimization problem, and the method is evaluated on several test images to demonstrate improved restoration quality over alternatives.

Significance. If the experimental gains can be reliably attributed to the MCP and reweighting choices rather than optimization artifacts, the work would offer a practical non-convex alternative to standard TV-ℓ0 or ℓ1 framelet models for blind deblurring, potentially improving edge and texture preservation in ill-posed inverse problems.

major comments (3)
  1. [§4] §4 (Optimization Algorithm): The iterative solver for the non-convex MCP-plus-reweighted-ℓ1 objective is presented without a convergence guarantee or any analysis of stationary points. Because the joint kernel-image estimation is already unstable, the absence of such analysis makes it impossible to attribute reported improvements to the regularizers rather than to particular initializations or early stopping.
  2. [§5] §5 (Experiments): No multi-start experiments, initialization-sensitivity study, or stability checks across random seeds are reported. Without evidence that the solver consistently reaches comparable minima, the visual and quantitative improvements cannot be confidently credited to the MCP approximation to ℓ0 or the bias-reduction effect of reweighting.
  3. [§3] §3 (Proposed Model), Eq. (7)–(9): The claim that MCP provides a closer approximation to the ℓ0-norm than alternatives is stated but not accompanied by a quantitative comparison (e.g., approximation error curves or recovery guarantees) that would justify its use over other non-convex penalties in the blind-deblurring setting.
minor comments (2)
  1. [Abstract, §5] The abstract and §5 omit concrete metrics, baseline methods, and error bars; these details should be added so readers can assess the magnitude of improvement.
  2. [§2, §3] Notation for the framelet transform and the reweighting schedule is introduced without a clear reference to the specific framelet basis or the exact reweighting formula used in each iteration.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive comments on our manuscript. We address each of the major comments below, indicating the revisions we plan to make to improve the paper.

read point-by-point responses

  1. Referee: [§4] §4 (Optimization Algorithm): The iterative solver for the non-convex MCP-plus-reweighted-ℓ1 objective is presented without a convergence guarantee or any analysis of stationary points. Because the joint kernel-image estimation is already unstable, the absence of such analysis makes it impossible to attribute reported improvements to the regularizers rather than to particular initializations or early stopping.

    Authors: We acknowledge that the manuscript does not include a formal convergence analysis for the non-convex optimization problem. The algorithm is constructed as an alternating minimization procedure, with the image update using the proximal mapping of the MCP and the kernel update incorporating reweighting. Although a complete proof of convergence to a stationary point is challenging due to the non-convexity and joint estimation, we will revise Section 4 to include a discussion of the conditions under which stationary points are reached and add plots demonstrating the convergence of the objective function in the experiments. This will provide better support for attributing the improvements to the proposed regularizers. revision: partial

  2. Referee: [§5] §5 (Experiments): No multi-start experiments, initialization-sensitivity study, or stability checks across random seeds are reported. Without evidence that the solver consistently reaches comparable minima, the visual and quantitative improvements cannot be confidently credited to the MCP approximation to ℓ0 or the bias-reduction effect of reweighting.

    Authors: We agree with the referee that additional validation of the solver's stability is necessary. In the revised manuscript, we will include new experiments in Section 5 consisting of multi-start runs with varied initializations for the blur kernel and different random seeds. We will report the average and standard deviation of the quantitative metrics (PSNR and SSIM) to demonstrate that the performance improvements are consistent across runs. revision: yes

  3. Referee: [§3] §3 (Proposed Model), Eq. (7)–(9): The claim that MCP provides a closer approximation to the ℓ0-norm than alternatives is stated but not accompanied by a quantitative comparison (e.g., approximation error curves or recovery guarantees) that would justify its use over other non-convex penalties in the blind-deblurring setting.

    Authors: The use of MCP is motivated by its theoretical properties as a non-convex penalty that approximates the ℓ0-norm more closely than the ℓ1-norm while maintaining computational tractability, as referenced in the literature. However, we did not provide a direct quantitative comparison in the current version. We will add to Section 3 a quantitative comparison, including approximation error curves for MCP versus other penalties such as ℓ1 and SCAD, to better justify its selection for the framelet-based model. revision: yes

Circularity Check

0 steps flagged

No circularity: forward model proposal, solver derivation, and experimental validation with no self-referential reductions.

full rationale

The paper proposes a framelet-based blind deblurring model that replaces ℓ0 with MCP plus reweighted ℓ1, derives an iterative numerical algorithm to solve the resulting non-convex problem, and reports experimental results on test images. No equation is shown to equal its own input by construction, no fitted parameter is relabeled as a prediction, and no load-bearing uniqueness claim rests on a self-citation chain. The derivation chain proceeds from modeling assumptions to algorithm to empirical checks without closing on itself.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The abstract relies on standard assumptions from convex optimization and image processing literature; no free parameters, new entities, or ad-hoc axioms are explicitly introduced or quantified.

axioms (2)
  • domain assumption The resulting non-convex optimization problem admits a practical numerical solution algorithm.
    Invoked when stating that a numerical algorithm is developed to solve the model.
  • ad hoc to paper MCP and reweighted l1 together yield better sparsity and detail preservation than standard TV-l0 in practice.
    Central modeling choice presented without derivation or proof in the abstract.

pith-pipeline@v0.9.0 · 5534 in / 1373 out tokens · 63651 ms · 2026-05-10T03:14:30.907551+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages

  1. [1]

    and Egiazarian, K

    Campisi, P . and Egiazarian, K. eds., 2017.Blind image deconvolution: theory and applications. CRC press

    work page 2017

  2. [2]

    and O’leary, D.P ., 2006.Deblurring images: matrices, spectra, and filtering

    Hansen, P .C., Nagy, J.G. and O’leary, D.P ., 2006.Deblurring images: matrices, spectra, and filtering. Society for Industrial and Applied Mathematics

    work page 2006

  3. [3]

    and Fatemi, E., 1992

    Rudin, L.I., Osher, S. and Fatemi, E., 1992. Nonlinear total variation based noise removal algorithms.Physica D: nonlinear phenomena, 60(1-4), pp.259-268

    work page 1992

  4. [4]

    and De Mol, C., 2004

    Daubechies, I., Defrise, M. and De Mol, C., 2004. An iterative thresholding algo- rithm for linear inverse problems with a sparsity constraint.Communications on Pure and Applied Mathematics: A Journal Issued by the Courant Institute of Mathematical Sciences, 57(11), pp.1413-1457

    work page 2004

  5. [5]

    Exact reconstruction of sparse signals via nonconvex min- imization.IEEE Signal Processing Letters, 14(10), pp.707-710

    Chartrand, R., 2007. Exact reconstruction of sparse signals via nonconvex min- imization.IEEE Signal Processing Letters, 14(10), pp.707-710

    work page 2007

  6. [6]

    and Jia, J., 2011, December

    Xu, L., Lu, C., Xu, Y. and Jia, J., 2011, December. Image smoothing via L 0 gra- dient minimization. InProceedings of the 2011 SIGGRAPH Asia conference(pp. 1-12)

    work page 2011

  7. [7]

    and Freeman, W .T ., 2006

    Fergus, R., Singh, B., Hertzmann, A., Roweis, S.T . and Freeman, W .T ., 2006. Re- moving camera shake from a single photograph. InAcm Siggraph 2006 Papers (pp. 787-794)

    work page 2006

  8. [8]

    and Lee, S., 2009

    Cho, S. and Lee, S., 2009. Fast motion deblurring. InACM SIGGRAPH Asia 2009 papers(pp. 1-8)

    work page 2009

  9. [9]

    and Chen, Y., 2020

    Dong, W ., Tao, S., Xu, G. and Chen, Y., 2020. Blind deconvolution for Poissonian blurred image with total variation and L 0-norm gradient regularizations.IEEE Transactions on Image Processing, 30, pp.1030-1043. 19

    work page 2020

  10. [11]

    and Gu, Y., 2014

    Chen, L. and Gu, Y., 2014. The convergence guarantees of a non-convex ap- proach for sparse recovery.IEEE Transactions on Signal Processing, 62(15), pp.3754-3767

    work page 2014

  11. [12]

    and Eldar, Y.C., 2023

    Zhang, X., Yu, L., Zheng, G. and Eldar, Y.C., 2023. Spiking sparse recovery with non-convex penalties.IEEE transactions on signal processing, 70, pp.6272-6285

    work page 2023

  12. [13]

    and Wahlberg, B., 2016

    Malek-Mohammadi, M., Rojas, C.R. and Wahlberg, B., 2016. A class of noncon- vex penalties preserving overall convexity in optimization-based mean filter- ing.IEEE Transactions on Signal Processing, 64(24), pp.6650-6664

    work page 2016

  13. [15]

    Nearly unbiased variable selection under minimax concave penalty.Annals of statistics, 38(2), pp.894-942

    Zhang, C.H., 2010. Nearly unbiased variable selection under minimax concave penalty.Annals of statistics, 38(2), pp.894-942

    work page 2010

  14. [16]

    Total variation denoising via the Moreau envelope.IEEE Sig- nal Processing Letters, 24(2), pp.216-220

    Selesnick, I., 2017. Total variation denoising via the Moreau envelope.IEEE Sig- nal Processing Letters, 24(2), pp.216-220

    work page 2017

  15. [17]

    and Ding, S., 2018

    Zou, J., Shen, M., Zhang, Y., Li, H., Liu, G. and Ding, S., 2018. Total variation denoising with non-convex regularizers.IEEE Access, 7, pp.4422-4431

    work page 2018

  16. [18]

    and Combettes, P .L., 2011.Convex analysis and moreau opera- tor theory in hilbert space

    Bauschke, H.H. and Combettes, P .L., 2011.Convex analysis and moreau opera- tor theory in hilbert space. Springer

    work page 2011

  17. [19]

    Society for Industrial and Applied Mathematics

    Beck, A., 2017.First-order methods in optimization. Society for Industrial and Applied Mathematics

    work page 2017

  18. [20]

    Sparse regularization via convex analysis.IEEE Transactions on Signal Processing, 65(17), pp.4481-4494

    Selesnick, I., 2017. Sparse regularization via convex analysis.IEEE Transactions on Signal Processing, 65(17), pp.4481-4494

    work page 2017

  19. [21]

    and Pesquet, J.C., 2011

    Combettes, P .L. and Pesquet, J.C., 2011. Proximal splitting methods in signal processing. InFixed-point algorithms for inverse problems in science and engi- neering(pp. 185-212). New York, NY: Springer New York

    work page 2011

  20. [22]

    and Zhang, Y., 2008

    Wang, Y., Yang, J., Yin, W . and Zhang, Y., 2008. A new alternating minimization algorithm for total variation image reconstruction.SIAM Journal on Imaging Sciences, 1(3), pp.248-272

    work page 2008

  21. [23]

    and Boyd, S.P ., 2008

    Candes, E.J., Wakin, M.B. and Boyd, S.P ., 2008. Enhancing sparsity by reweightedℓ 1 minimization.Journal of Fourier analysis and applications, 14(5), pp.877-905

    work page 2008

  22. [24]

    and Jia, J., 2008, October

    Liu, R. and Jia, J., 2008, October. Reducing boundary artifacts in image decon- volution. In2008 15th IEEE International Conference on Image Processing(pp. 505-508). IEEE. 20

    work page 2008

  23. [25]

    and Li, Q., 2010

    Wang, Z. and Li, Q., 2010. Information content weighting for perceptual im- age quality assessment.IEEE Transactions on image processing, 20(5), pp.1185- 1198

    work page 2010

  24. [26]

    and Bovik, A.C., 2003, November

    Wang, Z., Simoncelli, E.P . and Bovik, A.C., 2003, November. Multiscale struc- tural similarity for image quality assessment. InThe thrity-seventh asilomar conference on signals, systems & computers, 2003 (Vol. 2, pp. 1398-1402). Ieee

    work page 2003

  25. [27]

    and Zhang, D., 2011

    Zhang, L., Zhang, L., Mou, X. and Zhang, D., 2011. FSIM: A feature similarity in- dex for image quality assessment.IEEE transactions on Image Processing, 20(8), pp.2378-2386

    work page 2011

  26. [28]

    and Ziou, D., 2010, August

    Hore, A. and Ziou, D., 2010, August. Image quality metrics: PSNR vs. SSIM. In 201020th international conference on pattern recognition(pp. 2366-2369). IEEE

    work page 2010

  27. [29]

    and Freeman, W .T ., 2009, June

    Levin, A., Weiss, Y., Durand, F . and Freeman, W .T ., 2009, June. Understand- ing and evaluating blind deconvolution algorithms. In 2009IEEE conference on computer vision and pattern recognition(pp. 1964-1971). IEEE

    work page 2009

  28. [30]

    Pan, J., Hu, Z., Su, Z. and Yang, M.H., 2016.l 0-regularized intensity and gradi- ent prior for deblurring text images and beyond.IEEE transactions on pattern analysis and machine intelligence, 39(2), pp.342-355

    work page 2016

  29. [31]

    and Truong, T .K., 2020

    Wen, F ., Ying, R., Liu, Y., Liu, P . and Truong, T .K., 2020. A simple local mini- mal intensity prior and an improved algorithm for blind image deblurring.IEEE Transactions on Circuits and Systems for Video Technology, 31(8), pp.2923-2937

    work page 2020

  30. [32]

    Point spread function estimation for blind image deblurring problems based on framelet transform.The Visual Computer, 39(7), pp.2653- 2669

    Parvaz, R., 2023. Point spread function estimation for blind image deblurring problems based on framelet transform.The Visual Computer, 39(7), pp.2653- 2669

    work page 2023

  31. [33]

    and Harmeling, S., 2012, Octo- ber

    Köhler, R., Hirsch, M., Mohler, B., Schölkopf, B. and Harmeling, S., 2012, Octo- ber. Recording and playback of camera shake: Benchmarking blind deconvo- lution with a real-world database. InEuropean conference on computer vision (pp. 27-40). Berlin, Heidelberg: Springer Berlin Heidelberg

    work page 2012

  32. [34]

    and Freeman, W .T ., 2006

    Fergus, R., Singh, B., Hertzmann, A., Roweis, S.T . and Freeman, W .T ., 2006. Re- moving camera shake from a single photograph. InAcm Siggraph2006 Papers (pp. 787-794)

    work page 2006

  33. [35]

    and Jia, J., 2010, September

    Xu, L. and Jia, J., 2010, September. Two-phase kernel estimation for robust motion deblurring. InEuropean conference on computer vision(pp. 157-170). Berlin, Heidelberg: Springer Berlin Heidelberg. 21

    work page 2010