Pith. sign in

REVIEW 4 major objections 6 minor 65 references

SpikeRestormer is the first spiking-neural-network framework for all-in-one image restoration, matching ANN-based quality at a fraction of the estimated energy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 09:53 UTC pith:GDULAQ6H

load-bearing objection A plausible first directly-trained SNN for all-in-one restoration with competitive PSNR at T=1, but the headline energy claim hinges on an under-described SOP estimator that shows unexplained non-spiking costs. the 4 major comments →

arxiv 2608.02290 v1 pith:GDULAQ6H submitted 2026-08-03 cs.CV

SpikeRestormer: Towards Energy-Efficient All-in-One Image Restoration via Unified Event Reasoning

Shengkai Hu , Jie Shao , Jiaqi Ma , Xu Zhang , Keying Wu , Qilu Zhu , Beihang Song , Jun Wan This is my paper
classification cs.CV
keywords spiking neural networksall-in-one image restorationevent reasoningsubtractive attentionadditive attentionBayesian skip maskingenergy efficiencywavelet transform
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

SpikeRestormer aims to show that spiking neural networks — the low-power, event-driven class of networks — can do all-in-one image restoration, a task usually handled by heavy ANN models. The paper's claim is that a single spiking network operating at one time step, with no ANN-to-SNN conversion, can restore images degraded by haze, rain, noise, blur, and low light as well as ANN baselines, while consuming far less estimated energy. Restoration is reformulated as a chain of spike-based events: subtractive attention in the encoder turns degradation-induced differences into spike events; a hierarchical Bayesian module filters unreliable events; additive attention in the decoder builds restoration events. The paper reports state-of-the-art results among SNN methods and competitive PSNR/SSIM against ANN methods at energy costs of roughly 4–17 mJ depending on model size.

Core claim

On its own terms, SpikeRestormer is the first SNN designed for all-in-one image restoration. It asserts that, without relying on ANN-to-SNN conversion, a single spiking network using one temporal step can unify five restoration tasks — dehazing, deraining, denoising, deblurring, and low-light enhancement — and perform competitively with ANN methods on standard benchmarks while establishing new state-of-the-art results among SNN-based methods. The claimed efficiency is large: on the AiOIR-3 benchmark the full model reports 16.80 mJ estimated energy versus 83.64 mJ for a prior SNN and 485.28 mJ for a representative ANN, with better average PSNR. The mechanism is a three-stage event-reasoning p

What carries the argument

The load-bearing mechanism is a pair of attention reformulations for spike-based computation. Standard spike attention multiplies binary query–key responses, which keeps only co-activated spikes; the paper replaces this with subtractive interaction in the encoder (Subtractive Degradation Event Attention, SDEA) to expose degradation discrepancies, and additive interaction in the decoder (Additive Restoration Event Attention, AREA) to preserve unilateral spikes. Two theorems show for binary Bernoulli spike responses that subtraction has higher activation probability and conditional information gain than multiplication when firing is sparse, and similarly for addition. These operations work on

Load-bearing premise

The paper's central efficiency claim rests on an estimated energy model: it assumes that the listed per-operation spike costs fully capture the network's energy use, including the wavelet transform and the Bayesian gate's non-spiking computations, and that real neuromorphic hardware adds no large fixed overhead.

What would settle it

Measure SpikeRestormer's actual energy on neuromorphic hardware (or a detailed cycle-accurate simulator) and compare it against an ANN baseline of similar quality on the same data; if the SNN does not consume substantially less energy, the paper's headline efficiency claim fails even if the PSNR results stand.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Spiking networks, not just ANNs, can perform all-in-one restoration; the paper reports T=1 operation with no ANN-to-SNN conversion.
  • Estimated energy drops by roughly 5x versus prior SNN methods and by over 20x versus a representative ANN baseline on the same benchmark, while average PSNR improves.
  • The method extends to composited degradations (CDD11), remote sensing (MDRS-Landsat), and real-world weather (WeatherBench), suggesting the event-reasoning formulation transfers across domains.
  • Ablations attribute most of the gain to the subtractive/additive attention pairing; replacing multiplicative spike attention with these operations raises PSNR by 2.34 dB while cutting estimated energy by 67.8%.
  • The temporal-step study shows that a single step already saturates performance; adding steps triples energy for marginal PSNR gains, making the single-step design the practical operating point.

Where Pith is reading between the lines

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

  • The paper's energy numbers are estimates from spike-operation counts, not hardware measurements; a neuromorphic implementation could confirm or overturn the efficiency claim, especially because the wavelet transform and the Bayesian gate's Gaussian/softmax steps may add operations the estimator undercounts.
  • The subtractive/additive attention principle is more general than restoration: any SNN task where input contains both shared structure and task-relevant discrepancies (e.g., anomaly detection, change detection) could adopt the same encoder/decoder asymmetry.
  • The Bayesian skip-masking idea suggests a recipe for uncertainty-aware skip connections in other deep spiking or even ANN architectures: infer a latent reliability state per stage and use its posterior to gate information flow.
  • If the energy estimates survive measurement, the natural next test is scaling to higher resolutions and video, where static-image degradation cues become temporal event streams and the spiking advantage could grow.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. This paper proposes SpikeRestormer, an SNN for all-in-one image restoration (AiOIR) that operates at T=1 without ANN-to-SNN conversion. The architecture couples subtractive degradation-event attention (SDEA), hierarchical Bayesian skip masking (HBSM), and additive restoration-event attention (AREA) in a wavelet-based encoder–decoder. The authors report SOTA results among SNN baselines on AiOIR-3/5, CDD11, MDRS-Landsat, and WeatherBench, and claim competitive performance against the selected ANN methods with substantially lower energy consumption, using an SOP-based estimator from Zhou et al. (2024). Two theorems are given to motivate subtractive and additive spike interactions under a Bernoulli model.

Significance. If the empirical and efficiency claims hold, SpikeRestormer is a significant contribution: it is, to the authors' knowledge, the first direct-trained SNN for AiOIR, achieves strong PSNR/SSIM at T=1, and reports large energy reductions over ANN baselines. The paper covers multiple benchmarks, includes ablations of the three proposed components, and ships a theoretical justification, albeit a simplified one. The main weakness is that the headline energy-efficiency claim depends on an opaque estimation protocol, the ANN baselines are not current SOTA, and training/implementation details are missing. These issues are fixable but currently prevent full verification.

major comments (4)
  1. [Experimental Settings; Tables 7-8] The energy consumption numbers are not reproducible. The paper states energy is 'estimated following (Zhou et al. 2024)' but gives no formula or list of included operations. The ratio of reported energy to SOPs varies widely (Table 8a: 52.18 mJ / 0.479 G ≈ 109 mJ/G; Table 8d: 16.80 mJ / 0.314 G ≈ 54 mJ/G; Table 7 T=2: 30.06 mJ / 0.698 G ≈ 43 mJ/G), indicating unstated fixed costs or non-spiking operations. Since 'significantly lower energy consumption' is a central claim, the full estimation protocol must be provided, including whether the Haar transform, HBSM Gaussian/softmax, sigmoid, and elementwise residual operations are counted.
  2. [Experimental Settings; all result tables] The training configuration is missing. No loss function, optimizer, learning rate, schedule, batch size, epochs, resolution, data augmentation, or number of runs is reported. HBSM hyperparameters (K_s) and component initialization are not specified. Without these details, the reported improvements over ESDNet and VLIF are not reproducible and may be attributed to training choices rather than the proposed architecture. Code is also not provided. This is a major reproducibility gap for a paper claiming state-of-the-art results.
  3. [Tables 1-3] The ANN baseline selection is dated. The compared ANN methods (AirNet, Restormer, NAFNet, SwinIR, MPRNet, DGUNet, IDR, etc.) are mostly from 2021-2023, while the references include recent 2025-2026 ANN AiOIR methods such as ProRes, EvoIR, Perceive-IR, ClearAIR, and M2Restore. Without comparisons to these current ANN methods, the claim of being 'competitive against ANN-based methods' is not established against the current state of the art. The claim should be scoped to the tested baselines or the recent ANN methods should be added.
  4. [Rethinking Attention for Efficient AiOIR; Theorems 1-2] The theoretical justification is not connected to the actual network. The theorems assume independent Bernoulli(p) variables with identical p and require 0 < v_th <= 1 as a normalized threshold. In the real network, Q and K are derived from the same input via learned convolutions and tdBN, so they are not independent; the firing threshold is v_th=0.15, not normalized to the dynamic range of inputs; and the neuron's reset behavior is not modeled. Thus the theorems are a toy model that does not establish the claimed 'theoretical basis' for SDEA and AREA. The paper should either extend the analysis to correlated, non-identical inputs or explicitly present the theorems as intuition rather than a formal basis.
minor comments (6)
  1. [Table 1 caption] The caption says best and second-best results among SNN methods are highlighted in bold and underlined, but the table shows multiple bolds/underlines; please clarify the marking scheme.
  2. [Table 7] Define 'SOPs' (spike operations?) at first use and state how it is computed.
  3. [Eq. (6)] The stage-transition matrix T_s maps a posterior at stage s+1 to a prior at stage s; please specify the row/column convention and whether the posterior vector is a row or column vector.
  4. [Table 1] The variants SpikeRestormer-T and SpikeRestormer-S are used but not introduced in the architecture section; provide parameter counts or describe how they differ from the full model.
  5. [Abstract and Intro] The phrase 'significantly lower energy consumption' is not quantified; consider moving the quantitative comparison to the results section to avoid overstatement.
  6. [Eq. (2)] tdBN is used in the wavelet reference construction before being defined; define it on first use.

Circularity Check

0 steps flagged

No circularity: empirical restoration gains and Theorems 1–2 are self-contained; energy figures are an external estimator concern, not a derivation from the conclusion.

full rationale

I walked the paper's claimed derivation chain. Theorems 1 and 2 are explicit probability identities for Bernoulli spike interactions; they define S_sub and S_add and compute activation probabilities, entropies, and conditional information gains. They contain no fitted parameters, do not assume the experimental results, and are not used to generate the PSNR/SSIM numbers, so they are not circular. SDEA, HBSM, and AREA are architectural modules; the paper's claims about them are supported by ablations and external benchmark comparisons (Tables 1–8), not by self-referential derivation. The restoration results are measured against standard datasets and external ANN/SNN baselines, making them externally falsifiable. The energy-efficiency headline relies on the SOP-based estimator from Zhou et al. 2024 ('The energy consumption is estimated following (Zhou et al. 2024)' in Experimental Settings), and the paper does not disclose the exact formula or whether Haar, GAP, Gaussian/softmax, sigmoid, and elementwise operations are counted; that is a genuine validity/accounting limitation, but it is not an instance of a prediction reducing by construction to an input. The manuscript cites several works by the same authors (ClusIR, EvoIR, Proto-former) in related-work context; these citations are not load-bearing for the architecture, the theorems, or the reported benchmark numbers, so under the rule that self-citation is not itself circularity they do not increase the circularity score.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central claim depends on several hand-chosen hyperparameters (v_th, T, K_s) and on domain assumptions about the validity of the toy-model theorems and the SOP-based energy estimator. The Haar wavelet cost assumption is unstated but load-bearing for the efficiency claim. No external entities are introduced.

free parameters (3)
  • Firing threshold v_th = 0.15
    Set experimentally for all spiking neurons; not reported as swept or optimized, but it directly controls the activation probabilities in SDEA/AREA and LIF responses.
  • Number of latent reliability states K_s = not reported
    HBSM component count per stage; a design hyperparameter that determines the capacity of the learned Gaussian mixture. The paper never states its value.
  • Temporal step T = 1
    Chosen for efficiency; increasing to T=2,3 gives marginal PSNR gains (31.68, 31.76 vs 31.60). This is a deliberate trade-off setting rather than a fitted parameter, but it defines the operating point of the energy and accuracy claims.
axioms (4)
  • domain assumption Independent Bernoulli spike responses in Theorems 1 and 2 (q,k,f,g ~ Bernoulli(p), mutually independent)
    The theoretical justification of SDEA/AREA assumes independent identical spike distributions with p<1/2. Real spike responses are correlated, have spatial/channel structure, and are learned; therefore the theorems are toy-model motivations, not exact characterizations of the network.
  • standard math LIF fires whenever input >= threshold for inputs 0,1,2 with normalized threshold <=1
    Used to derive P(S_sub=1)=2p(1-p) and P(S_add=1)=2p-p^2. In the network, threshold is v_th=0.15 and inputs are integers 0/1/2 (or weighted sums), so firing for 0.15 is consistent, but membrane potential reset and weighted inputs are not modeled in the theorems.
  • domain assumption Energy estimation formula from Zhou et al. 2024 is valid for T=1 operations
    The energy comparisons rest on SOP-counting and per-op constants from prior work; no hardware measurements are provided. This premise is load-bearing for the 'significantly lower energy consumption' claim.
  • domain assumption Haar wavelet transform and inverse transform incur negligible or excluded energy cost
    The energy numbers do not appear to include the wavelet decomposition/reconstruction operations, which are non-spiking fixed transforms; if these have non-negligible cost on target hardware, the efficiency advantage changes.
invented entities (1)
  • Latent reliability state z_s in HBSM no independent evidence
    purpose: Discrete latent variable indexing K_s learned Gaussian components; used to compute a channel-wise reliability gate for skip connections.
    Internal latent variable with no falsifiable prediction outside the model; its posterior is a softmax over learned mixture components, so it acts as a gating mechanism rather than a physical entity.

pith-pipeline@v1.3.0-daily-deepseek · 15408 in / 17054 out tokens · 135647 ms · 2026-08-04T09:53:07.257839+00:00 · methodology

0 comments
read the original abstract

ANN-based All-in-One image restoration (AiOIR) unifies diverse degradation handling but incurs high computational costs, limiting its real-time deployment. While Spiking Neural Networks (SNNs) offer a low-power alternative, applying them to static images remains challenging. This difficulty arises because explicit event signals are absent, and degradation cues are heavily entangled with scene structures, hindering the learning of reliable restoration-oriented spike events. To address these issues, we propose SpikeRestormer, an energy-efficient SNN for AiOIR that performs event reasoning over internally generated spike cues. Specifically, we propose a degradation-event perception process to extract spike-based degradation events through Subtractive Degradation Event Attention (SDEA). Moreover, we introduce Hierarchical Bayesian Skip Masking (HBSM) and Additive Restoration Event Attention (AREA) processes for event-reliability inference and restoration-event construction, respectively. By integrating these complementary processes, SpikeRestormer formulates restoration as a unified process of degradation-event perception, degradation-event reliability inference, and restoration-event construction, liberating the potential of SNNs for energy-efficient AiOIR. Extensive experiments show that SpikeRestormer delivers competitive performance against ANN-based methods and establishes new state-of-the-art results among SNN-based methods with significantly lower energy consumption.

Figures

Figures reproduced from arXiv: 2608.02290 by Beihang Song, Jiaqi Ma, Jie Shao, Jun Wan, Keying Wu, Qilu Zhu, Shengkai Hu, Xu Zhang.

Figure 1
Figure 1. Figure 1: Motivation and evaluation of SpikeRestormer. (a) Different degradations produce distinct post-LIF responses under [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of SpikeRestormer. The encoder employs SWRBs with SDEA to perceive degradation events, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of events in SpikeRestormer. (a) [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

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

65 extracted references · 8 linked inside Pith

  1. [1]

    2025 , publisher=

    Beyond degradation redundancy: Contrastive prompt learning for all-in-one image restoration , author=. 2025 , publisher=

    2025

  2. [2]

    2025 , publisher=

    UniUIR: Considering underwater image restoration as an all-in-one learner , author=. 2025 , publisher=

    2025

  3. [3]

    2026 , publisher=

    Boosting Filter Optimization and Prompt-Guided Decoding for Mixed Degradation Image Restoration , author=. 2026 , publisher=

    2026

  4. [4]

    Acl: Activating capability of linear attention for image restoration , author=

  5. [5]

    Pattern Recognition , pages=

    Spiking pyramid wavelet transformation for high-efficient and low-energy image restoration , author=. Pattern Recognition , pages=. 2026 , publisher=

    2026

  6. [6]

    arXiv preprint arXiv:2405.06277 , year=

    Learning a spiking neural network for efficient image deraining , author=. arXiv preprint arXiv:2405.06277 , year=

    Pith/arXiv arXiv

  7. [7]

    Spiking meets attention: Efficient remote sensing image super-resolution with attention spiking neural networks , author=

  8. [8]

    Exploring the potentials of spiking neural networks for image deraining , author=

  9. [9]

    arXiv preprint arXiv:2512.10948 , year=

    ClusIR: Towards Cluster-Guided All-in-One Image Restoration , author=. arXiv preprint arXiv:2512.10948 , year=

    Pith/arXiv arXiv

  10. [10]

    arXiv preprint arXiv:2512.05104 , year=

    EvoIR: Towards All-in-One Image Restoration via Evolutionary Frequency Modulation , author=. arXiv preprint arXiv:2512.05104 , year=

    arXiv

  11. [11]

    IEEE Transactions on Image Processing , year=

    Perceive-ir: Learning to perceive degradation better for all-in-one image restoration , author=. IEEE Transactions on Image Processing , year=

  12. [12]

    arXiv preprint arXiv:2601.02763 , year=

    ClearAIR: A Human-Visual-Perception-Inspired All-in-One Image Restoration , author=. arXiv preprint arXiv:2601.02763 , year=

    arXiv

  13. [13]

    All-in-one image restoration for unknown corruption , author=

  14. [14]

    PromptIR: Prompting for All-in-One Blind Image Restoration , author=

  15. [15]

    2023 , publisher=

    Restoring vision in adverse weather conditions with patch-based denoising diffusion models , author=. 2023 , publisher=

    2023

  16. [16]

    arXiv preprint arXiv:2303.13739 , year=

    WM-MoE: Weather-aware multi-scale mixture-of-experts for blind adverse weather removal , author=. arXiv preprint arXiv:2303.13739 , year=

    Pith/arXiv arXiv

  17. [17]

    Efficient deweahter mixture-of-experts with uncertainty-aware feature-wise linear modulation , author=

  18. [18]

    Retinexformer: One-stage retinex-based transformer for low-light image enhancement , author=

  19. [19]

    2022 , organization=

    Wave-vit: Unifying wavelet and transformers for visual representation learning , author=. 2022 , organization=

    2022

  20. [20]

    2024 , organization=

    Restoring images in adverse weather conditions via histogram transformer , author=. 2024 , organization=

    2024

  21. [21]

    Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions , author =

  22. [22]

    arXiv preprint arXiv:2506.07814 , year=

    M2Restore: Mixture-of-Experts-based Mamba-CNN Fusion Framework for All-in-One Image Restoration , author=. arXiv preprint arXiv:2506.07814 , year=

    Pith/arXiv arXiv

  23. [23]

    Neural networks , volume=

    Networks of spiking neurons: the third generation of neural network models , author=. Neural networks , volume=. 1997 , publisher=

    1997

  24. [24]

    Nature , volume=

    Towards spike-based machine intelligence with neuromorphic computing , author=. Nature , volume=. 2019 , publisher=

    2019

  25. [25]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Rmp-snn: Residual membrane potential neuron for enabling deeper high-accuracy and low-latency spiking neural network , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  26. [26]

    IEEE transactions on neural networks and learning systems , volume=

    Advancing spiking neural networks toward deep residual learning , author=. IEEE transactions on neural networks and learning systems , volume=. 2024 , publisher=

    2024

  27. [27]

    arXiv preprint arXiv:2306.13653 , year=

    Prores: Exploring degradation-aware visual prompt for universal image restoration , author=. arXiv preprint arXiv:2306.13653 , year=

    Pith/arXiv arXiv

  28. [28]

    Neural Networks , volume=

    Restoration and enhancement on low exposure raw images by joint demosaicing and denoising , author=. Neural Networks , volume=. 2023 , publisher=

    2023

  29. [29]

    2024 , organization=

    Onerestore: A universal restoration framework for composite degradation , author=. 2024 , organization=

    2024

  30. [30]

    arXiv preprint arXiv:2209.15425 , year=

    Spikformer: When spiking neural network meets transformer , author=. arXiv preprint arXiv:2209.15425 , year=

    Pith/arXiv arXiv

  31. [31]

    Spike-driven transformer , author=

  32. [32]

    arXiv preprint arXiv:2006.13846 , year=

    Understanding ssim , author=. arXiv preprint arXiv:2006.13846 , year=

    Pith/arXiv arXiv 2006

  33. [33]

    Going deeper with directly-trained larger spiking neural networks , author=

  34. [34]

    2024 , organization=

    Spiking wavelet transformer , author=. 2024 , organization=

    2024

  35. [35]

    Qkformer: Hierarchical spiking transformer using qk attention , author=

  36. [36]

    Gradient as conditions: Rethinking HOG for all-in-one image restoration , author=

  37. [37]

    Ingredient-oriented multi-degradation learning for image restoration , author=

  38. [38]

    2022 , organization=

    Tape: Task-agnostic prior embedding for image restoration , author=. 2022 , organization=

    2022

  39. [39]

    Transweather: Transformer-based restoration of images degraded by adverse weather conditions , author=

  40. [40]

    2024 , organization=

    Mambair: A simple baseline for image restoration with state-space model , author=. 2024 , organization=

    2024

  41. [41]

    Restormer: Efficient transformer for high-resolution image restoration , author=

  42. [42]

    Swinir: Image restoration using swin transformer , author=

  43. [43]

    Deep generalized unfolding networks for image restoration , author=

  44. [44]

    2022 , organization=

    Simple baselines for image restoration , author=. 2022 , organization=

    2022

  45. [45]

    arXiv preprint arXiv:2310.01018 , volume=

    Controlling vision-language models for universal image restoration , author=. arXiv preprint arXiv:2310.01018 , volume=

    Pith/arXiv arXiv

  46. [46]

    Neural Networks , volume=

    Image denoising using deep CNN with batch renormalization , author=. Neural Networks , volume=. 2020 , publisher=

    2020

  47. [47]

    Dynamic scene deblurring with parameter selective sharing and nested skip connections , author=

  48. [48]

    FD-GAN: Generative adversarial networks with fusion-discriminator for single image dehazing , author=

  49. [49]

    2019 , publisher=

    A general decoupled learning framework for parameterized image operators , author=. 2019 , publisher=

    2019

  50. [50]

    Multi-stage progressive image restoration , author=

  51. [51]

    2024 , publisher=

    Neural degradation representation learning for all-in-one image restoration , author=. 2024 , publisher=

    2024

  52. [52]

    1999 , publisher=

    Elements of information theory , author=. 1999 , publisher=

    1999

  53. [53]

    Weatherbench: A real-world benchmark dataset for all-in-one adverse weather image restoration , author=

  54. [54]

    Information Fusion , volume=

    Ada4DIR: An adaptive model-driven all-in-one image restoration network for remote sensing images , author=. Information Fusion , volume=. 2025 , publisher=

    2025

  55. [55]

    Contour detection and hierarchical image segmentation , author=

  56. [56]

    Waterloo exploration database: New challenges for image quality assessment models , author=

  57. [57]

    A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics , author=

  58. [58]

    Single image super-resolution from transformed self-exemplars , author=

  59. [59]

    Deep joint rain detection and removal from a single image , author=

  60. [60]

    Benchmarking single-image dehazing and beyond , author=

  61. [61]

    Deep multi-scale convolutional neural network for dynamic scene deblurring , author=

  62. [62]

    Deep retinex decomposition for low-light enhancement , author=

  63. [63]

    FFDNet: Toward a fast and flexible solution for CNN-based image denoising , author=

  64. [64]

    Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising , author=

  65. [65]

    2026 , publisher=

    Proto-former: Unified facial landmark detection by prototype transformer , author=. 2026 , publisher=

    2026