Real-Time Blind Defocus Deblurring for Earth Observation: The IMAGIN-e Mission Approach

Alejandro D. Mousist

REVIEW 3 major objections 4 minor 12 references

Real-Time Blind Defocus Deblurring for Earth Observation: The IMAGIN-e Mission Approach

T0 review · 3 major / 4 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 A GAN trained on synthetic blur restores real defocused images from a camera on the ISS, with no reference images and within edge-computing limits.

desk verdict A genuinely deployed edge-deblurring pipeline for an ISS camera, but the 'validated real-world restoration' claim rests on NIQE/BRISQUE and no kernel characterization; needs stronger evidence before the strong claim stands. read the letter →

arxiv 2505.22128 v2 pith:VPHPXYMY submitted 2025-05-28 cs.CV cs.AI

Alejandro D. Mousist This is my paper · ORCID

classification cs.CVcs.AI

keywords blinddeblurringdefocusEarthobservationedgecomputinggenerativeadversarialnetworkSentinel-2imagerestorationno-referencequalityassessment

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

Mechanical defocus is blurring the images captured by the IMAGIN-e camera on the International Space Station, and the mission has no sharp reference frame from that camera to train a correction. This paper claims the blur can nevertheless be removed in real time onboard: patches from the sharper Sentinel-2 satellite are degraded with synthetic blur families (Gaussian, defocus, shot, motion, spin), and a MIMO-Unet++ generator trained inside a GAN learns the blurred-to-sharp mapping. On synthetic validation where clean references exist, SSIM rises from 0.4442 to 0.7662 and PSNR from 24.01 dB to 30.02 dB. On the real IMAGIN-e captures, where no reference exists, the no-reference perceptual metrics NIQE and BRISQUE improve by 60.66% and 48.38%, and the model runs as a deployed post-processing step under tight edge-computing limits. The value of the claim, if right, is that a compromised instrument becomes usable for water-body segmentation and contour detection without hardware fixes.

What carries the argument

The load-bearing mechanism is the MIMO-Unet++, a coarse-to-fine single-image deblurring network, repurposed as the generator of a GAN. Training data are 256×256 patches downsampled from 1024×1024 Sentinel-2 scenes and degraded with synthetic Gaussian, defocus, shot, motion, and spin blur, with a batch size of four over 3000 iterations. The training loss combines the adversarial signal of a multi-scale discriminator (inspired by Pix2pixHD, with self-attention and spectral normalization), an L1 term, an FFT-domain loss, and a perceptual loss from a VGG16 network pre-trained on Sentinel-2 imagery. A second loop is equally load-bearing: an initial model sharpened real images enough to allow georeferencing against Sentinel-2, which sharpened the characterization of the real noise and fed back into more realistic synthetic training data for the final deployed model.

What would settle it

Measure the IMAGIN-e sensor's on-orbit point-spread function — for instance from a star-field exposure or a known ground target — and compare its radial frequency profile with the family of synthetic blur kernels used in training. If the real PSF lies outside that family, or if deconvolving the measured PSF from a raw capture does not reproduce the reported NIQE/BRISQUE gains, the real-world restoration improvements cannot be attributed to inverting the actual defocus.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that blind defocus deblurring is feasible for the IMAGIN-e payload without any sharp reference from the target sensor. The central claim is that a model trained exclusively on synthetically degraded Sentinel-2 RGB patches generalizes to the real mechanical defocus of the IMAGIN-e camera, improving perceptual quality on actual captures and enabling coarse downstream analyses such as water-body segmentation and contour detection. Edge functionality is part of the claim: the network runs in the capture pipeline on shared CPUs with no acceleration hardware, processing a 2048×1536 image in roughly five minutes with about 600 MB of peak memory. The paper also claims the restoration is structurally meaningful, not merely cosmetic, citing clearer Sobel edge delineation on processed images and the usefulness of the outputs for map-generation tasks, while acknowledging that resolution remains too low for small-object detection or fine-grained segmentation of closely related classes.

Load-bearing premise

The load-bearing premise is that the synthetic blurs applied to Sentinel-2 patches (Gaussian, defocus, shot, motion, spin) adequately mimic the real mechanical defocus of the IMAGIN-e camera, a match the paper assumes but never measures quantitatively.

Editorial extensions

If this is right

If the claim holds, the IMAGIN-e instrument — otherwise largely unusable because of its defocus — can keep operating as-is, with an onboard software step restoring images before downstream applications consume them.
The reported 5-minute processing time for 2048×1536 captures on shared CPUs without acceleration hardware fits the mission's post-processing pipeline, which the paper presents as an operational deployment of generative restoration in an ISS Earth-observation context.
Reference-free validation can rely on no-reference perceptual metrics: since NIQE and BRISQUE improved substantially on real captures, the paper argues such metrics can stand in for absent ground truth in deployed settings.
The restored images support concrete applications — water-body segmentation and coarse contour detection for map generation — so downstream onboard applications can be designed around the deblurred product.
The iterative noise-characterization loop (model, then georeferencing, then better synthetic data, then a better model) is presented as the path that turns an initially unknown sensor defect into a trainable restoration problem.

Reading between the lines

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

An editorial extension: the same synthetic-to-real bootstrapping should transfer to other uncharacterized cameras with mechanical or thermal misalignment, provided the real blur stays inside the synthetic kernel family used at training time — the paper does not demonstrate that condition for sensors beyond IMAGIN-e.
An editorial caution that doubles as an experiment: the natural decisive check is measuring the IMAGIN-e point-spread function on orbit (for example from a star-field exposure) and comparing its radial profile to the synthetic kernel set; the paper reports no such comparison.
A testable extension in the paper's spirit: use the NIQE/BRISQUE scores as an onboard quality gate, flagging or re-restoring captures whose residual perceptual score remains high — a closed loop the deployed system does not yet implement.
An editorial extrapolation about the architecture: because the model is trained and applied at 256×256 with upscaling, its gains concentrate on coarse structure; restoring at native resolution with the same GAN objective is the most direct route to the small-object and fine-segmentation applications the paper leaves open.

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

Short version: this is a real engineering deployment, but the abstract's 'validating real-world onboard restoration' is not supported by the evidence. The paper is worth a serious referee, not because the central claim is proven, but because the system exists, runs in 5 minutes on a 3-core/300MB payload, and the approach—Sentinel-2-derived synthetic training for blind defocus—is a sensible template.

What's actually new: the IMAGIN-e mission result is new. Adapting MIMO-Unet++ into a GAN with a multi-scale discriminator, self-attention, spectral normalization, FFT/perceptual losses, trained on Sentinel-2 patches with synthetic blur, then deployed onboard, is a legitimate engineering contribution. The authors disclose operational constraints honestly: no clean reference, no dedicated hardware, 300MB RAM, virtual memory, occasional ringing, and resolution insufficient for small-object detection. That kind of concrete detail is useful.

Where it's soft: the main validation gap is exactly where the reader's strongest claim lands. The 72.47% SSIM / 25% PSNR gains on Sentinel-2 only show the model inverts its own synthetic degradation pipeline; they don't validate generalization to IMAGIN-e. The IMAGIN-e numbers are NIQE and BRISQUE, which are not defocus-specific and can improve with contrast/brightness changes or mild sharpening. Without matching to a ground truth, a calibrated sensor blur model, or a downstream task metric, the 60.66%/48.38% improvement doesn't establish that real defocus was removed. The kernel estimation procedure is also left unspecified—'estimates the defocus kernel' appears, but no method or comparison is given. And the 'edge constraints' claim is softer than presented: peak memory is 600MB, double the 300MB RAM budget, so it relies on virtual memory, and 5 minutes per 2048x1536 image is 'real-time' only in an onboard post-processing sense. No code or data are provided, which limits reproducibility but is common for mission-specific papers.

If you work on EO deblurring or onboard image restoration, this is a useful existence proof and a cautionary example about reference-free validation. It deserves peer review: a good reviewer would ask for kernel characterization and at least one independent verification, such as registration to Sentinel-2 after deblurring or downstream segmentation accuracy. My own verdict would be conditional, not accept.

Referee Report

3 major / 4 minor

Summary. The paper presents a blind deblurring method for the IMAGIN-e Earth observation payload, where mechanically defocused images are restored onboard under severe edge-computing constraints (no acceleration hardware, 300 MB RAM, 3 shared CPU cores). The approach uses Sentinel-2 imagery to generate synthetic degraded training data (Gaussian, defocus, shot, motion, and spin blur), trains a MIMO-Unet++ based GAN with a multi-scale discriminator, and deploys the model in the capture pipeline. On synthetic Sentinel-2 data the method reports SSIM improvement of 72.47% and PSNR improvement of 25.00%. On real IMAGIN-e data, the paper reports NIQE improvement of 60.66% and BRISQUE improvement of 48.38%, and claims these results validate real-world onboard restoration. The paper also discusses edge implementation, processing time of approximately 5 minutes per image, peak memory of 600 MB with virtual memory usage, and qualitative results including Sobel edge detection.

Significance. If the real-world deblurring claim were established, this would be a valuable contribution: a blind deblurring system running onboard an ISS payload without dedicated acceleration hardware, enabling exploitation of a defocused instrument. The integration with an operational mission and the use of Sentinel-2 as a training reference are interesting and potentially useful design choices. However, the significance is currently constrained by the weakness of the real-world validation and the circularity of the synthetic evaluation. The paper would be strengthened by a more direct assessment of actual defocus removal on IMAGIN-e images, for example through kernel estimation comparison or reference-free metrics that are sensitive to specific blur artifacts, rather than generic perceptual quality indices.

major comments (3)

[Section 4, Table 2; Abstract] The real-world validation relies solely on NIQE and BRISQUE, which are generic no-reference quality models not tuned for defocus or for remote-sensing imagery. These metrics can improve with global contrast stretching, denoising, or mild sharpening that does not restore lost spatial frequencies, so they do not by themselves demonstrate that mechanical defocus was removed. The abstract's statement that these results are 'validating real-world onboard restoration' is therefore an overstatement. Please provide additional evidence: ideally a comparison with a sharp reference (e.g., a georeferenced Sentinel-2 scene acquired close in time), or at least a defocus-sensitive measure such as estimated blur kernel width, edge profile, or frequency-spectrum restoration on the real IMAGIN-e images.
[Section 3.1 and Section 4] The synthetic training and synthetic evaluation use the same degradation pipeline (Gaussian, defocus, shot, motion, and spin blur) applied to Sentinel-2 patches. Consequently, the reported SSIM/PSNR improvements on synthetic data partly demonstrate that the model inverts its own training distribution, and they do not establish that the model generalizes to the real IMAGIN-e defocus. The paper never quantitatively compares the synthetic blur kernel with the real sensor's blur; for example, estimated kernels from real IMAGIN-e images versus the synthetic kernels used in training. Without such a comparison, the synthetic validation cannot be used as evidence for real-world performance. Please add a quantitative analysis of the synthetic-to-real gap, or clearly limit the synthetic claims to self-consistency.
[Section 4, Table 1] The paper claims 'Real-Time Blind Defocus Deblurring' and states that the model 'maintains processing viability despite resource limitations,' but the reported numbers in Section 4 indicate a processing time of about 5 minutes for a 2048x1536 image and peak memory consumption of 600 MB, which exceeds the 300 MB available RAM and requires virtual memory. These numbers are not reconciled with the title's 'real-time' claim, nor with the mission's stated constraints. Please either provide a quantitative definition of 'real-time' in the context of IMAGIN-e's capture pipeline, or revise the real-time claim to something more modest such as 'onboard processing within mission time constraints.'

minor comments (4)

[Section 3.1] The paper states that the perceptual loss uses 'a VGG16 model pre-trained on Sentinel-2 images.' VGG16 is normally pretrained on ImageNet, and there is no standard 'Sentinel-2 pretrained' VGG16. Please clarify whether this means a network fine-tuned on Sentinel-2 data or something else, and provide a reference or training description.
[Section 4] The quantitative evaluation on IMAGIN-e appears to be based on an unspecified number of images. Please state how many real IMAGIN-e images were used for the NIQE/BRISQUE results, and whether the improvements are consistent across images or driven by a few favorable cases.
[Section 5] The conclusions mention enabling applications such as water body segmentation and contour detection, but the paper does not present any segmentation or detection results. Either add a brief quantitative evaluation of these downstream tasks or remove the claim.
[Throughout] There are several typographical and readability issues, including 'deblurring without reference images' used loosely (the method does use reference images in training), inconsistent spacing in equations or captions, and duplicated phrasing such as 'mounted on the Columbus module, which is externally mounted on the Columbus module' in Section 2.1. A careful copyedit is needed.

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity in the synthetic training/evaluation loop; the real-world NIQE/BRISQUE claim is weak but not formally circular.

fitted input called prediction [Section 3.1 (Model Architecture and Training Strategy) and Section 4, Table 2]
"we first trained an early version of the MIMO-Unet++ model using RGB images generated from Sentinel2 products and augmented with various noise types (Gaussian, defocus, shot, motion, and spin blur). The outputs of this model allowed us to correlate the images relative to their Sentinel-2 counterparts, leading to improved noise characterization and the creation of more realistic synthetic training data. Subsequently, we used these synthetic images to train a refined MIMO-Unet++ model within a GAN framework."

The synthetic degradation distribution used for training is refined using outputs of an early version of the same model, and the refined model is then trained and evaluated on synthetic images from that same pipeline. The reported +72.47% SSIM and +25.00% PSNR therefore measure how well the model inverts its own training distribution, not whether the synthetic blur actually matches the real IMAGIN-e defocus. The paper provides no quantitative comparison between the synthetic degradation and the real sensor's blur spectrum, so the synthetic 'validation' is self-referential rather than an independent confirmation of real defocus removal.

full rationale

The paper has two validation arms. The synthetic arm (Section 3.1, Section 4, Table 2) is partially circular: the early MIMO-Unet++ model's outputs are used to improve the noise characterization and create the synthetic training data, and the refined model is trained and evaluated on the same synthetic pipeline. Thus the SSIM/PSNR improvements largely demonstrate inversion of a self-generated degradation distribution, with no independent check against the real IMAGIN-e defocus. The real-arm evidence (NIQE -60.66%, BRISQUE -48.38%) is not circular in the strict sense, because the model was not fitted to those metrics, but NIQE and BRISQUE are generic no-reference opinion-quality models rather than defocus-specific or reference-based measures; the abstract's phrase 'validating real-world onboard restoration' therefore overstates what these metrics establish. No self-citation chain, uniqueness theorem, or defined-in-terms-of-each-other construction is present. The partial circularity is confined to the synthetic evaluation loop, while the central real-world claim retains some independent, though weak, metric content, so a mid-range score is appropriate.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the untested transfer from synthetic degradations to real defocus, on an undisclosed kernel estimation procedure, and on the validity of no-reference metrics as evidence of actual deblurring. These assumptions are not independently supported in the paper.

free parameters (3)

Defocus kernel (unspecified model and parameters) = Not disclosed
The kernel is claimed to be estimated from IMAGIN-e data via an iterative alignment process, but its functional form and fitting procedure are not described in Section 3.1.
Synthetic degradation parameters (Gaussian, defocus, shot, motion, spin) = Not disclosed
The exact magnitudes and kernels used to corrupt Sentinel-2 images are not specified; these control the distribution the model is trained on and thus are central to the reported SSIM/PSNR gains.
Training hyperparameters = lr=1e-4, batch=4, 3000 iterations, MultiStepLR
Chosen by hand, not central to the claim, but included for completeness.

assumptions (4)

domain assumption Sentinel-2 downscaled RGB images at 40m GSD serve as sharp references for the IMAGIN-e defocused images
Used throughout Section 3; requires that the scenes are geometrically comparable and that the only major difference is defocus.
ad hoc to paper Synthetic degradations (Gaussian, defocus, shot, motion, spin) adequately mimic the real IMAGIN-e defocus
Assumed in Section 3.1 where the model is trained on these degradations and then applied to real images without direct validation of distribution match.
domain assumption NIQE and BRISQUE are valid proxies for deblurring quality on real images
Used in Section 4 to conclude real-world restoration; these metrics are generic naturalness measures and do not specifically measure sharpness or defocus removal.
ad hoc to paper A VGG16 pretrained on Sentinel-2 images provides a useful perceptual loss
Section 3.1 mentions this without details on the pretraining dataset or whether it benefits the task.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Real-Time Blind Defocus Deblurring for Earth Observation: The IMAGIN-e Mission Approach." pith.science (2026). https://pith.science/paper/VPHPXYMY

@misc{pith2026250522128,
  author       = {Pith},
  title        = {Pith review of: Real-Time Blind Defocus Deblurring for Earth Observation: The IMAGIN-e Mission Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VPHPXYMY}},
  note         = {Machine review of arXiv:2505.22128}
}

read the original abstract

This work addresses mechanical defocus in Earth observation images from the IMAGIN-e mission aboard the ISS, proposing a blind deblurring approach adapted to space-based edge computing constraints. Leveraging Sentinel-2 data, our method estimates the defocus kernel and trains a restoration model within a GAN framework, effectively operating without reference images. On Sentinel-2 images with synthetic degradation, SSIM improved by 72.47% and PSNR by 25.00%, confirming the model's ability to recover lost details when the original clean image is known. On IMAGIN-e, where no reference images exist, perceptual quality metrics indicate a substantial enhancement, with NIQE improving by 60.66% and BRISQUE by 48.38%, validating real-world onboard restoration. The approach is currently deployed aboard the IMAGIN-e mission, demonstrating its practical application in an operational space environment. By efficiently handling high-resolution images under edge computing constraints, the method enables applications such as water body segmentation and contour detection while maintaining processing viability despite resource limitations.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages

[1]

Machine learning models for eos sat-1 satellite image enhancing

Viacheslav Popika and Lidia Lelechenko. Machine learning models for eos sat-1 satellite image enhancing. In IGARSS 2024-2024 IEEE International Geoscience and Remote Sens- ing Symposium, pages 1095–1098. IEEE, 2024

work page 2024
[2]

Extrapolation, interpolation, and smoothing of stationary time series: With engineering applications

Norbert Wiener. Extrapolation, interpolation, and smoothing of stationary time series: With engineering applications. 1949

work page 1949
[3]

Bayesian-based iterative method of image restoration

William Hadley Richardson. Bayesian-based iterative method of image restoration. Journal of the Optical Society of America, 62(1):55–59, 1972

work page 1972
[4]

Deep multi-scale convolutional neural network for dynamic scene deblurring

Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on com- puter vision and pattern recognition , pages 3883–3891, 2017

work page 2017
[5]

Scale-recurrent network for deep image deblurring

Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8174–8182, 2018

work page 2018
[6]

Deblurgan: Blind motion deblurring using conditional adversarial networks

Orest Kupyn, V olodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Ji ˇr´ı Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8183–8192, 2018

work page 2018
[7]

Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better

Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. InProceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV), October 2019

work page 2019
[8]

Deblur- dinat: A lightweight and effective transformer for image de- blurring

Hanzhou Liu, Binghan Li, Chengkai Liu, and Mi Lu. Deblur- dinat: A lightweight and effective transformer for image de- blurring. arXiv e-prints, pages arXiv–2403, 2024

work page 2024

Show all 12 references

[9]

Mimo-uformer: A transformer-based image deblurring network for vehicle surveillance scenarios

Jian Zhang, Baoping Cheng, Tengying Zhang, Yongsheng Zhao, Tao Fu, Zijian Wu, and Xiaoming Tao. Mimo-uformer: A transformer-based image deblurring network for vehicle surveillance scenarios. Journal of Imaging, 10(11):274, 2024

2024
[10]

Rethinking coarse-to-fine approach in sin- gle image deblurring, 2021

Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in sin- gle image deblurring, 2021. URL https://arxiv.org/ abs/2108.05054

2021 arXiv
[11]

High-resolution image syn- thesis and semantic manipulation with conditional gans

Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image syn- thesis and semantic manipulation with conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8798–8807, 2018

2018
[12]

Self-attention generative adversarial networks

Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augus- tus Odena. Self-attention generative adversarial networks. In International Conference on Machine Learning , pages 7354– 7363, 2019

2019

Pith tools

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

[1] [1]

Machine learning models for eos sat-1 satellite image enhancing

Viacheslav Popika and Lidia Lelechenko. Machine learning models for eos sat-1 satellite image enhancing. In IGARSS 2024-2024 IEEE International Geoscience and Remote Sens- ing Symposium, pages 1095–1098. IEEE, 2024

work page 2024

[2] [2]

Extrapolation, interpolation, and smoothing of stationary time series: With engineering applications

Norbert Wiener. Extrapolation, interpolation, and smoothing of stationary time series: With engineering applications. 1949

work page 1949

[3] [3]

Bayesian-based iterative method of image restoration

William Hadley Richardson. Bayesian-based iterative method of image restoration. Journal of the Optical Society of America, 62(1):55–59, 1972

work page 1972

[4] [4]

Deep multi-scale convolutional neural network for dynamic scene deblurring

Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on com- puter vision and pattern recognition , pages 3883–3891, 2017

work page 2017

[5] [5]

Scale-recurrent network for deep image deblurring

Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8174–8182, 2018

work page 2018

[6] [6]

Deblurgan: Blind motion deblurring using conditional adversarial networks

Orest Kupyn, V olodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Ji ˇr´ı Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8183–8192, 2018

work page 2018

[7] [7]

Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better

Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. InProceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV), October 2019

work page 2019

[8] [8]

Deblur- dinat: A lightweight and effective transformer for image de- blurring

Hanzhou Liu, Binghan Li, Chengkai Liu, and Mi Lu. Deblur- dinat: A lightweight and effective transformer for image de- blurring. arXiv e-prints, pages arXiv–2403, 2024

work page 2024