pith. sign in

arxiv: 2604.12807 · v2 · submitted 2026-04-14 · 💻 cs.CV · cs.AI

Rethinking Satellite Image Restoration for Onboard AI: A Lightweight Learning-Based Approach

Adrien Dorise , Marjorie Bellizzi , Omar Hlimi This is my paper

Pith reviewed 2026-05-10 16:10 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords satellite image restorationonboard processingconvolutional neural networkFPGA deploymentPleiades-HR imageryPSNR improvementobject detectionlightweight model
0
0 comments X

The pith

A lightweight residual CNN trained on simulated data restores satellite images competitively while cutting onboard latency by 41 times.

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

The paper investigates whether a simple non-generative convolutional network can replace slow traditional restoration pipelines that correct noise and blur in satellite imagery. It shows that this lightweight model achieves higher image quality than the conventional ground-based approach and improves results on a following object-detection task. The network runs efficiently enough to be placed directly on satellite hardware, opening the door to real-time onboard AI without waiting for ground processing.

Core claim

ConvBEERS, a residual convolutional network trained only on simulated degradations, matches or exceeds a traditional sequential physical-model pipeline on both simulated test data and real Pleiades-HR imagery, delivering a 6.9 dB PSNR gain, up to 5.1 percent higher mAP@50 on downstream detection, and successful FPGA deployment with 41 times lower latency than the classical pipeline.

What carries the argument

ConvBEERS, a lightweight non-generative residual convolutional network that learns a direct mapping from degraded to restored satellite images.

If this is right

  • Restoration becomes fast enough to run before onboard AI tasks such as object detection.
  • Image quality gains of 6.9 dB PSNR translate into measurable improvements in detection accuracy of up to 5.1 percent mAP@50.
  • The same lightweight architecture fits on space-grade FPGA hardware and runs 41 times faster than the traditional pipeline.
  • Satellite systems can reduce reliance on ground-based processing for initial image correction.

Where Pith is reading between the lines

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

  • If simulation fidelity holds across other sensors, the approach could be reused for different satellite missions without per-mission retraining.
  • Lower latency and power draw on the FPGA may allow more complex AI models to run in parallel within the same onboard budget.
  • Replacing sequential physical models with a single learned residual network simplifies the overall processing chain and reduces the number of tunable parameters.

Load-bearing premise

Degradations created in simulation for training are close enough to the actual noise and blur present in real Pleiades-HR satellite captures that the learned mapping transfers without extra adaptation.

What would settle it

Retraining on a fresh set of real paired degraded-clean Pleiades-HR images and measuring whether PSNR or downstream mAP drops below the reported simulated results would falsify the claim that simulation alone is sufficient.

Figures

Figures reproduced from arXiv: 2604.12807 by Adrien Dorise, Marjorie Bellizzi, Omar Hlimi.

Figure 1
Figure 1. Figure 1: ConvBEERS architecture. It is composed of 3x3 size [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Image restoration on a simulated degraded image with [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of raw Pleiades image and restored results ´ using the traditional algorithm and ConvBEERS [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Object detection evaluation pipeline. The simulated degraded and restored datasets are created from the DIOR dataset using both our sensor simulation framework and ConvBEERS. Then, three YOLOv11n models are trained independently on (1) reference images, (2) degraded images, and (3) restored images using the training and validation sets. Their performance is compared on the test sets to evaluate the impact … view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of restored RGB images with [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
read the original abstract

Satellite image restoration aims to improve image quality by compensating for degradations (e.g., noise and blur) introduced by the imaging system and acquisition conditions. As a fundamental preprocessing step, restoration directly impacts both ground-based product generation and emerging onboard AI applications. Traditional restoration pipelines based on sequential physical models are computationally intensive and slow, making them unsuitable for onboard environments. In this paper, we introduce ConvBEERS: a Convolutional Board-ready Embedded and Efficient Restoration model for Space to investigate whether a light and non-generative residual convolutional network, trained on simulated satellite data, can match or surpass a traditional ground-processing restoration pipeline across multiple operating conditions. Experiments conducted on simulated datasets and real Pleiades-HR imagery demonstrate that the proposed approach achieves competitive image quality, with a +6.9dB PSNR improvement. Evaluation on a downstream object detection task demonstrates that restoration significantly improves performance, with up to +5.1% mAP@50. In addition, successful deployment on a Xilinx Versal VCK190 FPGA validates its practical feasibility for satellite onboard processing, with a ~41x reduction in latency compared to the traditional pipeline. These results demonstrate the relevance of using lightweight CNNs to achieve competitive restoration quality while addressing real-world constraints in spaceborne systems.

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 introduces ConvBEERS, a lightweight residual convolutional neural network for satellite image restoration. Trained exclusively on simulated degraded imagery, the model is evaluated on both simulated data and real Pleiades-HR acquisitions, claiming a +6.9 dB PSNR gain, up to +5.1% mAP@50 improvement on a downstream object detection task, and successful deployment on a Xilinx Versal VCK190 FPGA yielding ~41x latency reduction versus a traditional physical-model pipeline.

Significance. If the reported gains on real data are reliable, the work would show that simple non-generative CNNs can deliver competitive restoration quality while satisfying the strict latency and resource constraints of onboard satellite processing. The FPGA implementation and downstream task evaluation provide concrete evidence of end-to-end practicality for spaceborne AI pipelines.

major comments (3)
  1. [Abstract and Experiments] The central claims of +6.9 dB PSNR and +5.1% mAP@50 on real Pleiades-HR imagery (Abstract) rest on the unverified assumption that the simulated forward degradation model (blur, noise, atmospheric and sensor effects) produces statistics sufficiently close to actual Pleiades-HR acquisitions. No domain-gap metric, distribution comparison, or ablation isolating simulation fidelity is supplied, leaving the transfer results vulnerable to mismatch.
  2. [Hardware Deployment] The ~41x latency reduction on the Xilinx Versal VCK190 FPGA is reported relative to the traditional pipeline, yet no implementation details of that baseline (specific physical models, optimization flags, or hardware mapping) are given, nor are FPGA resource utilization, power figures, or timing methodology for ConvBEERS itself.
  3. [Evaluation Protocol] The evaluation protocol omits the object-detection architecture used for mAP measurement, the precise training protocol (data splits, augmentations, optimizer settings, and stopping criteria), and any statistical testing or variance reporting for the PSNR and mAP deltas.
minor comments (2)
  1. [Abstract] The abstract uses the phrase 'competitive image quality' without naming the reference methods or tabulating their scores.
  2. [Abstract] The acronym ConvBEERS is expanded only after first use; a parenthetical expansion on first appearance would improve readability.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thorough review and constructive suggestions. We address each of the major comments below, indicating where revisions will be made to the manuscript.

read point-by-point responses

  1. Referee: [Abstract and Experiments] The central claims of +6.9 dB PSNR and +5.1% mAP@50 on real Pleiades-HR imagery (Abstract) rest on the unverified assumption that the simulated forward degradation model (blur, noise, atmospheric and sensor effects) produces statistics sufficiently close to actual Pleiades-HR acquisitions. No domain-gap metric, distribution comparison, or ablation isolating simulation fidelity is supplied, leaving the transfer results vulnerable to mismatch.

    Authors: We agree that the manuscript would benefit from explicit analysis of the simulation-to-real domain gap. Although the degradation model was constructed based on the known characteristics of the Pleiades-HR sensor and acquisition conditions, and the model shows consistent improvements on real imagery, we did not include quantitative domain-gap metrics. In the revised version, we will add a dedicated subsection in the Experiments section providing distribution comparisons (e.g., mean and variance of pixel intensities, frequency domain analysis) and a domain-gap metric such as Fréchet Inception Distance (FID) between simulated and real degraded images. We will also include an ablation study varying the fidelity of the simulation components. revision: yes

  2. Referee: [Hardware Deployment] The ~41x latency reduction on the Xilinx Versal VCK190 FPGA is reported relative to the traditional pipeline, yet no implementation details of that baseline (specific physical models, optimization flags, or hardware mapping) are given, nor are FPGA resource utilization, power figures, or timing methodology for ConvBEERS itself.

    Authors: We acknowledge the lack of detailed hardware implementation information in the current manuscript. The traditional pipeline refers to the sequential application of physical models for deconvolution, denoising, and atmospheric correction as used in the Pleiades ground segment. In the revision, we will expand the Hardware Deployment section to include: (1) specific details on the baseline implementation, including the physical models employed and any optimizations; (2) FPGA resource utilization for ConvBEERS (e.g., LUTs, DSP slices, BRAM); (3) power consumption estimates; and (4) the timing methodology, including clock frequency and latency measurement approach on the VCK190. revision: yes

  3. Referee: [Evaluation Protocol] The evaluation protocol omits the object-detection architecture used for mAP measurement, the precise training protocol (data splits, augmentations, optimizer settings, and stopping criteria), and any statistical testing or variance reporting for the PSNR and mAP deltas.

    Authors: We agree that additional details on the evaluation protocol are necessary for reproducibility. In the revised manuscript, we will specify the object detection architecture (e.g., the particular detector and backbone used), provide the full training protocol including data splits, augmentations, optimizer, learning rate schedule, and stopping criteria, and report variance or standard deviations across multiple runs along with statistical significance tests (e.g., paired t-tests) for the reported PSNR and mAP improvements. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical results from held-out simulated and real data

full rationale

The paper trains a residual CNN (ConvBEERS) on simulated degradations and reports PSNR, mAP, and FPGA latency gains via direct measurement on separate test imagery and hardware execution. No equation, loss term, or architectural choice reduces the claimed +6.9 dB PSNR or +5.1 % mAP improvements to a fitted constant or self-citation by construction. The central claims rest on standard supervised learning and benchmarking, not on any self-definitional loop or imported uniqueness theorem.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on supervised learning of a residual CNN from paired simulated degradations and on the assumption that those simulations capture the statistics of real satellite imagery sufficiently for generalization.

free parameters (2)
  • CNN architecture hyperparameters
    Depth, width, and residual-block design choices are selected to meet onboard latency and resource limits.
  • Training hyperparameters
    Learning rate, batch size, and loss weighting are tuned on simulated data to maximize PSNR on validation pairs.
axioms (2)
  • domain assumption Simulated noise and blur distributions match real Pleiades-HR acquisition statistics
    The paper trains exclusively on simulated pairs and evaluates on real imagery without reporting domain-adaptation steps.
  • domain assumption Standard supervised regression loss is sufficient to learn a restoration mapping that also benefits downstream detection
    No auxiliary loss or perceptual term is mentioned; improvement in mAP is treated as a downstream consequence of PSNR gain.

pith-pipeline@v0.9.0 · 5526 in / 1432 out tokens · 47002 ms · 2026-05-10T16:10:06.288608+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

51 extracted references · 51 canonical work pages

  1. [1]

    Xilinx Doc- umentation, 2026

    Deep learning processor unit (dpu) — vitis ai. Xilinx Doc- umentation, 2026. https://xilinx.github.io/ Vitis-AI/. 7

    work page 2026

  2. [2]

    Physics-informed computer vision: A review and perspectives, 2024

    Chayan Banerjee, Kien Nguyen, Clinton Fookes, and George Karniadakis. Physics-informed computer vision: A review and perspectives, 2024. 8

    work page 2024

  3. [3]

    PhD thesis, 2025

    Maud Biquard.R ´egularisation avec des r´eseaux de neurones pour la restauration d’images satellites. PhD thesis, 2025. Th`ese de doctorat dirig ´ee par Oberlin, Thomas et Chabert, Marie Informatique et T´el´ecommunications Toulouse, ISAE

    work page 2025

  4. [4]

    R ´egularisation par mod `eles pseudo-g´en´eratifs pour la restauration d’images

    Maud Biquard, Marie Chabert, Florence Genin, Christophe Latry, and Thomas Oberlin. R ´egularisation par mod `eles pseudo-g´en´eratifs pour la restauration d’images. InColloque Gretsi 2023, Grenoble, France, 2023. Association Gretsi. 1

    work page 2023

  5. [5]

    Variational bayes image restora- tion with compressive autoencoders.IEEE Transactions on Image Processing, 34:2896–2909, 2025

    Maud Biquard, Marie Chabert, Florence Genin, Christophe Latry, and Thomas Oberlin. Variational bayes image restora- tion with compressive autoencoders.IEEE Transactions on Image Processing, 34:2896–2909, 2025. 2

    work page 2025

  6. [6]

    Deep priors for satellite image restoration with accurate uncertainties.IEEE Transactions on Geoscience and Remote Sensing, 63:1–16, 2025

    Maud Biquard, Marie Chabert, Florence Genin, Christophe Latry, and Thomas Oberlin. Deep priors for satellite image restoration with accurate uncertainties.IEEE Transactions on Geoscience and Remote Sensing, 63:1–16, 2025. 1, 2

    work page 2025

  7. [7]

    The perception-distortion tradeoff

    Yochai Blau and Tomer Michaeli. The perception-distortion tradeoff. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6228–6237,

    work page

  8. [8]

    Rareplanes detection using yolov5 on the versal adaptive soc

    Jacob Brown, Colton Yates, Jeffrey Goeders, and Michael Wirthlin. Rareplanes detection using yolov5 on the versal adaptive soc. In2025 IEEE Aerospace Conference, pages 1–9, 2025. 7

    work page 2025

  9. [9]

    Theses, Universit ´e Nice Sophia Antipolis, 2013

    Mika¨el Carlavan.Optimization of the compression/restoration chain for satellite images. Theses, Universit ´e Nice Sophia Antipolis, 2013. 1, 2

    work page 2013

  10. [10]

    Hat: Hy- brid attention transformer for image restoration, 2025

    Xiangyu Chen, Xintao Wang, Wenlong Zhang, Xiangtao Kong, Yu Qiao, Jiantao Zhou, and Chao Dong. Hat: Hy- brid attention transformer for image restoration, 2025. 2

    work page 2025

  11. [11]

    Chouteau, Laurent Gabet, Renaud Fraisse, T

    F. Chouteau, Laurent Gabet, Renaud Fraisse, T. Bonfort, Bouchra Harnoufi, V . Greiner, M. Goff, M. Ortner, and V . Paveau. Joint super-resolution and image restoration for pl ´Eiades neo imagery.The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLIII-B1-2022:9–15, 2022. 1, 2

    work page 2022

  12. [12]

    Chaine de fusion p+xs des images pl´eiades-hr, 2013

    Sebastien Fourest Christophe Latry, Gwendoline Blancher. Chaine de fusion p+xs des images pl´eiades-hr, 2013. 2, 5

    work page 2013

  13. [13]

    Senti wiki: S2 products, 2025

    ESA Copernicus. Senti wiki: S2 products, 2025. https://sentiwiki.copernicus.eu/web/s2-products, Accessed on February, 2026. 2

    work page 2025

  14. [14]

    Enhanced maritime monitoring via onboard processing of raw multi-spectral imagery by deep learning

    Roberto Del Prete, Gabriele Meoni, Manuel Salvoldi, Domenico Barretta, Maria Daniela Graziano, Nicolas Long´ep´e, and Alfredo Renga. Enhanced maritime monitoring via onboard processing of raw multi-spectral imagery by deep learning. InIGARSS 2024 - 2024 IEEE International Geo- science and Remote Sensing Symposium, pages 1713–1717,

    work page 2024

  15. [15]

    Oil spill identifica- tion from sar images for low power embedded systems using cnn.Remote Sensing, 13(18), 2021

    Lorenzo Diana, Jia Xu, and Luca Fanucci. Oil spill identifica- tion from sar images for low power embedded systems using cnn.Remote Sensing, 13(18), 2021. 1

    work page 2021

  16. [16]

    7nm finfet technology heavy ion sel evaluation using xilinx versal as case study

    Arnaud Dufour, J´erˆome Carron, Franc ¸ois Pierron, Matthieu Fongral, David Dangla, Guillaume Bascoul, Franc ¸oise Bez- erra, Julien Mekki, Florence Malou, and Pierre Maillard. 7nm finfet technology heavy ion sel evaluation using xilinx versal as case study. In2021 21th European Conference on Radia- tion and Its Effects on Components and Systems (RADECS),...

    work page 2021

  17. [17]

    Exploring ai-based satellite pose estimation: from novel syn- thetic dataset to in-depth performance evaluation

    Fabien Gallet, Christophe Marabotto, and Thomas Chambon. Exploring ai-based satellite pose estimation: from novel syn- thetic dataset to in-depth performance evaluation. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 6770–6778,

    work page

  18. [18]

    Garc´es-Socarr´as, Raudel Cuiman, Flor Ortiz, Juan A

    Luis M. Garc´es-Socarr´as, Raudel Cuiman, Flor Ortiz, Juan A. V´asquez-Peralvo, Jorge L. Gonz ´alez-Rios, Mouhamad Chehaitly, Arkadii Kazanskii, Sahar Malmir, Amirhossein Nik, Jan Thoemel, Sumit Kumar, Marcele Kuhfuss, Swetha Varadajulu, Eva Lagunas, Juan C. M. Duncan, Jorge Querol, and Symeon Chatzinotas. Onboard machine learning for satel- lite edge com...

    work page 2025

  19. [19]

    Fidmtf: An mtf estimator (qgis plugin)

    Jorge Gil. Fidmtf: An mtf estimator (qgis plugin). https://plugins.qgis.org/plugins/mtf estimator/. 5

    work page

  20. [20]

    Onboard anomaly detection for marine environmental protection.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:7918– 7931, 2024

    Thomas Goudemant, Benjamin Francesconi, Michelle Aubrun, Erwann Kervennic, Ingrid Grenet, Yves Bobichon, and Marjorie Bellizzi. Onboard anomaly detection for marine environmental protection.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:7918– 7931, 2024. 1

    work page 2024

  21. [21]

    Wavelet deep neural network for stripe noise removal.IEEE Access, 7:44544– 44554, 2019

    Juntao Guan, Rui Lai, and Ai Xiong. Wavelet deep neural network for stripe noise removal.IEEE Access, 7:44544– 44554, 2019. 2

    work page 2019

  22. [22]

    Deep residual learning for image recognition, 2015

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 3

    work page 2015

  23. [23]

    Deep learning- based ship detection on fpgas

    Younis Ibrahim, Li Chen, and Tian Haonan. Deep learning- based ship detection on fpgas. In2022 14th International Conference on Computational Intelligence and Communica- tion Networks (CICN), pages 454–459, 2022. 7

    work page 2022

  24. [24]

    Focal frequency loss for image reconstruction and synthesis,

    Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis,

    work page

  25. [25]

    Inves- tigating loss functions for extreme super-resolution

    Younghyun Jo, Sejong Yang, and Seon Joo Kim. Inves- tigating loss functions for extreme super-resolution. In 2020 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition Workshops (CVPRW), pages 1705–1712,

    work page 2020

  26. [26]

    Hadjim- itsis, Egbert Schwarz, and Stefan Wiehle

    Eleftheria Kalogirou, George Melillos, Diofantos G. Hadjim- itsis, Egbert Schwarz, and Stefan Wiehle. Maritime surveil- lance in Cyprus using Sentinel-1 SAR. InRemote Sensing of the Ocean, Sea Ice, Coastal Waters, and Large Water Regions 2023, page 127280N, 2023. 1

    work page 2023

  27. [27]

    Yolov11: An overview of the key architectural enhancements, 2024

    Rahima Khanam and Muhammad Hussain. Yolov11: An overview of the key architectural enhancements, 2024. 6

    work page 2024

  28. [28]

    Gen- erative ai

    Georges Labr`eche, Cesar Guzman, and Sam Bammens. Gen- erative ai... in space! adversarial networks to denoise images onboard the ops-sat-1 spacecraft. In2024 IEEE Aerospace Conference, pages 1–17, 2024. 2

    work page 2024

  29. [29]

    Latry, S

    C. Latry, S. Fourest, and C. Thiebaut. Restoration tech- nique for pleiades-hr panchromatic images.The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XXXIX-B1:555–560, 2012. 1, 2, 5, 7

    work page 2012

  30. [30]

    Im- plementation of the ”Non-Local Bayes” (NL-Bayes) Image Denoising Algorithm.Image Processing On Line, 3:1–42,

    Marc Lebrun, Antoni Buades, and Jean-Michel Morel. Im- plementation of the ”Non-Local Bayes” (NL-Bayes) Image Denoising Algorithm.Image Processing On Line, 3:1–42,

    work page

  31. [31]

    https://doi.org/10.5201/ipol.2013.16. 1, 2

    work page doi:10.5201/ipol.2013.16 2013

  32. [32]

    Photo-realistic single image super-resolution using a gen- erative adversarial network

    Christian Ledig, Lucas Theis, Ferenc Husz´ar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a gen- erative adversarial network. InProceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition (CVPR), pages ...

    work page 2017

  33. [33]

    Dpuv4e: High-throughput dpu architecture design for cnn on versal acap, 2025

    Guoyu Li, Pengbo Zheng, Jian Weng, and Enshan Yang. Dpuv4e: High-throughput dpu architecture design for cnn on versal acap, 2025. 7

    work page 2025

  34. [34]

    Object detection in optical remote sensing images: A survey and a new benchmark.ISPRS Journal of Photogrammetry and Remote Sensing, 159:296–307, 2020

    Ke Li, Gang Wan, Gong Cheng, Liqiu Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark.ISPRS Journal of Photogrammetry and Remote Sensing, 159:296–307, 2020. 6

    work page 2020

  35. [35]

    Enhanced Deep Residual Networks for Single Image Super-Resolution

    Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution.CoRR, abs/1707.02921, 2017. 2, 3

    work page Pith review arXiv 2017

  36. [36]

    Liu, L.F

    J.G. Liu, L.F. Zhang, and Q.X. Tong. Estimation of signal to noise ratio of remote sensing images.Journal of Basic Science and Engineering, 4:360–365, 1999. 5

    work page 1999

  37. [37]

    Gabriele Meoni, Roberto Del Prete, Federico Serva, Alix De Beusscher, Olivier Colin, and Nicolas Long´ep´e. Unlock- ing the use of raw multispectral earth observation imagery for onboard artificial intelligence.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:12521–12537, 2024. 1, 2, 8

    work page 2024

  38. [38]

    Virtanen, Ari-Pekka Jokinen, Joni Koskikala, Sakari V¨akev¨a, and Jenni Attila

    Janne M ¨ayr¨a, Elina A. Virtanen, Ari-Pekka Jokinen, Joni Koskikala, Sakari V¨akev¨a, and Jenni Attila. Mapping recre- ational marine traffic from sentinel-2 imagery using yolo object detection models.Remote Sensing of Environment, 326:114791, 2025. 1

    work page 2025

  39. [39]

    Openaerialmap

    OpenAerialMap. Openaerialmap. https : //openaerialmap.org. 4

    work page

  40. [40]

    Lvrnet: Lightweight image restoration for aerial images under low visibility, 2023

    Esha Pahwa, Achleshwar Luthra, and Pratik Narang. Lvrnet: Lightweight image restoration for aerial images under low visibility, 2023. 2

    work page 2023

  41. [41]

    Eval- uation of xilinx versal architecture for next-gen edge com- puting in space

    Noah Perryman, Christopher Wilson, and Alan George. Eval- uation of xilinx versal architecture for next-gen edge com- puting in space. In2023 IEEE Aerospace Conference, pages 1–11, 2023. 7

    work page 2023

  42. [42]

    Enhancing maritime situational awareness through end-to-end onboard raw data analysis, 2024

    Roberto Del Prete, Manuel Salvoldi, Domenico Bar- retta, Nicolas Long ´ep´e, Gabriele Meoni, Arnon Karnieli, Maria Daniela Graziano, and Alfredo Renga. Enhancing maritime situational awareness through end-to-end onboard raw data analysis, 2024. 1

    work page 2024

  43. [43]

    Enhancing cnn inference time and reducing latency on edge and resource-constrained systems through quantiza- tion

    Gaurav Upadhyay, C Siva Prasad, V Akash, Ayush Kumar Singh, R Monish Sai, Srijan Ghosal, Subrat Kar, S H Shan- tala, J Thakur LalitKrushna, S Srividhya, and Sunil Alias Bal- wantrao. Enhancing cnn inference time and reducing latency on edge and resource-constrained systems through quantiza- tion. In2024 Second International Conference on Networks, Multime...

    work page 2024

  44. [44]

    Spatial-frequency mutual learning for face super- resolution

    Chenyang Wang, Junjun Jiang, Zhiwei Zhong, and Xian- ming Liu. Spatial-frequency mutual learning for face super- resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22356–22366, 2023. 3

    work page 2023

  45. [45]

    Optical imaging degradation simulation and transformer-based image restoration for remote sensing

    Hua Wei, Kun Gao, Jing Wang, Qiuyan Tang, Xiongxin Tang, and Fanjiang Xu. Optical imaging degradation simulation and transformer-based image restoration for remote sensing. IEEE Geoscience and Remote Sensing Letters, 21:1–5, 2024. 2

    work page 2024

  46. [46]

    Deep memory connected neural network for optical remote sensing image restoration.Remote Sensing, 10(12), 2018

    Wenjia Xu, Guangluan Xu, Yang Wang, Xian Sun, Daoyu Lin, and Yirong Wu. Deep memory connected neural network for optical remote sensing image restoration.Remote Sensing, 10(12), 2018. 2

    work page 2018

  47. [47]

    Scalable neural pushbroom architectures for real-time denoising of hyperspectral images onboard satel- lites, 2026

    Ziyao Yi, Davide Piccinini, Diego Valsesia, Tiziano Bianchi, and Enrico Magli. Scalable neural pushbroom architectures for real-time denoising of hyperspectral images onboard satel- lites, 2026. 2

    work page 2026

  48. [48]

    Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising.IEEE Transactions on Image Processing, 26(7):3142–3155, 2017

    Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising.IEEE Transactions on Image Processing, 26(7):3142–3155, 2017. 3

    work page 2017

  49. [49]

    Learning deep cnn denoiser prior for image restoration, 2017

    Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration, 2017. 3

    work page 2017

  50. [50]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 3

    work page 2018

  51. [51]

    Denoising diffusion models for plug-and-play image restoration, 2023

    Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising diffusion models for plug-and-play image restoration, 2023. 2

    work page 2023