Pith. sign in

REVIEW 3 major objections 6 minor 88 references

CoRE-UIR claims that decomposing an all-in-one restorer into a shared dense expert and sparsely routed low-rank residual experts beats full-rank mixture-of-experts designs in quality per unit of compute.

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

2026-07-31 23:22 UTC pith:A2AZOAQY

load-bearing objection Solid all-in-one restoration system with a new benchmark, but the headline margin comes mostly from the frozen CLIP prior, not the CoRE block; still worth a serious referee. the 3 major comments →

arxiv 2607.27898 v1 pith:A2AZOAQY submitted 2026-07-30 cs.CV

CoRE-UIR: Prior-guided common and residual experts for efficient all-in-one remote sensing image restoration

Zaiyan Zhang , Qiangqiang Yuan , Jie Li , Ziyang Lihe , Yu Wan , Yuzeng Chen , Xin Su , Liangpei Zhang This is my paper
classification cs.CV
keywords all-in-one image restorationremote sensingmixture of expertslow-rank adaptationdegradation priorCLIPUAV imagerycompound degradation
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.

CoRE-UIR is a single network that restores remote sensing images degraded by fog, dust, rain, low light, motion blur, defocus blur, and their compound combinations. The paper's central claim is that restoration capacity should be split into a common dense expert that handles degradation-invariant operations (contrast, texture, structure) and low-rank residual experts that add degradation-specific corrections, selected by a prototype-guided router. A frozen CLIP encoder adapted by a lightweight adapter supplies the prior that modulates global feature statistics and routes local experts. On UAV imagery (MDVD-108K) the method gains 1.05 dB PSNR over the strongest baseline while running 11.83x faster and using 85.3% less peak memory, and it transfers to the satellite benchmark MDRS-Landsat. The significance is that all-in-one restoration models do not need to pay the full cost of replicating full-rank experts because most restoration behavior is shared across degradations and can be consolidated.

Core claim

The central discovery is that the common-residual expert decomposition yields a better quality-efficiency frontier than full-rank MoE. The paper argues that existing all-in-one restorers waste capacity because full-rank expert branches re-learn the same shared operations such as visibility enhancement, texture refinement, and brightness normalization. CoRE-UIR instead keeps the dense backbone block as an always-active common expert and adds a small library of low-rank residual experts (bottleneck rank 4) that are sparsely activated via cosine-similarity routing against learnable degradation prototypes. With this design, the model reaches 30.67 dB average PSNR on the compound MDVD-108K benchm

What carries the argument

The Common-and-Residual Expert Block (CoRE) is the primary structural innovation: it couples the original backbone block (the common dense expert) with N=6 low-rank residual expert pairs (A_n, B_n) that map channels down to rank r=4 and back, aggregated with Top-k=3 routing weights from the Prototype-Guided Router (PG-Router), which scores cosine similarity between the degradation embedding and learnable prototypes. Degradation Prior Embedding (DPE) adapts a frozen CLIP ViT-B/32 (global plus local views) through a two-layer adapter into a 384-dim restoration-oriented embedding; Global Feature Modulation (GFM) performs stage-entrance channel recalibration combining the prior with the current

Load-bearing premise

The entire system rests on the frozen CLIP encoder's embeddings of a global view and a local crop containing enough information to discriminate and control all six degradation types and their compounds, and to transfer to real-world degradations; the paper's own Table 9 shows PSNR collapses from 30.76 to 17.73 dB when the prior is wrong.

What would settle it

Feed CoRE-UIR a degradation that is visually distinct from the six training types but falls inside CLIP's semantic space, e.g., heavy snowfall on UAV imagery; if the PG-Router activates the wrong prototypes and PSNR drops several dB below a task-specific snow-removal model, the claim that CLIP priors span the needed degradation space is falsified. A simpler check: adversarially corrupt the DPE input so the predicted degradation label flips (as in Table 9's wrong bucket) and measure the PSNR drop; a drop of more than ~5 dB would confirm that quality is hostage to prior reliability rather than b

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

If this is right

  • A single CoRE-UIR model matches or exceeds task-specific models on each of six single degradations and six compound degradations in the UAV benchmark, suggesting that unified restorers need not sacrifice per-task accuracy.
  • On the satellite benchmark MDRS-Landsat, CoRE-UIR reaches the best overall average PSNR (40.22 dB) and best SSIM/LPIPS among all-in-one methods, including remote-sensing-specific ones, with the largest gains in darkening and haze removal.
  • Restoration with CoRE-UIR transfers to downstream object detection: a detector trained on clean images recovers 75.7% of the mAP50 gap and 73.4% of the mAP50:95 gap on synthetic test images, and improves detection on real-world degraded UAV images.
  • The common-residual decomposition keeps the model robust to unseen compound degradations (weather + defocus blur) better than baselines, indicating the low-rank residuals encode factorized degradation knowledge rather than memorized pairs.
  • The system's efficiency (11.83x faster and 85.3% less memory than BaryIR) suggests that prior-guided routing plus low-rank compensation is a scalable way to deploy all-in-one restoration on platforms with limited compute.

Where Pith is reading between the lines

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

  • The same common-residual decomposition could be applied to other universal restoration domains (e.g., natural image restoration, video restoration) where a single degradation-invariant backbone plus low-rank task experts would replace full-rank branches; the quality-efficiency gains are likely to transfer since the paper's argument is domain-agnostic, though the CLIP-prior assumption is tested her
  • A testable extension suggested by the paper's own analysis: the system's sensitivity to prior errors (Table 9) implies that making DPE uncertainty-aware — e.g., routing on a distribution over prototypes rather than a single embedding, or explicitly training the router to handle misclassified priors — could recover most of the lost PSNR in partial/wrong-prior cases.
  • Because the Phase-I classification head is discarded after training, the prior embedding is not directly supervised during restoration; an end-to-end variant where the DPE adapter is fine-tuned together with the backbone (with the CLIP encoder still frozen) might close the remaining gap to the exact-match oracle.
  • The paper evaluates downstream utility with object detection only; extending the same restoration output to semantic segmentation or change detection would test whether the common-residual specialization preserves the geometric and spectral cues those tasks require.

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

3 major / 6 minor

Summary. The paper proposes CoRE-UIR, a unified framework for all-in-one remote sensing image restoration. It combines a frozen CLIP-based Degradation Prior Embedding (DPE), a Global Feature Modulation (GFM) module, and a Common-and-Residual Expert (CoRE) block that couples an always-active dense expert with top-k selected low-rank residual experts. The model is trained in two phases: DPE is first adapted via a multi-label degradation classification objective, then frozen while the restoration backbone is trained. The authors also introduce MDVD-108K, a large-scale UAV degradation dataset with six single and six compound degradations plus a real-world qualitative test set, and evaluate on MDRS-Landsat. They report an average PSNR improvement of 1.05 dB over BaryIR with 11.83× lower latency and 85.3% lower peak memory.

Significance. If the reported results hold, the paper would be a solid contribution to efficient all-in-one image restoration in remote sensing, with a useful new dataset and a plausible mechanism for reducing expert redundancy. The strongest aspects are the comprehensive ablation suite, the large dataset, the downstream detection evaluation, and the clear architectural presentation. However, the central attribution of the performance gain to the CoRE block is not yet established: the frozen CLIP prior accounts for a larger share of the improvement than the CoRE block itself, and no baseline is provided with the same DPE/GFM conditioning. The efficiency claim also depends on measurement protocol. With additional controlled experiments, the contribution could be significant.

major comments (3)
  1. [§4.3.1/Table 5(b), §4.3.3/Table 7(a), §4.4.1/Table 8] The paper's central claim is that the common-residual expert decomposition (CoRE) yields the reported quality-efficiency trade-off. However, the ablations indicate that the largest component of the improvement over BaryIR comes from the frozen CLIP prior rather than CoRE. In Table 5(b), replacing the CLIP encoder with a trainable ResNet-50 reduces PSNR from 30.67 to 29.51 dB, a drop of 1.16 dB, which exceeds the entire 1.05 dB margin over BaryIR in Table 8. In Table 7(a), CoRE itself adds only 0.31 dB over the common-expert-only variant (30.36→30.67), and a full-rank FFN MoE is only 0.16 dB behind CoRE (30.51) despite using 2.3× more parameters. Because no baseline (BaryIR or MoCE-IR) is augmented with the same DPE/GFM conditioning, the reported advantage cannot be attributed to CoRE. The authors should either provide controlled experiments that ablate DPE/GFM on the strongest baselines
  2. [§4.4.2/Table 9, §4.2.3] Table 9 shows that restoration quality is critically dependent on DPE correctness: samples with a wrong DPE prediction drop to 17.73 dB PSNR, far below the 30.76 dB for exact matches. Since DPE is frozen in Phase II and the classification head is discarded, the model has no self-assessment or fallback when the prior is incorrect. The real-world evaluation (§4.2.3) is only qualitative on 500 unlabeled images, so the claim of generalizability to real-world degradations is not quantitatively supported. The authors should either provide a quantitative real-world benchmark (e.g., reference-based metrics on a subset with manual or pseudo-GT) or add a confidence/uncertainty mechanism to flag low-confidence DPE predictions.
  3. [§4.4.1/Table 8, §4.1.4] The efficiency comparison may be partially confounded by the use of 16-bit AMP for the DPE's CLIP encoder while other methods 'follow their official settings.' Since the 11.83× speedup over BaryIR is a headline result, the authors should report whether AMP is also applied to the baseline methods or whether the speedup persists under identical precision settings. Additionally, the memory measurement (1.45 GB) should clarify whether it includes the frozen CLIP encoder parameters and activations; the total parameter count of 110.7M vs. BaryIR's 53.5M suggests the memory advantage is nontrivial, but the paper should confirm it is not an artifact of precision.
minor comments (6)
  1. [§3.1.1] The mathematical notation for the degradation operator appears garbled in the rendering (e.g., stray '\left' commands in Eq. 1). Please check the LaTeX.
  2. [§3.4.2] In Eq. (14), the notation for L2-normalized vectors is dense. Consider clarifying that all vectors are normalized before the cosine-similarity computation and that τ is a learned temperature.
  3. [§4.3.3/Table 7(c)] The differences among sample-level, stage-level, and block-level routing are small (30.52, 30.67, 30.47 dB). Reporting standard deviations or significance would strengthen the claim that stage-level routing is best.
  4. [§4.4.2/Table 9] The 'wrong' bucket contains only 14 samples (0.13%). The 17.73 dB PSNR is based on a very small n and should be interpreted with caution; consider reporting confidence intervals or individual sample results.
  5. [§4.1.4] The AdamW optimizer is cited as Loshchilov and Hutter 2017a; the standard reference is ICLR 2019. Please update the reference.
  6. [Figure 12] The router-weight heatmaps are difficult to read at print size. Increasing font size or using a separate table for the activation values would improve clarity.

Circularity Check

0 steps flagged

No circular derivation; the only self-citations are contextual and non-load-bearing.

full rationale

CoRE-UIR is an empirical architecture paper. Its Eqs. (3)-(24) define a network and training objectives rather than deriving a prediction from fitted inputs, so no reduction-by-construction is apparent. The main potential candidate is DPE, which is trained in Phase I with multi-label degradation classification; however, the paper explicitly discards that head: 'After convergence, the classification head is discarded and the resulting DPE is used to condition both GFM and CoRE during restoration learning.' The restoration target is the clean image, not the degradation label, and test labels are never used to fit DPE or the router. Table 9's stratification by exact/partial/wrong prior prediction is an honest diagnostic, not a hidden training signal. The only author-overlapping citation is Ada4DIR/Lihe et al. (2025), used as a published benchmark and contextual related work; it is not an unverified uniqueness theorem or a load-bearing argument. The skeptic's point that the 1.05 dB margin may be confounded by the frozen CLIP prior is an empirical attribution concern, not circularity: it is falsifiable from the paper's own ablations (Table 5) rather than being true by construction. The conclusion's statement that 'CoRE-UIR still mainly addresses known degradation factors' further scopes the claims without circularity. Overall, no step equates an input with an output by definition, and no fitted parameter is renamed as a prediction. Score 2 reflects only the presence of minor, non-load-bearing self-citations.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The paper's central claims rest on the expressiveness of CLIP features for degradation, the low-rank structure of residual corrections, and the representativeness of the synthetic degradation models. The listed free parameters are hand-chosen hyperparameters that affect performance but are not fitted to the test set. No new physical entities are introduced.

free parameters (6)
  • Number of residual experts N = 6
    Chosen to match the number of base degradation types; set by hand, not fitted.
  • Low-rank bottleneck rank r = 4
    Chosen as a small rank for residual experts; central to the low-rank assumption.
  • Top-k activated experts k = 3
    Ablated; k=3 gives best trade-off. Set by hand.
  • GFM bottleneck reduction ratio rho = 16
    Architectural hyperparameter chosen without systematic tuning.
  • Degradation embedding dimension d_z = 384
    Chosen to match CLIP feature dimension after adapter; set by hand.
  • Loss weights (pixel/structural/perceptual) = 0.6/0.2/0.2
    Set without a sensitivity study; affects restoration trade-off.
axioms (4)
  • domain assumption Frozen CLIP ViT-B/32 features contain transferable degradation cues for remote sensing restoration
    The entire DPE design depends on CLIP's ability to encode degradation-relevant information; Table 9 shows catastrophic failure when the prior is wrong.
  • domain assumption Degradation-specific corrections are approximately low-rank, so a rank-4 residual branch suffices
    The CoRE block assumes that residual compensation can be captured by low-rank projections; no theoretical justification is given.
  • domain assumption The prototype-guided router with cosine similarity and k=3 generalizes to unseen compound degradations
    The router is trained on six base degradations and seen compound pairs; generalization to unseen pairs is assumed and tested only for a subset.
  • domain assumption The synthetic degradation models (fog, rain, dust, low-light, blur) are representative of real UAV degradations
    The benchmark is synthetic; real-world validation is qualitative only, yet the central performance claims are measured on this synthetic distribution.

pith-pipeline@v1.3.0-daily-deepseek · 32022 in / 10734 out tokens · 103482 ms · 2026-07-31T23:22:56.519327+00:00 · methodology

0 comments
read the original abstract

Remote sensing images acquired by unmanned aerial vehicles (UAVs) and satellites are often degraded by adverse weather, illumination variation, and imaging artifacts, which may co-occur and jointly induce global distribution shifts and local structural corruption. Although All-in-One image restoration offers an appealing unified alternative to task-specific pipelines, existing methods still suffer from weak or implicit degradation cues and parameter redundancy caused by full-rank multi-expert designs with overlapping restoration behaviors. We propose CoRE-UIR (Common and Residual Experts for Universal Image Restoration), a prior-guided global-local framework centered on the Common-and-Residual Expert Block (CoRE). CoRE explicitly decomposes restoration capacity into a common dense expert for degradation-invariant restoration and low-rank residual experts for degradation-specific compensation, enabling adaptive specialization without redundant expert replication. Built on this design, Degradation Prior Embedding (DPE) adapts frozen CLIP features into an explicit restoration-oriented prior, while Global Feature Modulation (GFM) aligns global feature statistics before local residual compensation. We also construct MDVD-108K (Multi-Degradation VisDrone), a large-scale UAV restoration dataset covering both single and compound degradations, together with a real-world test set. Extensive experiments on multiple datasets show that CoRE-UIR improves the overall average PSNR by 1.05 dB while running 11.83$\times$ faster and reducing peak memory by 85.3% relative to the strongest baseline, BaryIR, thereby maintaining a favorable quality-efficiency trade-off. Evaluations on downstream tasks and unseen degradation also validate the generalizability of CoRE-UIR. The code and dataset will be released at https://github.com/zzaiyan/CoRE-UIR.

Figures

Figures reproduced from arXiv: 2607.27898 by Jie Li, Liangpei Zhang, Qiangqiang Yuan, Xin Su, Yu Wan, Yuzeng Chen, Zaiyan Zhang, Ziyang Lihe.

Figure 1
Figure 1. Figure 1: The overall framework of the proposed CoRE-UIR. (a) Degradation Prior Embedding (DPE): a frozen CLIP image encoder processes multi-scale views of the input, and a lightweight adapter maps the features into restoration-oriented degradation embeddings. (b) Global-Local Adaptive Network: the unified restoration backbone integrates Global Feature Modulation (GFM) and Common-and-Residual Expert Blocks (CoRE). (… view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the proposed Global Feature Modula￾tion (GFM) module. GFM combines the degradation prior with the current feature state to perform lightweight degradation￾aware channel recalibration before CoRE processing. where both views are scaled to the CLIP encoder’s native input resolution. The global view 𝐼g preserves overall degra￾dation context, while the local view 𝐼l captures regional degradatio… view at source ↗
Figure 3
Figure 3. Figure 3: Architecture of the proposed Prototype-Guided Router (PG-Router), which selects the most relevant low-rank residual experts for each input sample based on degradation archetypes. ℝ𝐶×𝑟 are learnable projection matrices with bottleneck rank 𝑟 ≪ 𝐶, applying to the channel dimension at each spatial location. Each pair, therefore, forms a compact low-rank residual expert. In implementation, 𝐴𝑛 and 𝐵𝑛 are imple￾… view at source ↗
Figure 4
Figure 4. Figure 4: Representative samples of the MDVD-108K and MDRS-Landsat datasets, showing (a) single degradations UAV samples, (b) compound degradations UAV samples, and (c) satellite samples from MDRS-Landsat. Single Avg Fog Dust Low-light Rain Motion Defocus MDVD-108K Single Compound Avg Dust+ Low-light Dust+ Motion Fog+ Low-light Fog+ Motion Rain+ Low-light Rain+ Motion MDVD-108K Compound Avg Blur Haze Dark Noise MDRS… view at source ↗
Figure 5
Figure 5. Figure 5: Radar-chart comparison of PSNR for representative algorithms. From left to right, the three panels summarize MDVD-108K single degradations, MDVD-108K compound degradations, and MDRS-Landsat. Larger radius indicates higher PSNR. evaluates structural consistency in terms of luminance, con￾trast, and local pattern similarity, and LPIPS (Zhang et al., 2018a) measures perceptual distance in a deep feature space… view at source ↗
Figure 6
Figure 6. Figure 6: Visual comparison on weather-induced degradation cases from MDVD-108K. CoRE-UIR produces more natural visibility, contrast, and color than prior universal restoration baselines, which is consistent with the role of DPE and GFM in organizing global appearance modulation under weather-induced degradations. Zhang et al.: Preprint submitted to Elsevier Page 9 of 23 [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visual comparison on imaging-induced degradation cases from MDVD-108K. CoRE-UIR produces more uniform brightness correction and sharper structural recovery than prior universal restoration baselines, reflecting the benefit of low-rank residual specialization for local degradation compensation. degradation types, so compound samples can activate mul￾tiple labels. Precision, Recall, and F1 are macro-averaged… view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative comparison for compound degradation cases. CoRE-UIR produces cleaner structures and fewer residual compound artifacts than prior universal restoration baselines under overlapping degradations, where the common dense path handles common restoration while low-rank residual experts compensate degradation-specific residuals. randomly cropped to 256 × 256 patches. Data augmentation includes random h… view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison for real-world degradation results. CoRE-UIR preserves land-cover structures while suppressing authentic degradation artifacts more reliably than competing universal restoration methods. 4.2.4. Satellite-Domain Evaluation We further use MDRS-Landsat for satellite-domain eval￾uation, where the comparison is stronger because Ada4DIR￾d and PhyDAE are both tailored to remote sensing rest… view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison on MDRS-Landsat. The panel illustrates that CoRE-UIR preserves large-scale structures while suppressing domain-shifted degradation artifacts more reliably than competing universal restoration methods. 4.3.1. Degradation Prior Embedding We first examine whether DPE provides priors that are both discriminative and useful for restoration [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Feature visualization for DPE and GFM, showing (a) Raw CLIP features, (b) Adapted DPE priors, (c) Features before GFM, and (d) Features after GFM. (a)–(b) demonstrate that DPE improves the degradation discriminativeness of the raw CLIP feature, while (c)–(d) show that GFM reshapes the restoration feature space into a more compact and routable manifold. Prior encoder [PITH_FULL_IMAGE:figures/full_fig_p015… view at source ↗
Figure 12
Figure 12. Figure 12: Feature visualization for the first CoRE module at stage 2. From left to right, we show Input, CoRE Input Feature, Common Expert Feature, Router Weights, Residual Experts Feature, CoRE Output Feature, and Prediction. The common expert exhibits similar responses across degradations and builds a shared restoration basis, while the residual experts focus on degradation-specific compensation such as fog depth… view at source ↗
Figure 13
Figure 13. Figure 13: Quality-efficiency trade-off on MDVD-108K. Left: PSNR vs. inference time. Right: PSNR vs. peak memory. Both horizontal axes are logarithmic, and CoRE-UIR stays near the upper-left frontier in both views. but the set is incomplete or contains extra labels; wrong means that no ground-truth degradation label is recovered. Zhang et al.: Preprint submitted to Elsevier Page 17 of 23 [PITH_FULL_IMAGE:figures/fu… view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative downstream detection examples on syn￾thetic motion blur and real low-light scenes. Compared with degraded inputs, CoRE-UIR restores clearer object boundaries and illumination, yielding more complete detector responses [PITH_FULL_IMAGE:figures/full_fig_p018_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Synthetic-real comparison in the DPE feature space. Colored points are synthetic degradations and red stars are real UAV samples. The broader synthetic manifold covers the real samples. space [PITH_FULL_IMAGE:figures/full_fig_p021_15.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

88 extracted references · 1 canonical work pages

  1. [1]

    arXiv preprint arXiv:2410.15385 , year =

    LoRA-IR: Taming low-rank experts for efficient all-in-one image restoration , author =. arXiv preprint arXiv:2410.15385 , year =

    Pith/arXiv arXiv

  2. [2]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Retinexformer: One-stage retinex-based transformer for low-light image enhancement , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  3. [3]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages =

    Pre-trained image processing transformer , author =. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages =

  4. [4]

    arXiv preprint arXiv:2603.04114 , year =

    Any2any: Unified arbitrary modality translation for remote sensing , author =. arXiv preprint arXiv:2603.04114 , year =

    Pith/arXiv arXiv

  5. [5]

    Proceedings of the European Conference on Computer Vision , pages =

    Simple baselines for image restoration , author =. Proceedings of the European Conference on Computer Vision , pages =. 2022 , organization =

    2022

  6. [6]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages =

    Activating more pixels in image super-resolution transformer , author =. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages =

  7. [7]

    IEEE Transactions on Image Processing , year =

    Hyperspectral Video Tracking with Spectral-Spatial Fusion and Memory Enhancement , author =. IEEE Transactions on Image Processing , year =

  8. [8]

    Advances in Neural Information Processing Systems , volume =

    Cross aggregation transformer for image restoration , author =. Advances in Neural Information Processing Systems , volume =

  9. [9]

    Cloud removal for remotely sensed images by similar pixel replacement guided with a spatio-temporal MRF model , volume =

    Cheng, Qing and Shen, Huanfeng and Zhang, Liangpei and Yuan, Qiangqiang and Zeng, Chao , year =. Cloud removal for remotely sensed images by similar pixel replacement guided with a spatio-temporal MRF model , volume =

  10. [10]

    arXiv preprint arXiv:2604.05629 , year =

    A Unified Foundation Model for All-in-One Multi-Modal Remote Sensing Image Restoration and Fusion with Language Prompting , author =. arXiv preprint arXiv:2604.05629 , year =

    Pith/arXiv arXiv

  11. [11]

    ISPRS Journal of Photogrammetry and Remote Sensing , volume =

    Pansharpening via predictive filtering with element-wise feature mixing , author =. ISPRS Journal of Photogrammetry and Remote Sensing , volume =. 2025 , publisher =

    2025

  12. [12]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Enpowering Your Pansharpening Models with Generalizability: Unified Distribution is All You Need , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  13. [13]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    Image super-resolution using deep convolutional networks , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =. 2015 , publisher =

    2015

  14. [14]

    IEEE Transactions on Geoscience and Remote Sensing , year =

    PhyDAE: Physics-Guided Degradation-Adaptive Experts for All-in-One Remote Sensing Image Restoration , author =. IEEE Transactions on Geoscience and Remote Sensing , year =

  15. [15]

    IEEE Transactions on Circuits and Systems for Video Technology , volume =

    Frequency-oriented efficient transformer for all-in-one weather-degraded image restoration , author =. IEEE Transactions on Circuits and Systems for Video Technology , volume =. 2023 , publisher =

    2023

  16. [16]

    European conference on computer vision , pages =

    Mambair: A simple baseline for image restoration with state-space model , author =. European conference on computer vision , pages =. 2024 , organization =

    2024

  17. [17]

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

    Deep residual learning for image recognition , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =

  18. [18]

    IEEE transactions on geoscience and remote sensing , volume =

    Total-variation-regularized low-rank matrix factorization for hyperspectral image restoration , author =. IEEE transactions on geoscience and remote sensing , volume =. 2015 , publisher =

    2015

  19. [19]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages =

    Squeeze-and-excitation networks , author =. Proceedings of the IEEE conference on computer vision and pattern recognition , pages =

  20. [20]

    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

  21. [21]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    A survey on all-in-one image restoration: Taxonomy, evaluation and future trends , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

  22. [22]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year =

    MB-TaylorFormer V2: Improved multi-branch linear transformer expanded by Taylor formula for image restoration , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , year =

  23. [23]

    Remote Sensing , volume =

    Denoising diffusion probabilistic feature-based network for cloud removal in Sentinel-2 imagery , author =. Remote Sensing , volume =. 2023 , publisher =

    2023

  24. [24]

    arXiv preprint arXiv:2410.17725 , year =

    Yolov11: An overview of the key architectural enhancements , author =. arXiv preprint arXiv:2410.17725 , year =

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2401.03379 , year =

    Towards effective multiple-in-one image restoration: A sequential and prompt learning strategy , author =. arXiv preprint arXiv:2401.03379 , year =

    Pith/arXiv arXiv

  26. [26]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    All-in-one image restoration for unknown corruption , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  27. [27]

    IEEE Transactions on Image Processing , volume =

    Benchmarking single-image dehazing and beyond , author =. IEEE Transactions on Image Processing , volume =. 2019 , publisher =

    2019

  28. [28]

    Remote Sensing of Environment , volume =

    CloudRuler: Rule-based transformer for cloud removal in Landsat images , author =. Remote Sensing of Environment , volume =. 2025 , publisher =

    2025

  29. [29]

    ISPRS Journal of Photogrammetry and Remote Sensing , volume =

    Thin cloud removal in optical remote sensing images based on generative adversarial networks and physical model of cloud distortion , author =. ISPRS Journal of Photogrammetry and Remote Sensing , volume =. 2020 , publisher =

    2020

  30. [30]

    Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops , pages =

    SwinIR: Image Restoration Using Swin Transformer , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops , pages =

  31. [31]

    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

  32. [32]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    Fan Liu and Delong Chen and Zhangqingyun Guan and Xiaocong Zhou and Jiale Zhu and Qiaolin Ye and Liyong Fu and Jun Zhou , title =. IEEE Transactions on Geoscience and Remote Sensing , volume =

  33. [33]

    CrossMatch: Cross-View Matching for Semi-Supervised Remote Sensing Image Segmentation , year =

    Liu, Ruizhong and Luo, Tingzhang and Huang, Shaoguang and Wu, Yuwei and Jiang, Zhen and Zhang, Hongyan , journal =. CrossMatch: Cross-View Matching for Semi-Supervised Remote Sensing Image Segmentation , year =

  34. [34]

    IEEE Transactions on Image Processing , year =

    IHDCP: Single image dehazing using inverted haze density correction prior , author =. IEEE Transactions on Image Processing , year =

  35. [35]

    International Conference on Learning Representations , year =

    Decoupled Weight Decay Regularization , author =. International Conference on Learning Representations , year =

  36. [36]

    International Conference on Learning Representations , year =

    SGDR: Stochastic gradient descent with warm restarts , author =. International Conference on Learning Representations , year =

  37. [37]

    International Conference on Learning Representations , pages =

    Controlling vision-language models for universal image restoration , author =. International Conference on Learning Representations , pages =

  38. [38]

    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

  39. [39]

    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

  40. [40]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Deep generalized unfolding networks for image restoration , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  41. [41]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    Restoring vision in adverse weather conditions with patch-based denoising diffusion models , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =. 2023 , publisher =

    2023

  42. [42]

    2024 , issn =

    HDRSA-Net: Hybrid dynamic residual self-attention network for SAR-assisted optical image cloud and shadow removal , journal =. 2024 , issn =

    2024

  43. [43]

    Advances in Neural Information Processing Systems , year =

    PromptIR: Prompting for all-in-one image restoration , author =. Advances in Neural Information Processing Systems , year =

  44. [44]

    Proceedings of the International Conference on Machine Learning , pages =

    Learning transferable visual models from natural language supervision , author =. Proceedings of the International Conference on Machine Learning , pages =

  45. [45]

    Advances in Neural Information Processing Systems , volume =

    Scaling vision with sparse mixture of experts , author =. Advances in Neural Information Processing Systems , volume =

  46. [46]

    International Conference on Learning Representations , year =

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer , author =. International Conference on Learning Representations , year =

  47. [47]

    IEEE Geoscience and Remote Sensing Magazine , volume =

    Missing information reconstruction of remote sensing data: A technical review , author =. IEEE Geoscience and Remote Sensing Magazine , volume =. 2015 , publisher =

    2015

  48. [48]

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

    Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network , author =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages =

  49. [49]

    Remote Sensing of Environment , volume =

    RESTORE-DiT: Reliable satellite image time series reconstruction by multimodal sequential diffusion transformer , author =. Remote Sensing of Environment , volume =. 2025 , publisher =

    2025

  50. [50]

    IEEE Transactions on Image Processing , year =

    Vision Transformers for Single Image Dehazing , author =. IEEE Transactions on Image Processing , year =

  51. [51]

    IEEE Trans

    U-TILISE: A Sequence-to-Sequence Model for Cloud Removal in Optical Satellite Time Series , author =. IEEE Trans. Geosci. Remote Sens. , volume =. 2023 , publisher =

    2023

  52. [52]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    Diffusion Enhancement for Cloud Removal in Ultra-Resolution Remote Sensing Imagery , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2024 , publisher =

    2024

  53. [53]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year =

    Learning Continuous Wasserstein Barycenter Space for Generalized All-in-One Image Restoration , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , year =

  54. [54]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Transweather: Transformer-based restoration of images degraded by adverse weather conditions , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  55. [55]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    Region-based Progressive Alignment Network for Multimodal Remote Sensing Object Detection , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2026 , publisher =

    2026

  56. [56]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Residual diffusion bridge model for image restoration , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  57. [57]

    arXiv preprint arXiv:2505.21089 , year =

    Disasterm3: A remote sensing vision-language dataset for disaster damage assessment and response , author =. arXiv preprint arXiv:2505.21089 , year =

    arXiv

  58. [58]

    ISPRS Journal of Photogrammetry and Remote Sensing , volume =

    MT\_GAN: A SAR-to-optical image translation method for cloud removal , author =. ISPRS Journal of Photogrammetry and Remote Sensing , volume =. 2025 , publisher =

    2025

  59. [59]

    IEEE Transactions on Image Processing , volume =

    M2Restore: Mixture-of-experts-based mamba-cnn fusion framework for all-in-one image restoration , author =. IEEE Transactions on Image Processing , volume =. 2025 , publisher =

    2025

  60. [60]

    IEEE Transactions on Image Processing , volume =

    Image quality assessment: from error visibility to structural similarity , author =. IEEE Transactions on Image Processing , volume =. 2004 , publisher =

    2004

  61. [61]

    Proceedings of the 33rd ACM International Conference on Multimedia , pages =

    Robust Single Image Sand Removal by Leveraging Uncertainty-aware SAM Priors and Prompt Learning with Refined Perceptual Loss , author =. Proceedings of the 33rd ACM International Conference on Multimedia , pages =

  62. [62]

    IEEE Transactions on Geoscience and Remote Sensing , year =

    Cross-level interaction and intra-level fusion network for remote sensing image dehazing , author =. IEEE Transactions on Geoscience and Remote Sensing , year =

  63. [63]

    IEEE Transactions on Multimedia , year =

    All-in-one weather-degraded image restoration via adaptive degradation-aware self-prompting model , author =. IEEE Transactions on Multimedia , year =

  64. [64]

    IEEE Transactions on Multimedia , year =

    Structure-preserving frequency-regularized text-guided optimal transport for unpaired rain streaks and raindrops removal , author =. IEEE Transactions on Multimedia , year =

  65. [65]

    IEEE Transactions on Geoscience and Remote Sensing , volume =

    Encoder-free multiaxis physics-aware fusion network for remote sensing image dehazing , author =. IEEE Transactions on Geoscience and Remote Sensing , volume =. 2023 , publisher =

    2023

  66. [66]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    MP-HSIR: A multi-prompt framework for universal hyperspectral image restoration , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  67. [67]

    Advances in Neural Information Processing Systems , volume =

    Depth anything v2 , author =. Advances in Neural Information Processing Systems , volume =

  68. [68]

    IEEE Transactions on Circuits and Systems for Video Technology , volume =

    Multi-expert adaptive selection: Task-balancing for all-in-one image restoration , author =. IEEE Transactions on Circuits and Systems for Video Technology , volume =. 2024 , publisher =

    2024

  69. [69]

    Remote sensing of Environment , volume =

    Deep learning in environmental remote sensing: Achievements and challenges , author =. Remote sensing of Environment , volume =. 2020 , publisher =

    2020

  70. [70]

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

    Complexity experts are task-discriminative learners for any image restoration , author =. Proceedings of the Computer Vision and Pattern Recognition Conference , pages =

  71. [71]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Restormer: Efficient transformer for high-resolution image restoration , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  72. [72]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages =

    Multi-stage progressive image restoration , author =. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages =

  73. [73]

    arXiv preprint arXiv:2409.20197 , year =

    UIR-LoRA: achieving universal image restoration through multiple low-rank adaptation , author =. arXiv preprint arXiv:2409.20197 , year =

    Pith/arXiv arXiv

  74. [74]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Ingredient-oriented multi-degradation learning for image restoration , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  75. [75]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =

    Enhanced spatio-temporal interaction learning for video deraining: Faster and better , author =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =. 2022 , publisher =

    2022

  76. [76]

    IEEE Transactions on Image Processing , volume =

    Deep dense multi-scale network for snow removal using semantic and depth priors , author =. IEEE Transactions on Image Processing , volume =. 2021 , publisher =

    2021

  77. [77]

    IEEE Transactions on Circuits and Systems for Video Technology , volume =

    MC-Blur: A comprehensive benchmark for image deblurring , author =. IEEE Transactions on Circuits and Systems for Video Technology , volume =. 2023 , publisher =

    2023

  78. [78]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages =

    The unreasonable effectiveness of deep features as a perceptual metric , author =. Proceedings of the IEEE conference on computer vision and pattern recognition , pages =

  79. [79]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Efficient Deweather Mixture-of-Experts with Uncertainty-aware Feature-wise Linear Modulation , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =

  80. [80]

    Perceive-IR: Learning to Perceive Degradation Better for All-in-One Image Restoration , year =

    Zhang, Xu and Ma, Jiaqi and Wang, Guoli and Zhang, Qian and Zhang, Huan and Zhang, Lefei , journal =. Perceive-IR: Learning to Perceive Degradation Better for All-in-One Image Restoration , year =

Showing first 80 references.