Pith. sign in

REVIEW 3 major objections 6 minor 53 references

From Controlled Scenarios to Real-World: Cross-Domain Degradation Pattern Matching for All-in-One Image Restoration

T0 review · 3 major / 6 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 UDAIR claims that a single 11.24-million-parameter model, using a degradation codebook and correlation-alignment test-time adaptation, outperforms existing all-in-one image restoration methods on all five tasks across ten source and…

desk verdict Real cross-domain gains on most tasks, but the 'all metrics SOTA' claim is false on the paper's own tables and the CORAL-based TTA step is under-specified to the point of being non-runnable as written. read the letter →

arxiv 2505.22284 v1 pith:TGU67Z6P submitted 2025-05-28 cs.CV

Junyu Fan , Chuanlin Liao , Yi Lin This is my paper
classification cs.CV
keywords all-in-oneimagerestorationdomainadaptationtest-timedegradationcodebookcontrastivelearningcorrelationalignmentreal-worldgeneralization
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

All-in-one image restoration promises one model that removes any degradation, but existing methods trained on controlled datasets lose accuracy on real-world images. This paper claims that the gap can be closed by representing each degradation type as a discrete prototype in a codebook, learning those prototypes with cross-sample contrastive learning, and then adapting each test image to the source prototype space at inference time. If the claim holds, a single 11.24-million-parameter model can outperform existing all-in-one systems on five restoration tasks across ten datasets, including no-reference real-world benchmarks. The paper also claims the adapted features form tighter, well-separated degradation clusters, which supports the degradation-identification story.

What carries the argument

The central object is the degradation codebook inside the DAAM: an embedding matrix of size $[K,D]$ that maps each continuous feature vector $z_e$ to its nearest discrete code $e_{k^*}$, turning unknown degradations into matched prototypes in a shared cross-domain latent space. It is trained jointly with the MAE reconstruction loss and the cross-sample contrastive loss, and at test time the same codebook serves as the anchor set that the CORAL-based test-time adaptation pulls target features toward. The gating mechanism after quantization selects the most representative channels, and the Domain Adaptation Module wraps the degradation features with channel expansion and squeeze-excitation before sending them to the decoder.

What would settle it

Run the target-domain inference while logging the shape of the feature tensor feeding Eqs. 16-19: if the DAM returns a single 96-dimensional vector per test image, the target covariance $C_t$ is either undefined or rank-deficient, and the reported gains cannot come from the stated CORAL computation. The claim would also be testable by replacing the source 'anchor' of Eq. 16 with a fixed random covariance matrix; if the metric gains persist, the mechanism is not the one described.

Watch

Extended reading notes

Core claim

UDAIR treats cross-domain generalization as a degradation-pattern matching problem. A Degradation Aware and Analysis Module uses a codebook of $K$ discrete embeddings to quantize continuous degradation features into prototypes, while a cross-sample contrastive loss pulls together features from different images sharing the same degradation and pushes apart different degradations. During target-domain inference, a Domain Adaptation Module is activated and fine-tuned per sample with a CORAL loss that aligns the covariance of target features to the source feature statistics, pulling each target degradation toward its corresponding source cluster center. The paper reports that this combination achieves state-of-the-art performance on all five tasks on both source-domain and target-domain datasets, with the largest gains on deraining, and that the alignment reduces the KL divergence between source and target feature distributions.

Load-bearing premise

The load-bearing premise is that the test-time adaptation step can actually be computed as written: each single test image must supply enough feature vectors to form a valid covariance matrix to compare against the source cluster statistics; if only one vector comes out of the module, the stated alignment formula has no well-defined value, and the reported gains would have to come from a different mechanism.

Editorial extensions

If this is right

  • A single model can replace per-task restoration systems in real-world applications, since one weight set handles noise, haze, rain, low light, and underwater color shifts.
  • The codebook plus contrastive learning gives the model a label-free way to identify which degradation is present, which is what makes the unified restoration possible.
  • The reported target-domain gains indicate that test-time adaptation, not just a bigger training set, is a viable route to closing the synthetic-to-real gap for all-in-one restoration.
  • Removing either the codebook or the test-time adaptation step measurably worsens target-domain results, so the paper's improvement depends on both components working together.
  • The decrease in KL divergence after adaptation provides a direct feature-level check that the domain gap is being reduced, not merely that output metrics improve.

Reading between the lines

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

  • The codebook-as-prototype recipe is transferable: any task with discrete latent degradation states, such as video restoration or medical imaging, could reuse the same cross-sample contrastive learning and anchor-based alignment.
  • A natural stress test is composite degradation, which the paper itself flags as future work; the codebook would need to represent combinations such as rain-plus-haze-plus-low-light rather than single prototypes.
  • A testable extension is to freeze the restoration decoder during adaptation; if the reported metric gains persist, the alignment step alone is responsible, and if not, the two stages interact in a way the paper does not isolate.
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.

Referee Report

3 major / 6 minor

Summary. The paper proposes UDAIR, a unified all-in-one image restoration framework that combines a codebook of discrete degradation prototypes, cross-sample contrastive learning (CSCL), and a correlation-alignment-based test-time adaptation (TTA) module with a domain adaptation module (DAM). The method is evaluated on five restoration tasks (denoising, dehazing, deraining, low-light enhancement, underwater enhancement), each with a source-domain training set and a target-domain test set, and the authors claim state-of-the-art performance across all tasks and all reported metrics, along with improved degradation identification as shown by feature-cluster visualization and reduced KL divergence.

Significance. If the proposed mechanism works as described, the contribution is practically valuable: a single 11.24M-parameter model that handles five degradation types and adapts to unseen target domains at test time without labels would be a meaningful step for real-world all-in-one restoration. The codebook-based degradation prototype idea and the cross-sample contrastive learning strategy are well-motivated and are reasonable responses to the limitations of sample-wise feature learning. The paper also reports an extensive benchmark across ten datasets and includes ablations that isolate the contributions of the codebook, CSCL, and TTA. The complexity comparison (Table 5) is useful and suggests a favorable efficiency-accuracy trade-off. However, the headline claim of universal state-of-the-art performance is contradicted by the paper's own tables, and the test-time adaptation procedure is not specified precisely enough to be reproducible; these issues must be resolved before the contribution can be accepted at its claimed strength.

major comments (3)
  1. [§1 (contribution bullet), Abstract, Table 1, Table 2] The claim that UDAIR 'achieves new state-of-the-art performance in an AiOIR way for all five tasks on both 10 open-source closed and real-world scenario datasets, in terms of all the proposed metrics' is directly contradicted by the results in the paper's own tables. In Table 1, on the source-domain UIE task, DiffUIR achieves SSIM 0.840 and PSNR 21.962, while UDAIR achieves SSIM 0.833 and PSNR 21.208. In Table 2, on the target-domain LLIE task, PromptIR achieves NIQE 4.714 versus UDAIR's 4.814; on the target-domain UIE task, ROP+ achieves UIQM 3.051 and UCIQE 0.627 versus UDAIR's 2.811 and 0.626. Because the claim is universal over all metrics, these single counterexamples falsify it as written. The authors should either weaken the claim to 'state-of-the-art on most tasks/metrics' or provide a precise enumeration of which metrics are improved, together with an explanation for the exceptions.
  2. [§2.4, Eqs. (16)–(19), Algorithm 1, Fig. 1 caption] The CORAL-based test-time adaptation is not well-defined as described. The CORAL loss in Eqs. (16)–(19) requires covariance matrices of source and target feature matrices with at least two samples per domain. However, the text and Fig. 1 state that TTA fine-tunes the model 'for each sample,' and the CSCL algorithm treats each sample as one feature vector. With a single target feature vector, the covariance in Eq. (17) has n_t = 1, making the denominator zero and the covariance undefined. The authors must clarify whether the target 'feature matrix' is formed by pooling spatial vectors from the feature map, by accumulating features over multiple images in a batch, or by some other construction. Without this detail, the central adaptation mechanism cannot be reproduced or evaluated.
  3. [§3.8, Fig. 12(f), §3.6] The KL-divergence decrease reported in Fig. 12(f) is used as evidence that the domain adaptation strategy effectively aligns source and target distributions. However, the TTA objective is itself a correlation-alignment loss that directly minimizes a second-order distributional discrepancy on the same degradation features that are plotted. The observed decrease in KL divergence after adaptation is therefore partly a re-statement of the optimization objective rather than an independent confirmation of improved degradation identification. To support the claimed mechanism, the authors should report KL divergence (or another distributional measure) on features that are not directly optimized by the TTA loss, such as features from the frozen encoder or features of the restored images, or compare against a control that optimizes the same objective without the codebook.
minor comments (6)
  1. [§2.3, Eq. (4)] The definition of the global embedding E in Eq. (4) uses notation that is not fully explained; please specify the shape of x and the range of the summation variables, and check the placement of the square root.
  2. [§2.4, text after Eq. (15)] The word 'precessed' should be 'processed'.
  3. [§3.4] The TTA procedure is described as using five steps, but no learning rate, optimizer, or loss-weight schedule for the CORAL loss during TTA is given, and it is not stated whether the backbone weights are frozen while only DAM is updated; these hyperparameters are needed for reproducibility.
  4. [§3.1] The composition of the source-domain training set is given in detail, but the target-domain dataset sizes and the number of test images per task are not reported; please add them.
  5. [§3.2, Ref. [42]] The paper cites Zhou et al. for the PSNR/SSIM metrics; the canonical references for PSNR and SSIM (e.g., Wang et al., 2004) should be cited instead or in addition.
  6. [Figures 9 and 12] The t-SNE and density plots would benefit from labeled axes, legends, and a statement of which feature layer is visualized; currently the reader must infer the exact feature source.

Circularity Check

2 steps flagged · score 2.0 of 10

No construction-level circularity: the SOTA claim is anchored in external benchmarks; the only self-referential evidence is feature-alignment diagnostics that mirror the CORAL and contrastive training objectives.

  1. self definitional [Section 3.8, Fig.12(f); also contribution list in Section 1]
    "Fig.12 (f) shows the KL divergence of samples from different tasks. The divergence between the source and raw target domains is large, indicating a pronounced distribution shift. After domain adaptation, the divergence between the source and target domains decreases sharply. The substantial decrease in divergence confirms the effectiveness of the adaptation module in narrowing the source–target gap."

    The adaptation module is trained with the CORAL loss (Eqs. 16-19), which directly penalizes the covariance discrepancy between source and target degradation features. KL divergence between the same source and target feature distributions is another measure of the same distributional gap that the loss explicitly minimizes. A post-adaptation decrease is therefore a check that the optimization objective was met, not independent confirmation that the alignment strategy improves restoration. The target-domain restoration gains in Table 2 are the independent evidence; the Fig.12(f) 'effectiveness' claim is a self-referential verification of the training objective rather than a prediction.

  2. self definitional [Section 3.7.1, Fig.9]
    "The t-SNE is applied to project the degradation features from DAAM into a two-dimensions space. The resulting visualization, presented in Fig.9 (a) and (b), further demonstrates the pivotal role of CSCL in separating the different degradation categories."

    CSCL's loss (Eq. 12) is explicitly constructed to pull same-degradation features together and push different-degradation features apart. The t-SNE separation shown for the model trained with CSCL is a direct visualization of that same objective, so it cannot serve as an independent validation that the learned features identify degradation patterns. It confirms that the loss was optimized rather than testing degradation identification on unseen conditions. This evidence is not load-bearing for the restored-image quality numbers, which come from external datasets and baselines.

full rationale

The central restoration claims are tested on 10 open-source datasets against independent baselines, and no target-domain metric is used as a training loss, so the main empirical comparison is not circular by construction. The CORAL objective and the CSCL loss do not directly optimize PSNR/SSIM/NIQE/UIQM, and the reported gains in Tables 1 and 2 therefore have independent content. I found no load-bearing self-citation chain, no imported uniqueness theorem, and no ansatz smuggled in via citation; the one author self-citation ([50]) merely supports choice of UIE metrics. The two flagged steps are self-referential diagnostics: the KL-divergence decrease and t-SNE clustering largely restate the CORAL and contrastive objectives rather than independently validating the mechanism. These are minor and non-load-bearing, so the score is 2. Separately, the paper's own tables contradict the abstract's 'all proposed metrics' SOTA phrasing (e.g., UDAIR is not best on source UIE SSIM/PSNR or target LLIE NIQE), and the CORAL covariance computation is underspecified for per-sample TTA; both are correctness and reproducibility concerns, not circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The central claim rests on a fairly heavy set of unverified modeling choices. Six numeric settings are chosen by hand or left unspecified: contrastive temperature tau, codebook size K, TTA learning rate eta, TTA step count (5), loss weights alpha=1 and beta=0.2, and DAM channel expansion (2x). Five domain assumptions are load-bearing: degradations form K discrete prototypes shared across domains; covariance alignment is the right alignment target; source cluster centers stay valid anchors during TTA; random permutation within task groups yields degradation-level positives; and VQ nearest-neighbor assignment preserves restoration-relevant information. One invented entity, the degradation codebook, is postulated without external evidence. The paper releases no code, so none of these assumptions can be checked against artifacts.

free parameters (6)
  • contrastive temperature tau = not reported
    Temperature in Eq. 12 scales the contrastive loss; no value is given in Sections 2.3 or 3.4.
  • codebook size K = not reported
    Number of discrete degradation prototypes in DAAM is never specified; the codebook is [K, D] with D=96 in Section 3.4.
  • TTA learning rate eta = not reported
    Eq. 25 uses eta for the DAM update but no value or schedule is given in Section 3.4.
  • TTA steps = 5
    Section 3.4 sets TTA steps to 5 with no sensitivity analysis.
  • loss weights alpha, beta = 1, 0.2
    Section 2.5 sets MAE and CSCL weights by hand with a single sentence of justification.
  • DAM channel expansion = 2x
    Section 2.4 sets the 1x1 convolution channel expansion to twice the input dimension.
assumptions (5)
  • domain assumption Image degradations form a small discrete set of K prototypes that are shared across source and target domains
    The codebook design in Section 2.3 postulates that quantizing continuous features to K nearest-neighbor codes preserves degradation identity across domains; no external evidence supports the discreteness of degradation space.
  • domain assumption Aligning second-order statistics (covariance) of target features to source features reduces harmful domain shift
    Section 2.4 uses CORAL (Eqs. 16-19) as the TTA objective; the paper never tests whether covariance matching is the right alignment target for restoration quality.
  • domain assumption Source-domain cluster centers remain valid fixed anchors while the DAM is updated at test time
    Section 2.4 treats anchors as fixed points; if source prototypes are not representative of a given target sample, the CORAL pull could distort target features.
  • domain assumption Random permutation within same-task mini-batches produces degradation-level positive pairs, not spurious pairs
    The CSCL algorithm in Section 2.3 relies on task group labels encoded by batch ordering; if the dataloader mixes tasks, the permutation would create false pairs.
  • domain assumption VQ nearest-neighbor quantization (Eqs. 1-3) preserves information sufficient for high-quality restoration
    The paper acknowledges quantization errors and information loss in Section 2.4 but assumes they are tolerable and correctable by the DAM.
invented entities (1)
  • degradation pattern codebook (K discrete cross-domain prototypes)
    purpose: Stores shared degradation embeddings to bridge source and target domains and guide restoration
    No external evidence shows that image degradations form K discrete clusters; the codebook's existence is validated only through the model's own restoration metrics and internal t-SNE plots.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Controlled Scenarios to Real-World: Cross-Domain Degradation Pattern Matching for All-in-One Image Restoration." pith.science (2026). https://pith.science/paper/TGU67Z6P

@misc{pith2026250522284,
  author       = {Pith},
  title        = {Pith review of: From Controlled Scenarios to Real-World: Cross-Domain Degradation Pattern Matching for All-in-One Image Restoration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TGU67Z6P}},
  note         = {Machine review of arXiv:2505.22284}
}
read the original abstract

As a fundamental imaging task, All-in-One Image Restoration (AiOIR) aims to achieve image restoration caused by multiple degradation patterns via a single model with unified parameters. Although existing AiOIR approaches obtain promising performance in closed and controlled scenarios, they still suffered from considerable performance reduction in real-world scenarios since the gap of data distributions between the training samples (source domain) and real-world test samples (target domain) can lead inferior degradation awareness ability. To address this issue, a Unified Domain-Adaptive Image Restoration (UDAIR) framework is proposed to effectively achieve AiOIR by leveraging the learned knowledge from source domain to target domain. To improve the degradation identification, a codebook is designed to learn a group of discrete embeddings to denote the degradation patterns, and the cross-sample contrastive learning mechanism is further proposed to capture shared features from different samples of certain degradation. To bridge the data gap, a domain adaptation strategy is proposed to build the feature projection between the source and target domains by dynamically aligning their codebook embeddings, and a correlation alignment-based test-time adaptation mechanism is designed to fine-tune the alignment discrepancies by tightening the degradation embeddings to the corresponding cluster center in the source domain. Experimental results on 10 open-source datasets demonstrate that UDAIR achieves new state-of-the-art performance for the AiOIR task. Most importantly, the feature cluster validate the degradation identification under unknown conditions, and qualitative comparisons showcase robust generalization to real-world scenarios.

Figures

Figures reproduced from arXiv: 2505.22284 by the authors.

Figure 1
Figure 1. To bridge the gap between closed scenarios and real [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed framework. The blue pipeline represent the dynamic domain adaptation strategy activated exclusively [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Schematic diagram of Cross-Sample Contrastive Learn [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Schematic diagram of Domain Adaptation Module. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Overall comparisons on the source domain datasets. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Detail comparisons on the source domain datasets. [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Overall comparisons on the target domain datasets. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Comparisons of different variants. haze in distant scenes. Although ROP+ also removes haze effectively, it introduces additional artifacts. In the derain￾ing task, ROP+ and CAPTNet leaves rain streaks intact, and PromptIR and AdaIR still retain light streaks. Although …
Figure 9
Figure 9. Figure 9: (a) and (b) are t-SNE visualization of degradation features of w/ CSCL and w/o CSCL. [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Comparisons of different variants. discrepancies and improves both degradation pattern recog￾nition and restoration quality in real-world scenarios. Fig.7 presents qualitative results on the target domain datasets: (a) shows the results of denoising and derain￾ing tas…
Figure 11
Figure 11. Figure 11: Comparisons of different variants. dation, enabling the model to learn more separable repre￾sentations. The features constrained by CSCL guide the backbone to more effectively and precisely remove noise, haze, rain streaks, and other degradations. In addition, the t-S…
Figure 12
Figure 12. Figure 12: (a) to (e) are density plots analysis in denoising, dehazing, deraining, LLIE, and UIE tasks, respectively. (f) is KL divergence in [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 42 canonical work pages

  1. [1]

    Fast, efficient, and accu- rate neuro-imaging denoising via supervised deep learning,

    S. Chaudhary, S. Moon, and H. Lu, “Fast, efficient, and accu- rate neuro-imaging denoising via supervised deep learning,” Nature communications, vol. 13, no. 1, p. 5165, 2022

    work page 2022

  2. [2]

    Ad- vancing real-world image dehazing: Perspective, modules, and training,

    Y . Feng, L. Ma, X. Meng, F. Zhou, R. Liu, and Z. Su, “Ad- vancing real-world image dehazing: Perspective, modules, and training,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    work page 2024

  3. [3]

    Multiscale low-light image enhancement network with il- lumination constraint,

    G.-D. Fan, B. Fan, M. Gan, G.-Y . Chen, and C. P. Chen, “Multiscale low-light image enhancement network with il- lumination constraint,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 11, pp. 7403– 7417, 2022

    work page 2022

  4. [4]

    Bridg- ing the gap between haze scenarios: A unified image dehaz- ing model,

    Y . Feng, Z. Su, L. Ma, X. Li, R. Liu, and F. Zhou, “Bridg- ing the gap between haze scenarios: A unified image dehaz- ing model,”IEEE Transactions on Circuits and Systems for Video Technology, 2024

    work page 2024

  5. [5]

    A survey on all-in-one image restoration: Taxonomy, evaluation and future trends,

    J. Jiang, Z. Zuo, G. Wu, K. Jiang, and X. Liu, “A survey on all-in-one image restoration: Taxonomy, evaluation and future trends,”arXiv preprint arXiv:2410.15067, 2024

    arXiv 2024

  6. [6]

    A comprehensive review of deep learning-based real-world image restoration,

    L. Zhai, Y . Wang, S. Cui, and Y . Zhou, “A comprehensive review of deep learning-based real-world image restoration,” IEEE Access, vol. 11, pp. 21 049–21 067, 2023

    work page 2023

  7. [7]

    Vision language models in au- tonomous driving: A survey and outlook,

    X. Zhou, M. Liu, E. Yurtsever, B. L. Zagar, W. Zimmer, H. Cao, and A. C. Knoll, “Vision language models in au- tonomous driving: A survey and outlook,”IEEE Transac- tions on Intelligent Vehicles, 2024

    work page 2024

  8. [8]

    Maui: modular analytics of uas imagery for specialty crop research,

    K. Kanaley, M. J. Murdock, T. Qiu, E. Liu, S. E. Seyram, D. Starzmann, L. B. Smart, K. M. Gold, and Y . Jiang, “Maui: modular analytics of uas imagery for specialty crop research,”Plant Methods, vol. 21, no. 1, pp. 1–17, 2025

    work page 2025

Show all 53 references
  1. [9]

    Vision- based real-time marine and offshore structural health mon- itoring system using underwater robots,

    P. Jiao, X. Ye, C. Zhang, W. Li, and H. Wang, “Vision- based real-time marine and offshore structural health mon- itoring system using underwater robots,”Computer-Aided Civil and Infrastructure Engineering, vol. 39, no. 2, pp. 281– 299, 2024

    2024

  2. [10]

    Development and external valida- tion of an artificial intelligence-based method for scalable chest radiograph diagnosis: A multi-country cross-sectional study,

    Z. Liu, J. Xu, C. Yin, G. Han, Y . Che, G. Fan, X. Li, L. Xie, L. Bao, Z. Penget al., “Development and external valida- tion of an artificial intelligence-based method for scalable chest radiograph diagnosis: A multi-country cross-sectional study,”Research, vol. 7, p. 0426, 2024

    2024

  3. [11]

    Parallel driving with big models and foundation intelligence in cyber–physical–social spaces,

    X. Wang, J. Huang, Y . Tian, C. Sun, L. Yang, S. Lou, C. Lv, C. Sun, and F.-Y . Wang, “Parallel driving with big models and foundation intelligence in cyber–physical–social spaces,”Research, vol. 7, p. 0349, 2024

    2024

  4. [12]

    Spatial- frequency dual-domain feature fusion network for low-light remote sensing image enhancement,

    Z. Yao, G. Fan, J. Fan, M. Gan, and C. P. Chen, “Spatial- frequency dual-domain feature fusion network for low-light remote sensing image enhancement,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  5. [13]

    Spatial residual for underwater object detection,

    J. Zhou, Z. He, D. Zhang, S. Liu, X. Fu, and X. Li, “Spatial residual for underwater object detection,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  6. [14]

    Energy-efficient high-fidelity image reconstruction with memristor arrays for medical di- agnosis,

    H. Zhao, Z. Liu, J. Tang, B. Gao, Q. Qin, J. Li, Y . Zhou, P. Yao, Y . Xi, Y . Linet al., “Energy-efficient high-fidelity image reconstruction with memristor arrays for medical di- agnosis,”Nature Communications, vol. 14, no. 1, p. 2276, 2023

    2023

  7. [15]

    Swinir: Image restoration using swin transformer,

    J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Tim- ofte, “Swinir: Image restoration using swin transformer,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 1833–1844

    2021

  8. [16]

    Restormer: Efficient transformer for high-resolution image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M.-H. Yang, “Restormer: Efficient transformer for high-resolution image restoration,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5728–5739

    2022

  9. [17]

    Mambair: A simple baseline for image restoration with state-space model,

    H. Guo, J. Li, T. Dai, Z. Ouyang, X. Ren, and S.-T. Xia, “Mambair: A simple baseline for image restoration with state-space model,” inEuropean conference on computer vi- sion. Springer, 2024, pp. 222–241

    2024

  10. [18]

    Image restoration via frequency selection,

    Y . Cui, W. Ren, X. Cao, and A. Knoll, “Image restoration via frequency selection,”IEEE Transactions on Pattern Analy- sis and Machine Intelligence, vol. 46, no. 2, pp. 1093–1108, 2023

    2023

  11. [19]

    AdaIR: Adaptive all-in-one image restoration via fre- quency mining and modulation,

    Y . Cui, S. W. Zamir, S. Khan, A. Knoll, M. Shah, and F. S. Khan, “AdaIR: Adaptive all-in-one image restoration via fre- quency mining and modulation,” inThe Thirteenth Interna- tional Conference on Learning Representations, 2025

    2025

  12. [20]

    Prompt-based ingredient-oriented all-in-one image restora- tion,

    H. Gao, J. Yang, Y . Zhang, N. Wang, J. Yang, and D. Dang, “Prompt-based ingredient-oriented all-in-one image restora- tion,”IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  13. [21]

    Towards effective multiple-in-one image restoration: A sequential and prompt learning strategy,

    X. Kong, C. Dong, and L. Zhang, “Towards effective multiple-in-one image restoration: A sequential and prompt learning strategy,”arXiv preprint arXiv:2401.03379, 2024

    2024 arXiv

  14. [22]

    Controlling vision-language models for universal image restoration,

    Z. Luo, F. K. Gustafsson, Z. Zhao, J. Sj ¨olund, and T. B. Sch¨on, “Controlling vision-language models for universal image restoration,”arXiv preprint arXiv:2310.01018, 2023

    2023 arXiv

  15. [23]

    Real-world scene im- age enhancement with contrastive domain adaptation learn- ing,

    Y . Zhang, Y . Cai, D. Yan, and R. Lin, “Real-world scene im- age enhancement with contrastive domain adaptation learn- ing,”ACM Transactions on Multimedia Computing, Commu- nications and Applications, vol. 20, no. 12, pp. 1–23, 2024

    2024

  16. [24]

    Promp- tir: Prompting for all-in-one image restoration,

    V . Potlapalli, S. W. Zamir, S. Khan, and F. Khan, “Promp- tir: Prompting for all-in-one image restoration,” inThirty- seventh Conference on Neural Information Processing Sys- tems, 2023

    2023

  17. [25]

    All-in-one image restoration for unknown corruption,

    B. Li, X. Liu, P. Hu, Z. Wu, J. Lv, and X. Peng, “All-in-one image restoration for unknown corruption,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 17 452–17 462

    2022

  18. [26]

    Multi- modal prompt perceiver: Empower adaptiveness generaliz- ability and fidelity for all-in-one image restoration,

    Y . Ai, H. Huang, X. Zhou, J. Wang, and R. He, “Multi- modal prompt perceiver: Empower adaptiveness generaliz- ability and fidelity for all-in-one image restoration,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 25 432–25 444

    2024

  19. [27]

    Tent: Fully test-time adaptation by entropy minimization,

    D. Wang, E. Shelhamer, S. Liu, B. Olshausen, and T. Darrell, “Tent: Fully test-time adaptation by entropy minimization,” inInternational Conference on Learning Representations,

  20. [28]

    Adaptformer: Adapting vision transformers for scal- able visual recognition,

    S. Chen, C. Ge, Z. Tong, J. Wang, Y . Song, J. Wang, and P. Luo, “Adaptformer: Adapting vision transformers for scal- able visual recognition,”Advances in Neural Information Processing Systems, vol. 35, pp. 16 664–16 678, 2022

    2022

  21. [29]

    Test- time degradation adaptation for open-set image restoration,

    Y . Gou, H. Zhao, B. Li, X. Xiao, and X. Peng, “Test- time degradation adaptation for open-set image restoration,” arXiv preprint arXiv:2312.02197, 2023

    2023 arXiv

  22. [30]

    Continual test- time domain adaptation,

    Q. Wang, O. Fink, L. Van Gool, and D. Dai, “Continual test- time domain adaptation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 7201–7211

    2022

  23. [31]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gellyet al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  24. [32]

    Deep coral: Correlation alignment for deep domain adaptation,

    B. Sun and K. Saenko, “Deep coral: Correlation alignment for deep domain adaptation,” inComputer vision–ECCV 2016 workshops: Amsterdam, the Netherlands, October 8- 10 and 15-16, 2016, proceedings, part III 14. Springer, 2016, pp. 443–450

    2016

  25. [33]

    A high-quality denoising dataset for smartphone cameras,

    A. Abdelhamed, S. Lin, and M. S. Brown, “A high-quality denoising dataset for smartphone cameras,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 1692–1700

    2018

  26. [34]

    Benchmarking single-image dehazing and be- yond,

    B. Li, W. Ren, D. Fu, D. Tao, D. Feng, W. Zeng, and Z. Wang, “Benchmarking single-image dehazing and be- yond,”IEEE Transactions on Image Processing, vol. 28, no. 1, pp. 492–505, 2018

    2018

  27. [35]

    Toward real-world single image deraining: A new bench- mark and beyond,

    W. Li, Q. Zhang, J. Zhang, Z. Huang, X. Tian, and D. Tao, “Toward real-world single image deraining: A new bench- mark and beyond,”arXiv preprint arXiv:2206.05514, 2022

    2022 arXiv

  28. [36]

    Deep retinex decomposition for low-light enhancement,

    C. Wei, W. Wang, W. Yang, and J. Liu, “Deep retinex decomposition for low-light enhancement,”arXiv preprint arXiv:1808.04560, 2018

    2018 arXiv

  29. [37]

    An underwater image enhancement benchmark dataset and beyond,

    C. Li, C. Guo, W. Ren, R. Cong, J. Hou, S. Kwong, and D. Tao, “An underwater image enhancement benchmark dataset and beyond,”IEEE transactions on image process- ing, vol. 29, pp. 4376–4389, 2019

    2019

  30. [38]

    Real-world noisy image denoising: A new benchmark,

    J. Xu, H. Li, Z. Liang, D. Zhang, and L. Zhang, “Real-world noisy image denoising: A new benchmark,”arXiv preprint arXiv:1804.02603, 2018

    2018 arXiv

  31. [39]

    From sky to the ground: A large-scale benchmark and simple baseline towards real rain removal,

    Y . Guo, X. Xiao, Y . Chang, S. Deng, and L. Yan, “From sky to the ground: A large-scale benchmark and simple baseline towards real rain removal,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Octo- ber 2023, pp. 12 097–12 107

    2023

  32. [40]

    Lime: Low-light image en- hancement via illumination map estimation,

    X. Guo, Y . Li, and H. Ling, “Lime: Low-light image en- hancement via illumination map estimation,”IEEE Transac- tions on image processing, vol. 26, no. 2, pp. 982–993, 2016

    2016

  33. [41]

    Simultaneous enhance- ment and super-resolution of underwater imagery for im- proved visual perception,

    M. J. Islam, P. Luo, and J. Sattar, “Simultaneous enhance- ment and super-resolution of underwater imagery for im- proved visual perception,”arXiv preprint arXiv:2002.01155, 2020

    2002 arXiv

  34. [42]

    Underwater camera: Improving visual perception via adap- tive dark pixel prior and color correction,

    J. Zhou, Q. Liu, Q. Jiang, W. Ren, K.-M. Lam, and W. Zhang, “Underwater camera: Improving visual perception via adap- tive dark pixel prior and color correction,”International Journal of Computer Vision, pp. 1–19, 2023

    2023

  35. [43]

    Referenceless pre- diction of perceptual fog density and perceptual image de- fogging,

    L. K. Choi, J. You, and A. C. Bovik, “Referenceless pre- diction of perceptual fog density and perceptual image de- fogging,”IEEE Transactions on Image Processing, vol. 24, no. 11, pp. 3888–3901, 2015

    2015

  36. [44]

    Real-world non-homogeneous haze removal by sliding self-attention wavelet network,

    Y . Feng, X. Meng, F. Zhou, W. Lin, and Z. Su, “Real-world non-homogeneous haze removal by sliding self-attention wavelet network,”IEEE Transactions on Circuits and Sys- tems for Video Technology, vol. 33, no. 10, pp. 5470–5485, 2023

    2023

  37. [45]

    Making a “completely blind

    A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a “completely blind” image quality analyzer,”IEEE Signal processing letters, vol. 20, no. 3, pp. 209–212, 2012

    2012

  38. [46]

    Bilevel fast scene adaptation for low-light image enhance- ment,

    L. Ma, D. Jin, N. An, J. Liu, X. Fan, Z. Luo, and R. Liu, “Bilevel fast scene adaptation for low-light image enhance- ment,”International Journal of Computer Vision, pp. 1–19, 2023

    2023

  39. [47]

    No-reference image quality assessment in the spatial domain,

    A. Mittal, A. K. Moorthy, and A. C. Bovik, “No-reference image quality assessment in the spatial domain,”IEEE Transactions on image processing, vol. 21, no. 12, pp. 4695– 4708, 2012

    2012

  40. [48]

    Human-visual-system- inspired underwater image quality measures,

    K. Panetta, C. Gao, and S. Agaian, “Human-visual-system- inspired underwater image quality measures,”IEEE Journal of Oceanic Engineering, vol. 41, no. 3, pp. 541–551, 2015

    2015

  41. [49]

    An underwater color image qual- ity evaluation metric,

    M. Yang and A. Sowmya, “An underwater color image qual- ity evaluation metric,”IEEE Transactions on Image Process- ing, vol. 24, no. 12, pp. 6062–6071, 2015

    2015

  42. [50]

    See through wa- ter: Heuristic modeling towards color correction for under- water image enhancement,

    J. Fan, J. Xu, J. Zhou, D. Meng, and Y . Lin, “See through wa- ter: Heuristic modeling towards color correction for under- water image enhancement,”IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  43. [51]

    Rank-one prior: Real- time scene recovery,

    J. Liu, R. W. Liu, J. Sun, and T. Zeng, “Rank-one prior: Real- time scene recovery,”IEEE Transactions on Pattern Analy- sis and Machine Intelligence, vol. 45, no. 7, pp. 8845–8860, 2022

    2022

  44. [52]

    Selective hourglass mapping for universal im- age restoration based on diffusion model,

    D. Zheng, X.-M. Wu, S. Yang, J. Zhang, J.-F. Hu, and W.- S. Zheng, “Selective hourglass mapping for universal im- age restoration based on diffusion model,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 25 445–25 455

    2024

  45. [2021]

    Available: https://openreview.net/forum?id= uXl3bZLkr3c

    [Online]. Available: https://openreview.net/forum?id= uXl3bZLkr3c

Pith tools

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