Pith. sign in

REVIEW 4 major objections 4 minor 56 references

CLUIE: Clustering-Aware Recurrent Propagation with Local Structural Compensation for Underwater Image Enhancement

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Underwater restoration improves when recurrent states travel along content-aware cluster paths.

desk verdict Solid, novel architecture paper for UIE; the grouping control is missing, so the semantic-trajectory claim is not fully pinned down. read the letter →

arxiv 2607.21467 v1 pith:GAVSO3HJ submitted 2026-07-23 cs.CV

Kui Jiang , Zefan Feng , Laibin Chang , Yan Luo , Junjun Jiang , Xiaopeng Fan This is my paper
classification cs.CV
keywords underwaterimageenhancementcontent-adaptivetokenreorderingclustering-awarerecurrentpropagationvisualRWKVlocalstructuralcompensationdark-responsemodulationrestorationstate-spacemodels
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

Underwater images degrade unevenly: distant patches can have similar color casts while neighbors differ, so a single global correction or a fixed scan order is a poor match. The paper argues that recurrent sequence models for restoration should not use fixed spatial or spectral token orders; instead, the token order itself should be derived from the image content. It proposes CLUIE, which clusters feature tokens into semantic groups, orders the groups by inter-cluster relations, and only then lets the recurrent WKV states accumulate along that content-adaptive path, with a dark-response-modulated local branch to preserve edges. On paired benchmarks it reports best PSNR/MSE on UIEB and EUVP and competitive results on LSUI with only 4.39M parameters, suggesting the gain comes from the traversal order rather than model capacity.

What carries the argument

The central mechanism is the Clustering-aware Semantic Dynamic Reordering (CSDR) module, which converts a 2D feature map into a 1D recurrent trajectory by K-means clustering of L2-normalized tokens, inter-cluster relation-based greedy ordering, and spatially coherent intra-cluster traversal. Alongside it, Dark-response Modulated Local Propagation (DMLP) extracts depth-wise convolution local responses and gates them by a per-pixel pseudo-dark response statistic. Together they define the order in which the WKV state update operates, without changing the WKV operator itself.

What would settle it

Train CLUIE on a set of high-turbidity or extremely low-texture underwater images and compare against a fixed-scan baseline with identical capacity: if CLUIE does not outperform (or underperforms) there, the clustering's relevance to degradation would be contradicted. Alternatively, replace CSDR with a fixed but input-independent random order and observe whether the PSNR advantage vanishes.

Watch

Extended reading notes

Core claim

The paper's core claim is that the propagation path of a recurrent token-state model—not just its weights—can be learned per input, and that this matters for underwater restoration. CLUIE's CSDR module L2-normalizes tokens, runs K-means into K=8 clusters, computes a relation matrix mixing cluster-feature cosine similarity with spatial proximity, and greedily visits clusters in that relation-guided order; within each cluster, tokens are traversed in spatially coherent connected components. The authors show that replacing a raster scan with this clustered trajectory raises PSNR from 24.23 to 25.53 dB on UIEB, and that randomizing the inter-cluster order degrades results, so the ordering itself

Load-bearing premise

The method depends on K-means clustering over learned features yielding stable, semantically meaningful groups that align with degradation-relevant regions; if clusters are unstable or unrelated to restoration needs, the whole benefit collapses.

Editorial extensions

If this is right

  • If the ordering of recurrent states is as important as their aggregation, other recurrent/state-space restoration models can adopt input-dependent reordering to gain accuracy without adding parameters.
  • The 4.39M-parameter model's competitive results suggest that token-order adaptation is an efficiency lever: small models can match or beat larger fixed-scan ones.
  • The pseudo-dark response modulation provides a feature-space alternative to physical transmission estimates, opening a route to combining learned local cues with content-adaptive long-range modeling.
  • The method's relation matrix is parameter-free, so the trajectory adapts automatically at inference time; this could transfer to video or high-resolution images where fixed scans are standard.

Reading between the lines

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

  • Should the K-means groupings prove unstable in turbid or textureless scenes, a differentiable soft-clustering variant might retain the ordering benefit without the hard-assignment brittleness; the paper's own conclusion flags this risk.
  • The same reordering principle could be tested on other spatially heterogeneous degradations, such as dehazing or deraining, where the fixed-scan assumption is equally questionable.
  • A direct ablation swapping K-means for a cheaper grouping (e.g., superpixels or grid partitions with learned permutations) would tell whether the semantic clustering per se or just any adaptive order drives the gain.
  • Since the gain is attributed to order rather than capacity, comparing CLUIE against fixed-scan base models with identical parameter counts on more diverse benchmarks would sharpen the attribution.
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

4 major / 4 minor

Summary. The paper proposes CLUIE, a visual-RWKV architecture for underwater image enhancement. The main idea is to replace fixed raster/spatial/spectral RWKV scan orders with a content-adaptive trajectory: Clustering-aware Semantic Dynamic Reordering (CSDR) groups L2-normalized tokens by K-means, builds a cluster-level order from a relation matrix combining feature and spatial proximity, and traverses tokens inside each cluster by connected components to preserve locality. A second module, Dark-response Modulated Local Propagation (DMLP), compensates for locality loss by injecting depth-wise-convolution local responses modulated by a feature-space pseudo-dark response map. The model is trained on UIEB+LSUI and evaluated on UIEB, LSUI, and EUVP, reporting the best PSNR/MSE on UIEB and EUVP and competitive LSUI results, with 4.39M parameters. Ablations on UIEB support the contribution of each module.

Significance. If the central claim is correct, CLUIE is a meaningful contribution: it is a compact, linear-complexity recurrent architecture that adapts token propagation to content-dependent degradation patterns, an underexplored direction in visual RWKV for restoration. The paper is also commendable for releasing code, reporting model efficiency, and providing per-module ablations. CSDR introduces no learnable parameters, and the 0.25-0.70 dB gains over raster-order and random-order baselines suggest the trajectory matters. However, the evidence that the benefit comes specifically from "semantic" clustering rather than from any content-dependent permutation is incomplete, and the headline UIEB result is affected by hyperparameter selection on the same benchmark. These points are addressable but essential for the paper's central claim.

major comments (4)
  1. [Sec. III-B.1 / Table IV] The central claim that semantic grouping, not merely any content-adaptive reordering, drives the improvement is not established. The 'CSDR w random-sorting' row randomizes only the inter-cluster order while keeping the K-means partition; it does not control for the clustering itself. A baseline with random token assignment to K groups (or spatial blocks) under the same relation-guided traversal and intra-cluster logic is missing. Without it, the observed gain could come from any input-dependent permutation rather than from semantically meaningful clusters. Since K-means assignments are piecewise-constant in the features, the restoration loss does not directly shape cluster membership, and the visualizations in Figs. 7-9 are computed from the model's own clusters/responses rather than an independent ground-truth degradation measure. Please add this control and report repeated-seed statist
  2. [Sec. IV.A.2 / Tables V and VII] The hyperparameters K=8 and the layer-dependent 3/5/7 local windows are selected based on UIEB performance, and the same UIEB test set is then used for the headline result. This is a selection-circularity concern: the reported 25.53 dB PSNR is a selected maximum over the considered K and window settings. Please either use a held-out validation split for hyperparameter selection, or report the UIEB performance for all K/window settings with a clear protocol. At minimum, state this limitation and provide confidence intervals or repeated-seed means for the selected configuration.
  3. [Tables I-III / Sec. IV] All quantitative results are single-run, with no error bars or significance testing. Several comparisons are close: on UIEB, CLUIE's SSIM is 0.921 versus MaIR's 0.923; on LSUI, CLUIE's PSNR is 29.60 versus MaIR's 29.76. Without multiple seeds or statistical tests, the claim of state-of-the-art performance in Table I is not robustly supported. The authors should report mean +/- std over at least three seeds for the main tables, or justify why the differences are meaningful.
  4. [Eq. (5) / Sec. III-B.2] The relation-balance parameter alpha in Eq. (5) is a free hyperparameter that directly controls the cluster trajectory, but its value is never reported and no sensitivity analysis is given. Since the whole method depends on this trade-off between feature affinity and spatial proximity, a small ablation over alpha (or a statement of the chosen value and its stability) is needed.
minor comments (4)
  1. [Abstract / Title] The abstract names the model 'CRWKV' while the title, rest of the paper, and Fig. 3 use 'CLUIE'. Please reconcile this naming inconsistency.
  2. [Sec. IV.A.1 / Fig. 3] The encoder-decoder in Fig. 3 has CRWKV blocks at four resolutions, but it is not stated whether K=8 is used at every scale or whether clustering is performed independently per block. Please clarify.
  3. [Fig. 6] The diagnostic measures 'Chromaticity L2' and 'long-range distance' are used to support claims about CSDR/DMLP but are not formally defined in the text. Please define them or provide formulas.
  4. [Tables III/IV] The 'Baseline' and 'raster order' variants need precise definitions: which Q-shift variant, number of blocks, and which stages use the proposed CRWKV block. This is important for reproducibility.

Circularity Check

2 steps flagged · score 4.0 of 10

Partial circularity from UIEB test-set hyperparameter selection; central trajectory mechanism retains independent support.

  1. fitted input called prediction [Section IV-C.2 (Table V), Section IV-A.2 (Implementation Details), Section IV-B.1 (Table I)]
    "Table V analyzes the grouping granularity. The best performance is obtained when K=8. ... Unless otherwise specified, the number of semantic groups in CSDR is set to K=8. ... CLUIE (Ours) - 25.53 0.921 67.87"

    K is selected as the argmax of UIEB PSNR over K in {6,8,10} (25.22, 25.53, 25.18 dB), and the same UIEB score is then reported as the headline result in Table I. The UIEB figure is therefore the best of the searched hyperparameter values on that benchmark, not an independent test of the configuration. The gain of K=8 over K=6 and K=10 on UIEB is forced by the selection rule. EUVP and LSUI provide some independent support, but the UIEB claim is partly self-selected.

  2. fitted input called prediction [Section IV-C.3 (Table VII), Section IV-A.2 (Implementation Details)]
    "Table VII further shows that the layer-based 3/5/7 window outperforms fixed 3×3 and 5×5 windows ... the depth-wise local propagation windows are set to 3×3, 5×5, and 7×7 according to layer depth."

    The layer-dependent window sizes are selected on UIEB because they give the highest PSNR in Table VII (24.95 dB vs 24.61 and 24.68 dB for fixed windows), and the final CLUIE result on UIEB is reported with this selected configuration. The observed superiority of 3/5/7 over fixed windows on UIEB is the direct result of choosing the argmax on that dataset; it does not independently validate the window schedule. As with K, EUVP results provide an external check, but the UIEB component is selection-inflated.

full rationale

The core mechanism is not circular by construction: CSDR is an explicit algorithm (Eqs. 2-9) that builds a token order from K-means on L2-normalized features and feeds the reordered sequence into the standard WKV recurrence; DMLP is a convolutional modulation branch (Eqs. 10-15). Neither module's output is defined in terms of the metrics it is claimed to improve, and the ablations (Tables III, IV, VI) compare concrete architectural choices rather than renaming a fitted quantity. No load-bearing self-citation chain is present; the RWKV and Q-shift components are cited to external work ([14], [15], [37]). The genuine circularity-like element is the selection of hyperparameters (K and the local window sizes) on the UIEB evaluation set, after which the same UIEB scores are presented as headline achievements. This makes the UIEB portion partly self-confirming. The paper itself concedes in the conclusion that hard clustering 'may become less stable in extremely low-texture or highly turbid scenes,' and the absence of a random-partition control weakens the causal attribution of the gain to semantic grouping, but those are evidence and experimental-design limitations rather than definitional circularity. Because EUVP (30.74 dB PSNR) and LSUI results, plus the random-sorting ablation, provide non-forced evidence, the central claim retains independent content; score 4 reflects the partial UIEB selection circularity rather than a full collapse of the derivation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The paper's central contribution is an empirical architecture built from trained network weights and a small set of hand-set hyperparameters (K, alpha, windows, s_d). No new physical or mathematical entities are introduced; the only invented construct is the pseudo-dark response modulation cue, which lacks independent evidence.

free parameters (4)
  • Number of semantic clusters K = 8
    Selected via ablation on UIEB (Table V); the final UIEB result is reported from this selection.
  • Relation balance alpha in Eq. (5) = not reported
    Balances feature affinity and spatial proximity in cluster ordering; the value is unspecified and affects the constructed trajectory.
  • Pseudo-dark response window s_d = 3
    Fixed to 3x3 in implementation and not ablated, but governs the spatial extent of dark-response modulation.
  • Local propagation windows s_l = 3/5/7
    Layer-based window schedule selected over fixed 3x3/5x5 by UIEB ablation (Table VII).
assumptions (5)
  • domain assumption K-means on L2-normalized features yields stable, semantically meaningful clusters that reflect degradation-related regions.
    Section III-B.1 (Eq. 3); the CSDR mechanism assumes cluster assignments are meaningful. The conclusion admits instability in low-texture/turbid scenes.
  • domain assumption The order of token serialization determines which tokens interact earlier and more directly in WKV aggregation.
    Section III-B opening; central mechanism of CSDR, supported only by ablations, not by formal proof.
  • domain assumption Spatially distant but degradation-similar regions should share recurrent state interactions in underwater image enhancement.
    Introduction and Fig. 1 motivate content-adaptive ordering; if false, the CSDR benefit would be incidental.
  • ad hoc to paper The pseudo-dark response (channel-minimum averaged statistic) is a valid spatial modulation cue for local structural injection.
    Section III-C.2 (Eqs. 12-14); an invented feature-space statistic with no independent physical grounding, and the authors acknowledge interpretability is limited.
  • ad hoc to paper Depth-wise convolution windows 3/5/7 by layer depth are an appropriate local-context schedule.
    Section III-C.1 (Eq. 11); selected via UIEB ablations (Table VII).
invented entities (1)
  • Pseudo-dark response map A
    purpose: Modulates the injection strength of local structural responses in DMLP before recurrent aggregation.
    Defined in Eqs. (12)-(14) as a feature-space statistic, not a physical transmission estimate. No external falsifiable handle is provided, and its interpretability is acknowledged as limited.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CLUIE: Clustering-Aware Recurrent Propagation with Local Structural Compensation for Underwater Image Enhancement." pith.science (2026). https://pith.science/paper/GAVSO3HJ

@misc{pith2026260721467,
  author       = {Pith},
  title        = {Pith review of: CLUIE: Clustering-Aware Recurrent Propagation with Local Structural Compensation for Underwater Image Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GAVSO3HJ}},
  note         = {Machine review of arXiv:2607.21467}
}
read the original abstract

Underwater image enhancement remains challenging due to wavelength-dependent light absorption, scattering, and backscattering, which jointly cause color distortion, contrast degradation, and detail loss. Since these degradations vary with scene depth and imaging conditions, different regions within the same image often exhibit heterogeneous degradation patterns and thus require region-adaptive restoration. Although visual RWKV models offer an efficient linear-complexity solution for long-range dependency modeling, their predefined scanning orders are content-agnostic and therefore fail to adapt recurrent state propagation to spatially non-uniform restoration demands. To address this limitation, we propose a Clustering-aware RWKV framework, termed CRWKV, which reformulates the fixed recurrent propagation path of conventional RWKV into a content-adaptive token trajectory. Specifically, we introduce Clustering-aware Semantic Dynamic Reordering (CSDR), which groups tokens according to semantic feature similarity and derives a dynamic traversal order from inter-cluster contextual relations. This design enables WKV states to be accumulated along semantically correlated regions rather than fixed spatial or spectral orders. Since dynamic reordering may disrupt the local continuity of original spatial neighborhoods, we further propose Dark-response Modulated Local Propagation (DMLP), which extracts local structural responses via depth-wise convolution and adaptively modulates their propagation strength using a neighborhood-aware pseudo-dark response map. In this way, local structural cues are compensated before recurrent aggregation while preserving content-adaptive long-range modeling. Extensive experiments on multiple underwater image enhancement benchmarks demonstrate that CRWKV achieves state-of-the-art quantitative performance and superior visual quality.

Figures

Figures reproduced from arXiv: 2607.21467 by the authors.

Figure 1
Figure 1. Illustration of spatially non-uniform underwater degradation based [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparison of sequence organization strategies in visual RWKV. (a) RWKV-IR builds dependencies through horizontal and vertical spatial scans. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overall architecture of the proposed CLUIE framework for underwater image enhancement. The model adopts an encoder-decoder structure built [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Full-reference visual comparison on UIEB, LSUI, and EUVP. CLUIE produces color-consistent results with relatively clear local structures under [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: No-reference visual comparison on real-world underwater images. CLUIE provides balanced enhancement under different illumination and color [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Visual and statistical comparison of CSDR and DMLP on UIEB using [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 8
Figure 8. Figure 8: Feature-level visualization of clustering effectiveness. The target [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Statistical analysis of feature-response alignment. Left: restoration [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 11
Figure 11. Figure 11: Qualitative ablation of DMLP designs. Combining pseudo-dark [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 10
Figure 10. Figure 10: Feature visualization of different DMLP designs. Our CLUIE [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 5 linked inside Pith

  1. [1]

    Under- water image enhancement via minimal color loss and locally adaptive contrast enhancement,

    W. Zhang, P. Zhuang, H.-H. Sun, G. Li, S. Kwong, and C. Li, “Under- water image enhancement via minimal color loss and locally adaptive contrast enhancement,”IEEE Transactions on Image Processing, vol. 31, pp. 3997–4010, 2022

    2022

  2. [2]

    Underwater image enhancement via medium transmission-guided multi-color space embedding,

    C. Li, S. Anwar, J. Hou, R. Cong, C. Guo, and W. Ren, “Underwater image enhancement via medium transmission-guided multi-color space embedding,”IEEE Transactions on Image Processing, vol. 30, pp. 4985– 5000, 2021

    2021

  3. [3]

    Underwater image enhancement by dehazing with minimum information loss and histogram distribution prior,

    C.-Y . Li, J.-C. Guo, R.-M. Cong, Y .-W. Pang, and B. Wang, “Underwater image enhancement by dehazing with minimum information loss and histogram distribution prior,”IEEE Transactions on Image Processing, vol. 25, no. 12, pp. 5664–5677, 2016

    2016

  4. [4]

    Sea-thru: A method for removing water from underwater images,

    D. Akkaynak and T. Treibitz, “Sea-thru: A method for removing water from underwater images,” in2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 1682–1691

    2019

  5. [5]

    Underwater scene prior inspired deep underwater image and video enhancement,

    C. Li, S. Anwar, and F. Porikli, “Underwater scene prior inspired deep underwater image and video enhancement,”Pattern recognition, vol. 98, p. 107038, 2020

    2020

  6. [6]

    Domain adaptation for underwater image enhancement,

    Z. Wang, L. Shen, M. Xu, M. Yu, K. Wang, and Y . Lin, “Domain adaptation for underwater image enhancement,”IEEE Transactions on Image Processing, vol. 32, pp. 1442–1457, 2023

    2023

  7. [7]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems, vol. 30, 2017

    2017

  8. [8]

    Swinir: Image restoration using swin transformer,

    J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte, “Swinir: Image restoration using swin transformer,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1833–1844

    2021

Show all 56 references
  1. [9]

    Uformer: A general u-shaped transformer for image restoration,

    Z. Wang, X. Cun, J. Bao, W. Zhou, J. Liu, and H. Li, “Uformer: A general u-shaped transformer for image restoration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 17 683–17 693

    2022

  2. [10]

    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

  3. [11]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,”arXiv preprint arXiv:2312.00752, 2023

    2023 arXiv

  4. [12]

    Vmamba: Visual state space model,

    Y . Liu, Y . Tian, Y . Zhao, H. Yu, L. Xie, Y . Wang, Q. Ye, J. Jiao, and Y . Liu, “Vmamba: Visual state space model,”Advances in Neural Information Processing Systems, vol. 37, pp. 103 031–103 063, 2024

    2024

  5. [13]

    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,” in European Conference on Computer Vision (ECCV), 2024, pp. 222–241

    2024

  6. [14]

    Rwkv: Reinventing rnns for the transformer era,

    B. Peng, E. Alcaide, Q. Anthony, A. Albalak, S. Arcadinho, S. Bi- derman, H. Cao, X. Cheng, M. Chung, L. Derczynskiet al., “Rwkv: Reinventing rnns for the transformer era,” inFindings of the association for computational linguistics: EMNLP 2023, 2023, pp. 14 048–14 077

    2023

  7. [15]

    Exploring real&synthetic dataset and linear attention in image restoration,

    Y . Du, T. Hu, J. Zhang, R. Y . C. Xu, X. Hu, K. Wu, D. Luo, Y . Wang, and L. Ma, “Exploring real&synthetic dataset and linear attention in image restoration,”arXiv preprint arXiv:2412.03814, 2024

    2024 arXiv

  8. [16]

    Fourier-rwkv: A multi- state perception network for efficient image dehazing,

    L. Zheng, Y . Li, R. Yu, and K. Zhang, “Fourier-rwkv: A multi- state perception network for efficient image dehazing,”arXiv preprint arXiv:2512.08161, 2025

    2025

  9. [17]

    A revised underwater image formation model,

    D. Akkaynak and T. Treibitz, “A revised underwater image formation model,” in2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 6723–6732

    2018

  10. [18]

    Single image haze removal using dark channel prior,

    K. He, J. Sun, and X. Tang, “Single image haze removal using dark channel prior,”IEEE transactions on pattern analysis and machine intelligence, vol. 33, no. 12, pp. 2341–2353, 2010

    2010

  11. [19]

    Underwater image enhancement by wavelength compensation and dehazing,

    J. Y . Chiang and Y .-C. Chen, “Underwater image enhancement by wavelength compensation and dehazing,”IEEE Transactions on Image Processing, vol. 21, no. 4, pp. 1756–1769, 2012

    2012

  12. [20]

    Underwater image restoration via depth map and illumination estimation based on a single image,

    J. Zhou, T. Yang, W. Ren, D. Zhang, and W. Zhang, “Underwater image restoration via depth map and illumination estimation based on a single image,”Optics Express, vol. 29, no. 19, pp. 29 864–29 886, 2021

    2021

  13. [21]

    Enhancing underwa- ter images and videos by fusion,

    C. Ancuti, C. O. Ancuti, T. Haber, and P. Bekaert, “Enhancing underwa- ter images and videos by fusion,” in2012 IEEE conference on computer vision and pattern recognition, 2012, pp. 81–88

    2012

  14. [22]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  15. [23]

    Multi-stage progressive image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M.-H. Yang, and L. Shao, “Multi-stage progressive image restoration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 14 821–14 831

    2021

  16. [24]

    Simple baselines for image restoration,

    L. Chen, X. Chu, X. Zhang, and J. Sun, “Simple baselines for image restoration,” inEuropean Conference on Computer Vision (ECCV), 2022, pp. 17–33

    2022

  17. [25]

    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 Processing, vol. 29, pp. 4376–4389, 2019

    2019

  18. [26]

    Ugif- net: An efficient fully guided information flow network for underwater image enhancement,

    J. Zhou, B. Li, D. Zhang, J. Yuan, W. Zhang, Z. Cai, and J. Shi, “Ugif- net: An efficient fully guided information flow network for underwater image enhancement,”IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–17, 2023

    2023

  19. [27]

    Uialn: En- hancement for underwater image with artificial light,

    M. Li, K. Wang, L. Shen, Y . Lin, Z. Wang, and Q. Zhao, “Uialn: En- hancement for underwater image with artificial light,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 8, pp. 3622– 3637, 2023

    2023

  20. [28]

    Underwater image co-enhancement with correlation feature matching and joint learning,

    Q. Qi, Y . Zhang, F. Tian, Q. J. Wu, K. Li, X. Luan, and D. Song, “Underwater image co-enhancement with correlation feature matching and joint learning,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 3, pp. 1133–1147, 2021

    2021

  21. [29]

    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

  22. [30]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 10 012–10 022

    2021

  23. [31]

    Pre-trained image processing transformer,

    H. Chen, Y . Wang, T. Guo, C. Xu, Y . Deng, Z. Liu, S. Ma, C. Xu, C. Xu, and W. Gao, “Pre-trained image processing transformer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 12 299–12 310

    2021

  24. [32]

    Water- mamba: Visual state space model for underwater image enhancement,

    M. Guan, H. Xu, G. Jiang, M. Yu, Y . Chen, T. Luo, and Y . Song, “Water- mamba: Visual state space model for underwater image enhancement,” arXiv preprint arXiv:2405.08419, 2024

    2024 arXiv

  25. [33]

    Mamba-uie: Enhancing underwater images with physical model constraint,

    S. Zhang, Y . Duan, D. Li, and R. Zhao, “Mamba-uie: Enhancing underwater images with physical model constraint,”arXiv preprint arXiv:2407.19248, 2024

    2024 arXiv

  26. [34]

    O-mamba: O- shape state-space model for underwater image enhancement,

    C. Dong, C. Zhao, W. Cai, B. Yang, and Y . Guo, “O-mamba: O- shape state-space model for underwater image enhancement,” inChinese Conference on Pattern Recognition and Computer Vision (PRCV), 2025, pp. 168–182

    2025

  27. [35]

    Mair: A locality-and continuity-preserving mamba for image restoration,

    B. Li, H. Zhao, W. Wang, P. Hu, Y . Gou, and X. Peng, “Mair: A locality-and continuity-preserving mamba for image restoration,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 7491–7501

    2025

  28. [36]

    Efficient visual state space model for image deblurring,

    L. Kong, J. Dong, J. Tang, M.-H. Yang, and J. Pan, “Efficient visual state space model for image deblurring,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 12 710–12 719

    2025

  29. [37]

    Vision-rwkv: Efficient and scalable visual perception with rwkv-like architectures,

    Y . Duan, W. Wang, Z. Chen, X. Zhu, L. Lu, T. Lu, Y . Qiao, H. Li, J. Dai, and W. Wang, “Vision-rwkv: Efficient and scalable visual perception with rwkv-like architectures,” inInternational Conference on Learning Representations, 2025, pp. 83 166–83 182

    2025

  30. [38]

    Restore- rwkv: Efficient and effective medical image restoration with rwkv,

    Z. Yang, J. Li, H. Zhang, D. Zhao, B. Wei, and Y . Xu, “Restore- rwkv: Efficient and effective medical image restoration with rwkv,”IEEE Journal of Biomedical and Health Informatics, 2025

    2025

  31. [39]

    Urwkv: Unified rwkv model with multi-state perspective for low-light image restoration,

    R. Xu, Y . Niu, Y . Li, H. Xu, W. Liu, and Y . Chen, “Urwkv: Unified rwkv model with multi-state perspective for low-light image restoration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 21 267–21 276. IEEE TRANSACTIONS ON IMA...

    2025

  32. [40]

    U-shape transformer for underwater image enhancement,

    L. Peng, C. Zhu, and L. Bian, “U-shape transformer for underwater image enhancement,”IEEE Transactions on Image Processing, vol. 32, pp. 3066–3079, 2023

    2023

  33. [41]

    A generalized physical-knowledge-guided dynamic model for underwater image en- hancement,

    P. Mu, H. Xu, Z. Liu, Z. Wang, S. Chan, and C. Bai, “A generalized physical-knowledge-guided dynamic model for underwater image en- hancement,” inProceedings of the 31st ACM international conference on multimedia, 2023, pp. 7111–7120

    2023

  34. [42]

    Contrastive semi- supervised learning for underwater image restoration via reliable bank,

    S. Huang, K. Wang, H. Liu, J. Chen, and Y . Li, “Contrastive semi- supervised learning for underwater image restoration via reliable bank,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 18 145–18 155

    2023

  35. [43]

    Underwater organism color fine-tuning via decomposition and guidance,

    X. Cong, J. Gui, and J. Hou, “Underwater organism color fine-tuning via decomposition and guidance,” inProceedings of the AAAI conference on artificial intelligence, vol. 38, no. 2, 2024, pp. 1389–1398

    2024

  36. [44]

    Wavelet-based fourier infor- mation interaction with frequency diffusion adjustment for underwater image restoration,

    C. Zhao, W. Cai, C. Dong, and C. Hu, “Wavelet-based fourier infor- mation interaction with frequency diffusion adjustment for underwater image restoration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 8281–8291

    2024

  37. [45]

    Hclr-net: Hybrid contrastive learning regularization with locally randomized perturbation for underwater image enhancement,

    J. Zhou, J. Sun, C. Li, Q. Jiang, M. Zhou, K.-M. Lam, W. Zhang, and X. Fu, “Hclr-net: Hybrid contrastive learning regularization with locally randomized perturbation for underwater image enhancement,” International Journal of Computer Vision, vol. 132, no. 10, pp. 4132– 4156, 2024

    2024

  38. [46]

    Efficient concertormer for image deblurring and beyond,

    P.-H. Kuo, J. Pan, S.-Y . Chien, and M.-H. Yang, “Efficient concertormer for image deblurring and beyond,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 14 665–14 675

    2025

  39. [47]

    Multi-prior fusion transfer plugin for adapting in-air models to underwater image enhance- ment and detection,

    J. Zhou, D. Zhang, Z. He, Q. Gai, and Q. Jiang, “Multi-prior fusion transfer plugin for adapting in-air models to underwater image enhance- ment and detection,”IEEE Transactions on Image Processing, vol. 34, pp. 7773–7785, 2025

    2025

  40. [48]

    Pyuie: A coarse- to-fine deep pyramid network for underwater image enhancement,

    W. Jiang, Y . Tan, Z. Qiu, Z. Wang, Y . Yu, and Q. Jiang, “Pyuie: A coarse- to-fine deep pyramid network for underwater image enhancement,”IEEE Transactions on Multimedia, vol. 28, pp. 3054–3067, 2026

    2026

  41. [49]

    Color correction meets cross- spectral refinement: A distribution-aware diffusion for underwater image restoration,

    L. Chang, Y . Wang, B. Du, and C. Xu, “Color correction meets cross- spectral refinement: A distribution-aware diffusion for underwater image restoration,”IEEE Transactions on Multimedia, vol. 28, pp. 3792–3806, 2026

    2026

  42. [50]

    Downstream task- inspired underwater image enhancement: A perception-aware study from dataset construction to network design,

    B. Lin, F. Gao, Y . Yu, J. Dong, and Q. Du, “Downstream task- inspired underwater image enhancement: A perception-aware study from dataset construction to network design,”IEEE Transactions on Image Processing, vol. 35, pp. 2873–2888, 2026

    2026

  43. [51]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  44. [52]

    An underwater color image quality evaluation metric,

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

    2015

  45. [53]

    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 Engi- neering, vol. 41, no. 3, pp. 541–551, 2015

    2015

  46. [54]

    Musiq: Multi-scale image quality transformer,

    J. Ke, Q. Wang, Y . Wang, P. Milanfar, and F. Yang, “Musiq: Multi-scale image quality transformer,” in2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 5128–5137

    2021

  47. [55]

    Nima: Neural image assessment,

    H. Talebi and P. Milanfar, “Nima: Neural image assessment,”IEEE Transactions on Image Processing, vol. 27, no. 8, pp. 3998–4011, 2018

    2018

  48. [56]

    Devil is in the uniformity: Exploring diverse learners within transformer for image restoration,

    S. Zhou, D. Li, J. Pan, J. Zhou, J. Shi, and J. Yang, “Devil is in the uniformity: Exploring diverse learners within transformer for image restoration,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 12 307–12 317

    2025

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.