Pith. sign in

REVIEW 3 major objections 5 minor 51 references

This paper claims that conditioning underwater image enhancement on reliable, calibration-gated spatial priors—learned by a teacher and distilled to a student—surpasses existing methods on restoration quality and downstream perception.

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

2026-08-04 01:10 UTC pith:DZ6WALNP

load-bearing objection Solid integrative method, but the SOTA numbers rest on a non-standard split and no code, so I'd hold the results until the protocol is clarified. the 3 major comments →

arxiv 2608.00137 v1 pith:DZ6WALNP submitted 2026-07-31 eess.IV cs.CV

RPL-UIE: Reliable Prior Learning for Underwater Image Enhancement

Yifan Chen , Jiaming Liu , Ye Zheng , Zhe Sun , Tao Chen This is my paper
classification eess.IV cs.CV
keywords underwater image enhancementteacher-student learningspatial priorsRetinex decompositiondiffusion modelfrequency-aware calibrationimage restorationdownstream perception
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.

This paper claims that the main obstacle to high-quality underwater image enhancement is not the generator but the unreliability of priors extracted from badly degraded inputs. It proposes RPL-UIE, a two-stage teacher–student network: a teacher learns trustworthy spatial priors from paired degraded/reference images, a student estimates those priors from the degraded image alone, and two new modules—residual prior refinement diffusion and frequency-aware prior residual calibration—close the gap between student and teacher. On public benchmarks the method reports the best PSNR, SSIM, and LPIPS scores, and it also improves downstream object detection and instance segmentation. If these results hold under fair evaluation, the key insight is that conditioning an enhancement network on reliable, calibration-gated priors can prevent semantic drift without needing a reference at test time.

Core claim

The central claim is that prior reliability, not architecture alone, determines how faithfully an underwater enhancement network restores color, contrast, and texture. The paper constructs two complementary spatial priors—appearance and photometric (Retinex-based) priors—and trains a student encoder to predict them from degraded input only, with teacher-prior supervision. Residual Prior Refinement Diffusion (RPRD) refines the coarse student priors by diffusing in the residual correction space anchored by the coarse prior, and Frequency-Aware Prior Residual Calibration (FPRC) keeps stable low-frequency residual components while gating high-frequency ones. The calibrated priors modulate recons

What carries the argument

Residual Prior Refinement Diffusion (RPRD) is the core mechanism: instead of predicting the full teacher prior from the degraded image, the student's coarse prior is treated as an anchor and a diffusion model learns only the residual correction toward the teacher prior, starting from a zero-residual state and running deterministic reverse steps at inference. Frequency-Aware Prior Residual Calibration (FPRC) then decomposes the residual correction into low- and high-frequency parts, retains the low-frequency compensation, and adaptively gates high-frequency detail residuals with a learned spatial-channel reliability map. Together they produce calibrated priors that are injected into the recon

Load-bearing premise

The reported state-of-the-art results rest on the assumption that the random 8:2 train/test split with images resized to 256×256 yields a test set comparable to the splits used for the published comparison methods, so that the numbers in Table I are apples-to-apples.

What would settle it

A direct reproduction run: take each compared method, train and test it on the exact same random 8:2 split and 256×256 resize that RPL-UIE uses, and check whether the 26.85 dB / 33.06 dB PSNR gaps (especially the 5.88 dB LSUI gap) persist; if another method matches or beats it under identical data, the central SOTA claim is refuted.

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

If this is right

  • Reference-free inference: at test time only the degraded image is needed, yet the network behaves as if it had access to reference-derived priors.
  • Better fidelity and perception: the method reports the highest PSNR/SSIM/LPIPS on UIEB and LSUI among compared methods.
  • Downstream gains: enhanced images improve YOLOv9s object detection (mAP50 65.4 vs 63.0 raw) and SOLOv2 instance segmentation (mAP50 36.5 vs 34.0 raw).
  • Real-world utility: enhanced frames yield more SIFT feature matches in turbid ROV video, supporting navigation.
  • Ablations show each component contributes: both priors together outperform either alone; RPRD and FPRC each add gains.

Where Pith is reading between the lines

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

  • The gain on LSUI (5.88 dB over the next method) is so large that it may indicate the evaluation protocol matters more than the algorithm; a fair head-to-head with identical test images and preprocessing for every method would be the clearest test of the claim.
  • If the prior-learning mechanism generalizes, the same teacher–student residual diffusion recipe could be applied to other restoration tasks where paired data exist but test-time references are unavailable, such as dehazing, deraining, or low-light enhancement.
  • The paper leaves open the regime of extreme information loss (severe turbidity or near-zero visibility); in that regime the student's coarse prior may be too weak an anchor, and the diffusion refinement may hallucinate. A stress test with synthetically masked inputs would reveal the failure boundary.
  • Because FPRC gates high-frequency residuals by learned reliability, one could probe whether the gate is actually suppressing artifacts or merely acting as a soft attention; an analysis of gate statistics versus local degradation would make the mechanism interpretable.

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 / 5 minor

Summary. The paper proposes RPL-UIE, a two-stage teacher–student framework for underwater image enhancement. In Stage I, two encoders learn appearance and photometric spatial priors from paired degraded/reference images. In Stage II, student encoders estimate coarse priors from degraded inputs alone; Residual Prior Refinement Diffusion (RPRD) refines these coarse priors in a residual diffusion process, and Frequency-Aware Prior Residual Calibration (FPRC) gates high-frequency residual components. The calibrated priors condition a reconstruction network via Spatial Prior Modulation. The authors report state-of-the-art PSNR/SSIM/LPIPS on UIEB and LSUI, competitive results on EUVP subsets, gains on downstream object detection and instance segmentation, and an ROV-based real-world validation. The central claim is that the proposed prior-reliability mechanism yields consistently better restoration and downstream utility than eight published methods.

Significance. If the reported numbers are reliable, RPL-UIE would be a substantive empirical advance: the framework is well-motivated, the ablations (Tables V–VI) are internally consistent, and the design addresses a real limitation of prior-guided generative UIE. The residual-space diffusion and frequency-aware gating are interesting and potentially transferable. However, the significance is currently contingent on an evaluation protocol that the manuscript does not specify sufficiently to establish comparability with the literature.

major comments (3)
  1. [Sec. IV-A.1, Sec. IV-B.1, Table I] This is a load-bearing issue: if the protocol is not calibrated to the standard splits, the main empirical conclusion could collapse.
  2. [Sec. III-C, Eq. (12)] The diffusion update is under-specified. Eq. (12) defines Z_{τ-1} = \tilde{\mu}_\tau(Z_τ, \hat{∆P}_τ), but the posterior mean \tilde{\mu}_\tau is never given a concrete formula in terms of the predicted residual, the noise schedule, and the current state. As written, the equation is not implementable and prevents reproducibility. The reader is left to infer a DDIM-style update, but the manuscript should state it explicitly, including the coefficient for the predicted clean residual and the deterministic initialization Z_T=0. This is a key component of the method, not a cosmetic omission.
  3. [Tables III and IV] The downstream experiments are reported without error bars, multiple seeds, or statistical tests. Several differences are small: in Table III, RPL-UIE's mAP50 (65.4) is within 0.3 of GuidedHybSensUIR (65.1); in Table IV, mAP50 (36.5) is within 0.6 of several baselines. The paper should provide variance over at least three training runs for the downstream detectors/segmenters, and ideally for the restoration metrics as well. As it stands, the practical-utility claim is not quantitatively distinguished from noise.
minor comments (5)
  1. [Eq. (9)] There is a typographical error: '[∆Pb' should be '∆Pb'.
  2. [Eq. (10)] The product notation 'Q' should be the standard uppercase pi (∏), and the subscript limits should be clarified.
  3. [Table V] The text says adding either prior 'improved PSNR and LPIPS', but Baseline has SSIM 0.925 and configuration (A) has SSIM 0.922. Please report that SSIM slightly decreases for the appearance-only prior, or explain the discrepancy.
  4. [Sec. IV-C] The ROV real-world validation reports match counts on three frame pairs only. This is anecdotal; consider reporting aggregate statistics over a longer sequence or a standard feature-matching recall metric.
  5. [Sec. IV-F] The limitations paragraph is welcome, but the claims about robustness under extreme conditions are not quantified. If possible, add a small experiment or at least a qualitative failure case to support the stated limitations.

Circularity Check

0 steps flagged

No circularity: RPL-UIE is a supervised teacher-student restoration method whose losses and ablations are self-contained empirical training objectives

full rationale

The paper's claimed derivation chain is a standard supervised learning pipeline, not a formal derivation that could reduce to its own inputs. Teacher priors are computed from paired degraded/reference images (Eqs. 1-4), student priors are estimated from degraded inputs alone (Eqs. 6-7), RPRD refines coarse priors in residual space toward the teacher target (Eqs. 9-12), FPRC gates the resulting residuals (Eqs. 13-16), and the reconstruction network is trained with L1 and teacher-prior losses (Eqs. 19-21). None of these equations defines a quantity in terms of the metric it is later claimed to predict; the final PSNR/SSIM/LPIPS numbers are held-out test-set evaluations, not fitted parameters renamed as predictions. The word 'reliable' is a design goal supported by ablations, not a self-referential proof. Self-citations ([2], [3], [7], [17], [18]) are contextual background regarding computational imaging and do not carry the central claim. The Limitations section honestly states that extreme degradation 'may compromise reliable prior learning in Stage II,' which is an acknowledged scope caveat, not evidence of circularity. The largest concerns in the manuscript are external-validity and reproducibility issues—the unspecified random 8:2 split, 256x256 resizing, and the 'unified evaluation protocol' that does not state whether baselines were retrained—but these are correctness/comparability risks, not circularity. Thus no circular step can be exhibited from the paper's own equations or citation chain.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim is architectural and empirical. RPRD and FPRC introduce no physical entities but depend on several untested domain assumptions about prior reliability, Retinex decomposition, and the stability of low-frequency residuals.

free parameters (4)
  • λprior = 0.35
    Weight of the multi-step prior loss in Stage II (Eq. 21); set by hand without sensitivity analysis.
  • reverse steps T = 6
    Number of RPRD reverse refinement steps; no ablation over T is reported.
  • per-step supervision weights ω_k = not reported
    Introduced in Eq. (20) to weight the calibrated prior at each reverse step; values are not stated.
  • training iterations = 300k (S1), 400k/800k (S2)
    Stage-wise schedules chosen by hand; no convergence analysis.
axioms (5)
  • standard math DDPM forward/reverse formulation with posterior mean (Eqs. 10-12)
    Borrowed from Ho et al. [41]; the explicit posterior mean formula is not stated.
  • domain assumption Retinex decomposition provides complementary photometric cues
    Assumed in Sec. III-B; decomposition quality is never independently evaluated.
  • domain assumption Teacher priors learned from paired images are reliable supervision targets
    The teacher priors are used as ground truth for the student (Eq. 20), but their correctness is never measured.
  • ad hoc to paper Low-frequency residual components are stable, high-frequency components are unreliable
    Central to FPRC design (Eqs. 14-16); motivated heuristically, not quantitatively justified.
  • ad hoc to paper Residual prediction is easier than full-prior prediction
    Motivates RPRD's residual target (Eq. 9); supported only by the Table VI ablation on one dataset.

pith-pipeline@v1.3.0-alltime-deepseek · 16675 in / 15261 out tokens · 152213 ms · 2026-08-04T01:10:46.978366+00:00 · methodology

0 comments
read the original abstract

Underwater image enhancement (UIE) aims to recover clear images from observations affected by wavelength-dependent absorption, scattering, and spatially nonuniform degradation. Although existing generative methods can handle complex degradations, severe information loss may lead to semantic drift in the restored results. To address this issue, we propose RPL-UIE, a two-stage teacher--student framework for reliable prior learning. In the teacher stage, the network learns reliable and complementary spatial priors characterizing appearance and photometric properties from paired degraded and reference images. In the student stage, the network takes only degraded images as input and learns to emulate the teacher's prior extraction capability, thereby providing more reliable restoration guidance for the enhancement process without requiring reference images at inference. To reduce the prior-learning discrepancy between the teacher and student models, we further develop Residual Prior Refinement Diffusion (RPRD) and Frequency-Aware Prior Residual Calibration (FPRC). RPRD uses the coarse priors as anchors and progressively predicts the necessary corrections in the residual space. FPRC retains stable low-frequency residual components and selectively modulates high-frequency detail residuals, producing calibrated priors to support high-quality reconstruction. Experiments on multiple UIE benchmarks demonstrate competitive restoration performance. Downstream underwater object detection and instance segmentation experiments further demonstrate the improved utility of enhanced images for visual perception, while tests on real-world data captured by a remotely operated vehicle (ROV) support the practical applicability of RPL-UIE.

Figures

Figures reproduced from arXiv: 2608.00137 by Jiaming Liu, Tao Chen, Ye Zheng, Yifan Chen, Zhe Sun.

Figure 1
Figure 1. Figure 1: Overall framework of RPL-UIE. (a) In Stage I, the Appearance Spatial Prior Encoder (ASPE) and Photometric Spatial Prior Encoder (PSPE) learn [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Detailed architectures of Residual Prior Refinement Diffusion (RPRD) and Frequency-Aware Prior Residual Calibration (FPRC), illustrated with the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visual comparison on the UIEB dataset. rate in Stage I was 2 × 10−4 and was adjusted using cosine annealing. During Stage II, the prior-learning warm-up used an initial learning rate of 1×10−4 . During joint optimization, the reconstruction-network and prior parameter groups used initial learning rates of 2 × 10−4 and 5 × 10−5 , respectively. The learning rates were decayed every 80k iterations. The number… view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparison on the LSUI dataset. ImageNet, RPL-UIE achieved the best PSNR and SSIM values of 26.58 dB and 0.872, respectively, while ranking second in LPIPS and UCIQE with values of 0.146 and 0.598. Compared with CCL-Net, which achieved the best UCIQE, RPL-UIE improved PSNR and SSIM by 3.63 dB and 0.082 and reduced LPIPS by 0.066, with a UCIQE difference of only 0.002. On EUVP-Scenes, RPL-UIE achieve… view at source ↗
Figure 5
Figure 5. Figure 5: Visual comparison on the EUVP dataset. underwater video in Dushu Lake and enhanced the recorded frames. The scene exhibited severe water turbidity, nonuniform illumination, and viewpoint changes caused by ROV motion. Object boundaries and background structures were indistinct in the raw frames, posing substantial challenges to subsequent inter-frame matching and visual perception [PITH_FULL_IMAGE:figures/… view at source ↗
Figure 6
Figure 6. Figure 6: SIFT feature matching between adjacent frames from a real-world [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visual comparison of downstream object detection on UDD. Green, blue, and red boxes indicate correct detections, false positives, and missed objects, [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visual comparison of downstream instance segmentation on the UIIS dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Visual ablation of the main components in RPL-UIE. [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visual ablation of RPRD under different experimental configurations. [PITH_FULL_IMAGE:figures/full_fig_p012_10.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

51 extracted references

  1. [1]

    A survey on underwater computer vision,

    S. P. González-Sabbagh and A. Robles-Kelly, “A survey on underwater computer vision,”ACM Computing Surveys, vol. 55, no. 13s, pp. 1–39, 2023

    2023

  2. [2]

    Water-related optical imaging: From algorithm to hardware,

    Z. Sun and X. Li, “Water-related optical imaging: From algorithm to hardware,”Science China Technological Sciences, vol. 68, no. 1, p. 1100401, 2025

    2025

  3. [3]

    Computational ghost imaging in tur- bulent water based on self-supervised information extraction network,

    Y . Chen, Z. Sun, C. Li, and X. Li, “Computational ghost imaging in tur- bulent water based on self-supervised information extraction network,” Optics & Laser Technology, vol. 167, p. 109735, 2023

    2023

  4. [4]

    A revised underwater image formation model,

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

    2018

  5. [5]

    Underwater color disparities: Cues for enhancing underwater images toward natural color consistencies,

    H. Wang, S. Sun, and P. Ren, “Underwater color disparities: Cues for enhancing underwater images toward natural color consistencies,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 2, pp. 738–753, 2024

    2024

  6. [6]

    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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 1682–1691

    2019

  7. [7]

    Extreme-depth water-related optical imaging: Conquering ultra-low illumination environments from epipelagic zone to mariana trench,

    Z. Sun, T. Tian, H. Hu, Y . He, M. Shangguan, T. Yu, Q. Yang, M. Chen, X. Wang, Y . Chenet al., “Extreme-depth water-related optical imaging: Conquering ultra-low illumination environments from epipelagic zone to mariana trench,”PhotoniX, vol. 7, no. 1, p. 7, 2026

    2026

  8. [8]

    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

  9. [9]

    A semi-supervised domain-adaptive framework for real-world underwater image enhancement,

    J. Wen, G. Yang, B. Zhao, D. Huang, L. Lei, B. Zhang, Z. Gao, X. Chen, and B. M. Chen, “A semi-supervised domain-adaptive framework for real-world underwater image enhancement,”IEEE Transactions on Geo- science and Remote Sensing, vol. 63, p. 4209715, 2025

    2025

  10. [10]

    Color balance and fusion for underwater image enhancement,

    C. O. Ancuti, C. Ancuti, C. De Vleeschouwer, and P. Bekaert, “Color balance and fusion for underwater image enhancement,”IEEE Transac- tions on Image Processing, vol. 27, no. 1, pp. 379–393, 2018

    2018

  11. [11]

    Underwater image enhancement with hyper-laplacian reflectance priors,

    P. Zhuang, J. Wu, F. Porikli, and C. Li, “Underwater image enhancement with hyper-laplacian reflectance priors,”IEEE Transactions on Image Processing, vol. 31, pp. 5442–5455, 2022

    2022

  12. [12]

    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 Analysis and Machine Intelligence, vol. 45, no. 7, pp. 8845–8860, 2023

    2023

  13. [13]

    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

  14. [14]

    Toward sufficient spatial- frequency interaction for gradient-aware underwater image enhance- ment,

    C. Zhao, W. Cai, C. Dong, and Z. Zeng, “Toward sufficient spatial- frequency interaction for gradient-aware underwater image enhance- ment,” in2024 IEEE International Conference on Acoustics, Speech and Signal Processing. IEEE, 2024, pp. 3220–3224

    2024

  15. [15]

    DiffUIE: Learning latent global priors in diffusion models for underwater image enhancement,

    Y . Qing, S. Liu, H. Wang, and Y . Wang, “DiffUIE: Learning latent global priors in diffusion models for underwater image enhancement,” IEEE Transactions on Multimedia, vol. 27, pp. 2516–2529, 2025

    2025

  16. [16]

    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

  17. [17]

    Attention- enhanced computational ghost imaging,

    Y . Chen, T. Tian, X. Lu, C. Li, R. Zhu, Z. Sun, and X. Li, “Attention- enhanced computational ghost imaging,”Science China Information Sciences, vol. 68, no. 6, p. 162104, 2025

    2025

  18. [18]

    Large model enhanced computational ghost imaging,

    Y . Chen, H. An, Z. Sun, T. Tian, M. Chen, C. Spielmann, and X. Li, “Large model enhanced computational ghost imaging,”Science China Technological Sciences, vol. 68, no. 11, p. 2120403, 2025

    2025

  19. [19]

    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

  20. [20]

    DCD-UIE: Decoupled chromatic diffusion model for underwater image enhancement,

    G. Fan, Y . Zhou, J. Zhou, Y . Ju, G.-Y . Chen, J. Li, and A. C. Kot, “DCD-UIE: Decoupled chromatic diffusion model for underwater image enhancement,”IEEE Transactions on Image Processing, vol. 35, pp. 449–464, 2026

    2026

  21. [21]

    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

  22. [22]

    Underwater image restoration through a prior guided hybrid sense approach and extensive benchmark analysis,

    X. Guo, X. Chen, S. Wang, and C.-M. Pun, “Underwater image restoration through a prior guided hybrid sense approach and extensive benchmark analysis,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 5, pp. 4784–4800, 2025

    2025

  23. [23]

    SeaDiff: Underwater image enhancement with degradation-aware dif- fusion model,

    H. Bi, L. Chen, J. Cao, J. Wang, J. Sun, Y . Rao, and J. Dong, “SeaDiff: Underwater image enhancement with degradation-aware dif- fusion model,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 12, pp. 12 212–12 226, 2025

    2025

  24. [24]

    Learning a physical-aware diffusion model based on transformer for underwater image enhance- ment,

    C. Zhao, C. Dong, W. Cai, and Y . Wang, “Learning a physical-aware diffusion model based on transformer for underwater image enhance- ment,”IEEE Transactions on Geoscience and Remote Sensing, vol. 64, pp. 1–14, 2026

    2026

  25. [25]

    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

  26. [26]

    Automatic red- channel underwater image restoration,

    A. Galdran, D. Pardo, A. Picón, and A. Alvarez-Gila, “Automatic red- channel underwater image restoration,”Journal of Visual Communica- tion and Image Representation, vol. 26, pp. 132–145, 2015

    2015

  27. [27]

    Underwater image enhancement via piecewise color correction and dual prior optimized contrast enhancement,

    W. Zhang, S. Jin, P. Zhuang, Z. Liang, and C. Li, “Underwater image enhancement via piecewise color correction and dual prior optimized contrast enhancement,”IEEE Signal Processing Letters, vol. 30, pp. 229–233, 2023

    2023

  28. [28]

    Underwater image enhancement via weighted wavelet visual perception fusion,

    W. Zhang, L. Zhou, P. Zhuang, G. Li, X. Pan, W. Zhao, and C. Li, “Underwater image enhancement via weighted wavelet visual perception fusion,”IEEE Transactions on Circuits and Systems for Video Technol- ogy, vol. 34, no. 4, pp. 2469–2483, 2024

    2024

  29. [29]

    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

  30. [30]

    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

  31. [31]

    Underwater image enhancement with cascaded contrastive learning,

    Y . Liu, Q. Jiang, X. Wang, T. Luo, and J. Zhou, “Underwater image enhancement with cascaded contrastive learning,”IEEE Transactions on Multimedia, vol. 27, pp. 1512–1525, 2025

    2025

  32. [32]

    FDCE-Net: underwater image enhancement with embedding frequency and dual color encoder,

    Z. Cheng, G. Fan, J. Zhou, M. Gan, and C. P. Chen, “FDCE-Net: underwater image enhancement with embedding frequency and dual color encoder,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 2, pp. 1728–1744, 2025

    2025

  33. [33]

    TAFormer: A transmission- aware transformer for underwater image enhancement,

    Y . Li, Z. Mi, Y . Wang, S. Jiang, and X. Fu, “TAFormer: A transmission- aware transformer for underwater image enhancement,”IEEE Transac- tions on Circuits and Systems for Video Technology, vol. 35, no. 1, pp. 601–616, 2025

    2025

  34. [34]

    WWE-UIE: A wavelet & white balance efficient network for underwater image enhancement,

    C.-H. Cheng, J.-W. Lee, C.-M. Lee, and C.-C. Hsu, “WWE-UIE: A wavelet & white balance efficient network for underwater image enhancement,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2026, pp. 2135–2145

    2026

  35. [35]

    Underwater image enhance- ment by transformer-based diffusion model with non-uniform sampling for skip strategy,

    Y . Tang, H. Kawasaki, and T. Iwaguchi, “Underwater image enhance- ment by transformer-based diffusion model with non-uniform sampling for skip strategy,” inProceedings of the 31st ACM International Con- ference on Multimedia, 2023, pp. 5419–5427

    2023

  36. [36]

    Speed-up DDPM for real-time underwater image enhancement,

    S. Lu, F. Guan, H. Zhang, and H. Lai, “Speed-up DDPM for real-time underwater image enhancement,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 5, pp. 3576–3588, 2024

    2024

  37. [37]

    Adap- tive double-branch fusion conditional diffusion model for underwater image restoration,

    Y . Wang, K. He, Q. Qu, X. Du, T. Liu, T. Lei, and A. K. Nandi, “Adap- tive double-branch fusion conditional diffusion model for underwater image restoration,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 35, no. 10, pp. 9723–9735, 2025

    2025

  38. [38]

    UADiff: a deep underwater image enhancement network using generative diffusion prior and uncertainty-aware learning,

    Y . Ou, A. Esmaeilzehi, M. O. Ahmad, and M. Swamy, “UADiff: a deep underwater image enhancement network using generative diffusion prior and uncertainty-aware learning,”IEEE Transactions on Geoscience and Remote Sensing, vol. 63, pp. 1–14, 2025. 14

    2025

  39. [39]

    Reti-Diff: Illumination degradation image restoration with retinex- based latent diffusion model,

    C. He, C. Fang, Y . Zhang, L. Tang, J. Huang, K. Li, X. Li, S. Farsiuet al., “Reti-Diff: Illumination degradation image restoration with retinex- based latent diffusion model,” inInternational Conference on Learning Representations, 2025, pp. 43 332–43 352

    2025

  40. [40]

    Single-step latent diffusion for underwater image restoration,

    J. Wu, T. Wang, M. A. B. Siddique, M. J. Islam, C. Fermuller, Y . Aloimonos, and C. A. Metzler, “Single-step latent diffusion for underwater image restoration,”IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1–11, 2025

    2025

  41. [41]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 6840–6851

    2020

  42. [42]

    Fast underwater image enhancement for improved visual perception,

    M. J. Islam, Y . Xia, and J. Sattar, “Fast underwater image enhancement for improved visual perception,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 3227–3234, 2020

    2020

  43. [43]

    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. [44]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018, pp. 586–595

    2018

  45. [45]

    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. [46]

    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

  47. [47]

    Distinctive image features from scale-invariant keypoints,

    D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” International Journal of Computer Vision, vol. 60, no. 2, pp. 91–110, 2004

    2004

  48. [48]

    A new dataset, poisson gan and aquanet for underwater object grabbing,

    C. Liu, Z. Wang, S. Wang, T. Tang, Y . Tao, C. Yang, H. Li, X. Liu, and X. Fan, “A new dataset, poisson gan and aquanet for underwater object grabbing,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 5, pp. 2831–2844, 2021

    2021

  49. [49]

    YOLOv9: Learning what you want to learn using programmable gradient information,

    C.-Y . Wang, I.-H. Yeh, and H.-Y . Mark Liao, “YOLOv9: Learning what you want to learn using programmable gradient information,” in European Conference on Computer Vision. Springer, 2024, pp. 1–21

    2024

  50. [50]

    Watermask: Instance segmentation for underwater imagery,

    S. Lian, H. Li, R. Cong, S. Li, W. Zhang, and S. Kwong, “Watermask: Instance segmentation for underwater imagery,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 1305–1315

    2023

  51. [51]

    SOLOv2: Dynamic and fast instance segmentation,

    X. Wang, R. Zhang, T. Kong, L. Li, and C. Shen, “SOLOv2: Dynamic and fast instance segmentation,” vol. 33, pp. 17 721–17 732, 2020

    2020