The reviewed record of science sign in
Pith

arxiv: 2606.06176 · v1 · pith:YTYFX2QC · submitted 2026-06-04 · cs.CV

RQUL-UIE: Revitalizing Quality-Unstable Labels for Underwater Image Enhancement via In-Dataset Self-Supervision

Haochen Hu , Yanrui Bin , Chih-yung Wen , Bing Wang This is my paper

Reviewed by Pith2026-06-28 02:46 UTCgrok-4.3pith:YTYFX2QCopen to challenge →

classification cs.CV
keywords underwater image enhancementself-supervised learningdiffusion modelslabel quality assessmentimage restorationFourier refinementdenoising supervision
0
0 comments X

The pith

A diffusion model scores training label quality to guide level-wise denoising and improve underwater image enhancement.

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

The paper presents a method that extracts quality scores for underwater image labels directly from a pre-trained diffusion model without any additional training. These scores are turned into noise-level indices that control a multi-step denoising process during supervised training, shielding the model from low-quality examples while still extracting value from them. A separate Fourier-based network is added to recover high-frequency details that standard losses often miss. The approach is tested on standard underwater datasets and shown to exceed current state-of-the-art restoration performance. The central idea is that label quality can be treated as an internal signal rather than an external problem to be filtered out.

Core claim

RQUL-UIE evaluates each training label's quality via semantic perception embeddings taken from a frozen diffusion model, quantizes the resulting scores into discrete noise-level indices, and uses those indices to supply level-specific denoising supervision; a Fourier refinement stage then reconstructs missing high-frequency content, allowing the model to train on the full set of quality-unstable pairs without letting poor labels dominate the learned mapping.

What carries the argument

In-dataset self-supervised quality scoring that converts diffusion embeddings into quantized noise levels for level-wise denoising supervision.

If this is right

  • Low-quality labels stop harming overall model performance during training.
  • Every available paired example contributes to learning at an appropriate noise level.
  • High-frequency detail recovery becomes an explicit, separate stage rather than an implicit side effect.
  • The same label set yields higher restoration metrics than methods that discard or reweight unstable examples.

Where Pith is reading between the lines

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

  • The same quality-scoring step could be inserted into other paired restoration tasks that suffer from label noise.
  • If the diffusion embeddings capture semantic structure, the method might extend to domains where human quality judgments are expensive.
  • Replacing the fixed diffusion model with one fine-tuned on underwater data could tighten the quality proxy further.

Load-bearing premise

Embeddings from a pre-trained diffusion model give a reliable measure of how good each underwater training label actually is.

What would settle it

A direct comparison showing that the diffusion-derived quality scores do not correlate with human perceptual ratings of the same labels, or that training with the scores produces no measurable gain over training without them.

Figures

Figures reproduced from arXiv: 2606.06176 by Bing Wang, Chih-Yung Wen, Haochen Hu, Yanrui Bin.

Figure 1
Figure 1. Figure 1: An overview of quantitative comparison among the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Label-wise diffusion step configuration. For high-quality labels (top row), the model undergoes full-step diffusion supervision. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The framework of proposed RQUL-UIE. The label quality is reassessed and quantized in a training-free manner based on pre [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of two extreme examples about semantic [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The proposed TRFDM module. The high-frequency [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Illustration of the coarse result (b) after LLSD (Sec. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visual Comparison on the enhanced results of different methods. the images ”Raw” means the raw underwater images. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Illustration of final enhanced results compared with the [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Visual Comparison of the enhanced results among different methods with respect to UIEB [ [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visual Comparison of the enhanced results among different methods with respect to LSUI [ [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Visual Comparison of the enhanced results among different methods with respect to EUVP [ [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗
read the original abstract

Underwater Image Enhancement (UIE) is essential for mitigating degradations caused by water medium. Although learning-based methods have advanced significantly, most rely on paired datasets with unstable label quality, which bottlenecks model performance. This paper proposes a diffusion-based, in-dataset self-supervised learning strategy designed to exploit the quality distribution of training labels. Specifically, we evaluate label quality via semantic perception embeddings from a pre-trained diffusion model in a training-free manner. These quality scores are subsequently quantized into noise-level indices, guiding a multi-step denoising process for level-wise supervision. This mechanism prevents low-quality labels from degrading the model while maximizing their utility during training. Furthermore, a Fourier-based refinement network is incorporated to explicitly reconstruct high-frequency components. Extensive evaluations demonstrate that our method consistently outperforms SOTA approaches in restoration quality. The code and pre-trained model will be available once accepted in link.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes RQUL-UIE, a diffusion-based in-dataset self-supervised strategy for underwater image enhancement (UIE). It extracts semantic perception embeddings from a pre-trained diffusion model in a training-free manner to score the quality of unstable paired training labels, quantizes these scores into noise-level indices, and uses them to drive level-wise denoising supervision that avoids contaminating the model with low-quality labels. A Fourier-based refinement network is added to reconstruct high-frequency components. The central claim is that this approach consistently outperforms existing SOTA methods on restoration quality.

Significance. If the embedding-based quality proxy is shown to be reliable for underwater data, the method could meaningfully advance UIE by extracting more value from existing paired datasets whose labels vary in quality, a recognized bottleneck. The training-free scoring step and explicit high-frequency refinement are conceptually attractive. However, the significance is currently limited by the absence of any reported quantitative results, ablations, or validation of the proxy against domain-specific degradations.

major comments (2)
  1. [Abstract] Abstract: the assertion that 'extensive evaluations demonstrate that our method consistently outperforms SOTA approaches in restoration quality' is unsupported by any quantitative metrics, dataset names, ablation studies, or error analysis. Without these, the central empirical claim cannot be assessed.
  2. [Method] Method description (quality scoring and quantization steps): the assumption that semantic perception embeddings from a natural-image-pretrained diffusion model provide an accurate ranking of perceptual quality for underwater labels is load-bearing. Domain shift from absorption, scattering, and color casts is not addressed, and no correlation with human rankings or established underwater metrics (e.g., UIQM) is reported. If the proxy mis-ranks labels, the noise-level assignment fails and low-quality supervision persists.
minor comments (1)
  1. [Abstract] Abstract: the statement that code and pre-trained models 'will be available once accepted in link' does not provide an actual repository or placeholder URL.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We agree that the empirical support for our claims must be strengthened with explicit quantitative results, ablations, and validation of the quality proxy. The revised manuscript will incorporate these elements while preserving the core technical contributions. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion that 'extensive evaluations demonstrate that our method consistently outperforms SOTA approaches in restoration quality' is unsupported by any quantitative metrics, dataset names, ablation studies, or error analysis. Without these, the central empirical claim cannot be assessed.

    Authors: We acknowledge the abstract claim requires explicit backing. The initial submission omitted clear presentation of the supporting experiments. In revision we will update the abstract to name the evaluation datasets (UIEB, EUVP) and report key metrics (PSNR, SSIM, UIQM) showing consistent gains over SOTA. Detailed tables, ablation studies on the scoring and Fourier modules, and error analysis will be added to the results section and supplementary material. revision: yes

  2. Referee: [Method] Method description (quality scoring and quantization steps): the assumption that semantic perception embeddings from a natural-image-pretrained diffusion model provide an accurate ranking of perceptual quality for underwater labels is load-bearing. Domain shift from absorption, scattering, and color casts is not addressed, and no correlation with human rankings or established underwater metrics (e.g., UIQM) is reported. If the proxy mis-ranks labels, the noise-level assignment fails and low-quality supervision persists.

    Authors: We agree that explicit validation of the proxy is necessary. Although semantic embeddings are intended to capture high-level content less sensitive to low-level underwater degradations, we did not report correlations in the original submission. The revision will include new experiments correlating the diffusion-based quality scores against UIQM values and human preference rankings on underwater images, directly addressing domain-shift concerns and demonstrating the proxy's reliability for label quantization. revision: yes

Circularity Check

0 steps flagged

No circularity: external pre-trained proxy and training-free scoring keep derivation independent

full rationale

The paper's central mechanism extracts semantic embeddings from a pre-trained diffusion model (explicitly training-free) and quantizes them to drive level-wise denoising supervision on unstable labels. No equations, fitted parameters, or self-citations are shown that reduce the claimed outperformance to a quantity computed from the target training data itself. The quality proxy is described as external to the model being trained, satisfying the condition for an independent benchmark. No self-definitional, fitted-input, or uniqueness-imported patterns appear in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The quantization of quality scores into noise levels is an implicit modeling choice whose details are not supplied.

pith-pipeline@v0.9.1-grok · 5693 in / 1111 out tokens · 37888 ms · 2026-06-28T02:46:04.716834+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 4 canonical work pages · 3 internal anchors

  1. [1]

    A revised underwater image formation model

    Derya Akkaynak and Tali Treibitz. A revised underwater image formation model. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 6723–6732, 2018. 3

    2018

  2. [2]

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

    Derya Akkaynak and Tali Treibitz. Sea-thru: A method for removing water from underwater images. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1682–1691, 2019. 3

    2019

  3. [3]

    Unveiling the under- water world: Clip perception model-guided underwater im- age enhancement.Pattern Recognition, 162:111395, 2025

    Jiangzhong Cao, Zekai Zeng, Xu Zhang, Huan Zhang, Chun- ling Fan, Gangyi Jiang, and Weisi Lin. Unveiling the under- water world: Clip perception model-guided underwater im- age enhancement.Pattern Recognition, 162:111395, 2025. 7

    2025

  4. [4]

    Underwater image en- hancement by wavelength compensation and dehazing.IEEE transactions on image processing, 21(4):1756–1769, 2011

    John Y Chiang and Ying-Ching Chen. Underwater image en- hancement by wavelength compensation and dehazing.IEEE transactions on image processing, 21(4):1756–1769, 2011. 2

    2011

  5. [5]

    Pugan: Physi- cal model-guided underwater image enhancement using gan with dual-discriminators.IEEE Transactions on Image Pro- cessing, 32:4472–4485, 2023

    Runmin Cong, Wenyu Yang, Wei Zhang, Chongyi Li, Chun- Le Guo, Qingming Huang, and Sam Kwong. Pugan: Physi- cal model-guided underwater image enhancement using gan with dual-discriminators.IEEE Transactions on Image Pro- cessing, 32:4472–4485, 2023. 1, 3, 7

    2023

  6. [6]

    Underwater depth estimation and image restoration based on single im- ages.IEEE computer graphics and applications, 36(2):24– 35, 2016

    Paulo LJ Drews, Erickson R Nascimento, Silvia SC Botelho, and Mario Fernando Montenegro Campos. Underwater depth estimation and image restoration based on single im- ages.IEEE computer graphics and applications, 36(2):24– 35, 2016. 1

    2016

  7. [7]

    Twice mixing: A rank learning based quality assessment ap- proach for underwater image enhancement.Signal Process- ing: Image Communication, 102:116622, 2022

    Zhenqi Fu, Xueyang Fu, Yue Huang, and Xinghao Ding. Twice mixing: A rank learning based quality assessment ap- proach for underwater image enhancement.Signal Process- ing: Image Communication, 102:116622, 2022. 1, 6

    2022

  8. [8]

    Fine-tuning image-conditional diffusion models is easier than you think

    Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan De Geus, Alexander Hermans, and Bastian Leibe. Fine-tuning image-conditional diffusion models is easier than you think. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 753–

  9. [9]

    A survey on underwater computer vision.ACM Computing Surveys, 55(13s):1–39, 2023

    Salma P Gonz ´alez-Sabbagh and Antonio Robles-Kelly. A survey on underwater computer vision.ACM Computing Surveys, 55(13s):1–39, 2023. 1, 2

    2023

  10. [10]

    Underwater ranker: Learn which is better and how to be better

    Chunle Guo, Ruiqi Wu, Xin Jin, Linghao Han, Weidong Zhang, Zhi Chai, and Chongyi Li. Underwater ranker: Learn which is better and how to be better. InProceedings of the AAAI conference on artificial intelligence, pages 702–709,

  11. [11]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 3

    2020

  12. [12]

    Underwater sequential images enhancement via diffusion and physics priors fusion.Information Fusion, page 103365,

    Haochen Hu, Yanrui Bin, Chih-yung Wen, and Bing Wang. Underwater sequential images enhancement via diffusion and physics priors fusion.Information Fusion, page 103365,

  13. [13]

    Contrastive semi-supervised learning for underwa- ter image restoration via reliable bank

    Shirui Huang, Keyan Wang, Huan Liu, Jun Chen, and Yun- song Li. Contrastive semi-supervised learning for underwa- ter image restoration via reliable bank. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18145–18155, 2023. 1, 3, 6, 7

    2023

  14. [14]

    Fast un- derwater image enhancement for improved visual percep- tion.IEEE Robotics and Automation Letters, 5(2):3227– 3234, 2020

    Md Jahidul Islam, Youya Xia, and Junaed Sattar. Fast un- derwater image enhancement for improved visual percep- tion.IEEE Robotics and Automation Letters, 5(2):3227– 3234, 2020. 1, 6, 7, 8, 14

    2020

  15. [15]

    Computer modeling and the design of opti- mal underwater imaging systems.IEEE Journal of Oceanic Engineering, 15(2):101–111, 1990

    Jules S Jaffe. Computer modeling and the design of opti- mal underwater imaging systems.IEEE Journal of Oceanic Engineering, 15(2):101–111, 1990. 2

    1990

  16. [16]

    Five a+ network: You only need 9k parameters for underwater image enhance- ment.arXiv preprint arXiv:2305.08824, 2023

    Jingxia Jiang, Tian Ye, Jinbin Bai, Sixiang Chen, Wenhao Chai, Shi Jun, Yun Liu, and Erkang Chen. Five a+ network: You only need 9k parameters for underwater image enhance- ment.arXiv preprint arXiv:2305.08824, 2023. 7

    work page arXiv 2023

  17. [17]

    Repurpos- ing diffusion-based image generators for monocular depth estimation

    Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Met- zger, Rodrigo Caye Daudt, and Konrad Schindler. Repurpos- ing diffusion-based image generators for monocular depth estimation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492– 9502, 2024. 3

    2024

  18. [18]

    Musiq: Multi-scale image quality transformer

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. InProceedings of the IEEE/CVF international conference on computer vision, pages 5148–5157, 2021. 1, 6

    2021

  19. [19]

    Rad: Region-aware diffusion models for image inpainting

    Sora Kim, Sungho Suh, and Minsik Lee. Rad: Region-aware diffusion models for image inpainting. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 2439–2448, 2025. 3

    2025

  20. [20]

    Auto-Encoding Variational Bayes

    Diederik P Kingma. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013. 4

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  21. [21]

    An underwater image enhancement benchmark dataset and beyond.IEEE transac- tions on image processing, 29:4376–4389, 2019

    Chongyi Li, Chunle Guo, Wenqi Ren, Runmin Cong, Junhui Hou, Sam Kwong, and Dacheng Tao. An underwater image enhancement benchmark dataset and beyond.IEEE transac- tions on image processing, 29:4376–4389, 2019. 1, 2, 3, 6, 7, 8, 12

    2019

  22. [22]

    Underwater image enhance- ment via medium transmission-guided multi-color space em- bedding.IEEE Transactions on Image Processing, 30:4985– 5000, 2021

    Chongyi Li, Saeed Anwar, Junhui Hou, Runmin Cong, Chunle Guo, and Wenqi Ren. Underwater image enhance- ment via medium transmission-guided multi-color space em- bedding.IEEE Transactions on Image Processing, 30:4985– 5000, 2021. 1, 2, 3, 6, 7, 8, 13

    2021

  23. [23]

    Underwater image enhancement with cascaded con- trastive learning.IEEE Transactions on Multimedia, 2024

    Yi Liu, Qiuping Jiang, Xinyi Wang, Ting Luo, and Jingchun Zhou. Underwater image enhancement with cascaded con- trastive learning.IEEE Transactions on Multimedia, 2024. 1, 3, 7

    2024

  24. [24]

    Visual-instructed degradation diffusion for all-in-one image restoration

    Wenyang Luo, Haina Qin, Zewen Chen, Libin Wang, Dan- dan Zheng, Yuming Li, Yufan Liu, Bing Li, and Weiming Hu. Visual-instructed degradation diffusion for all-in-one image restoration. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 12764–12777,

  25. [25]

    A computer model for underwater camera systems

    BL McGlamery. A computer model for underwater camera systems. InOcean Optics VI, pages 221–231. SPIE, 1980. 2

    1980

  26. [26]

    Dpf-net: Physical imaging model embedded data-driven underwater image enhancement.ISPRS Jour- nal of Photogrammetry and Remote Sensing, 228:679–693,

    Han Mei, Kunqian Li, Shuaixin Liu, Chengzhi Ma, and Qianli Jiang. Dpf-net: Physical imaging model embedded data-driven underwater image enhancement.ISPRS Jour- nal of Photogrammetry and Remote Sensing, 228:679–693,

  27. [27]

    Human-visual- system-inspired underwater image quality measures.IEEE Journal of Oceanic Engineering, 41(3):541–551, 2015

    Karen Panetta, Chen Gao, and Sos Agaian. Human-visual- system-inspired underwater image quality measures.IEEE Journal of Oceanic Engineering, 41(3):541–551, 2015. 1, 6

    2015

  28. [28]

    Adaptive dual-domain learn- ing for underwater image enhancement

    Lintao Peng and Liheng Bian. Adaptive dual-domain learn- ing for underwater image enhancement. InProceedings of the AAAI Conference on Artificial Intelligence, pages 6461– 6469, 2025. 1, 3, 7

    2025

  29. [29]

    Ce-vae: Capsule enhanced variational autoencoder for underwater image enhancement

    Rita Pucci and Niki Martinel. Ce-vae: Capsule enhanced variational autoencoder for underwater image enhancement. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2113–2123. IEEE, 2025. 1, 3, 7

    2025

  30. [30]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2, 3, 4

    2022

  31. [31]

    Denoising Diffusion Implicit Models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 6

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  32. [32]

    Uveb: A large-scale bench- mark and baseline towards real-world underwater video en- hancement

    Yaofeng Xie, Lingwei Kong, Kai Chen, Ziqiang Zheng, Xiao Yu, Zhibin Yu, and Bing Zheng. Uveb: A large-scale bench- mark and baseline towards real-world underwater video en- hancement. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22358– 22367, 2024. 1, 3

    2024

  33. [33]

    Style-Decoupled Adaptive Routing Network for Underwater Image Enhancement

    Hang Xu, Chen Long, Bing Wang, Hao Chen, and Zhen Dong. Style-decoupled adaptive routing network for underwater image enhancement.arXiv preprint arXiv:2604.12257, 2026. 3, 7

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  34. [34]

    Underwater image enhance- ment via minimal color loss and locally adaptive contrast en- hancement.IEEE Transactions on Image Processing, 31: 3997–4010, 2022

    Weidong Zhang, Peixian Zhuang, Hai-Han Sun, Guohou Li, Sam Kwong, and Chongyi Li. Underwater image enhance- ment via minimal color loss and locally adaptive contrast en- hancement.IEEE Transactions on Image Processing, 31: 3997–4010, 2022. 1, 3

    2022

  35. [35]

    Wavelet-based fourier information interaction with fre- quency diffusion adjustment for underwater image restora- tion

    Chen Zhao, Weiling Cai, Chenyu Dong, and Chengwei Hu. Wavelet-based fourier information interaction with fre- quency diffusion adjustment for underwater image restora- tion. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 8281–8291,

  36. [36]

    Jingchun Zhou, Jiaming Sun, Chongyi Li, Qiuping Jiang, Man Zhou, Kin-Man Lam, Weishi Zhang, and Xianping Fu. Hclr-net: hybrid contrastive learning regularization with locally randomized perturbation for underwater image en- hancement.International Journal of Computer Vision, 132 (10):4132–4156, 2024. 3, 7

    2024

  37. [37]

    Underwater image enhancement with hyper-laplacian re- flectance priors.IEEE Transactions on Image Processing, 31:5442–5455, 2022

    Peixian Zhuang, Jiamin Wu, Fatih Porikli, and Chongyi Li. Underwater image enhancement with hyper-laplacian re- flectance priors.IEEE Transactions on Image Processing, 31:5442–5455, 2022. 3 RQUL-UIE: Revitalizing Quality-Unstable Labels for Underwater Image Enhancement via In-Dataset Self-Supervision Supplementary Material (a) (b) (c) (d) Figure 10. Visua...

    2022