A Generative Data Framework with Reliable Supervision for Underwater Image Restoration and Enhancement
Authors: Yufeng Tian, Yifan Chen, Zhe Sun, Libang Chen, Mingyu Dou, Jijun Lu, Ye Zheng, Xuelong Li
Organizations: Institute of Artificial Intelligence (TeleAI), China Telecom, Shanghai, 200232, China · Department of Automation, Shanghai Jiao Tong University, Shanghai, 200240, China · School of Artificial Intelligence, Optics and ElectroNics (iOPEN), Northwestern Polytechnical University, Xi’an, 710072, China · Guangdong Provincial Key Laboratory of Quantum Metrology and Sensing, School of Physics and Astronomy, Sun Yat-Sen University (Zhuhai Campus), Zhuhai, 519082, China · Xi’an Institute of Optics and Precision Mechanics, Chinese Academy of Sciences, Xi’an, 710119, China
Learning-based underwater image restoration and enhancement rely heavily on high-quality paired training data. However, obtaining clear reference images for real underwater scenes remains challenging. Existing datasets often use the outputs of enhancement algorithms as references, which may introduce uncertainty into supervision targets. To address this issue, we propose a generative data framework that preserves clear source images as restoration references and synthesizes corresponding underwater inputs through unpaired image-to-image translation. Using representative underwater appearances to guide the translation, we construct UWNature from natural photographs. We systematically examine the reliability of the generated supervision through its agreement with real underwater appearances, structural consistency, and depth-related degradation characteristics. Experiments across 6 restoration architectures and 3 test sets evaluate its effectiveness through objective metrics, human perceptual assessment, and the correspondence between synthetic degradation and learned restoration behavior. The results demonstrate competitive restoration performance and favorable perceptual quality on real underwater images. Further analysis reveals that greater spatial plausibility of synthetic degradation does not necessarily improve restoration performance, highlighting the importance of evaluating both the intrinsic properties of synthetic data and their effectiveness for model learning. These findings support a scalable framework for constructing and evaluating reliable paired supervision for underwater image restoration and enhancement. The generated datasets are available at: https://github.com/tianyufengcn/SynUIEDatasets.
Figures & tables
Fig. 1: Overview of the proposed generative data framework. 6 representative underwater appearances guide unpaired image translation to construct synthetic degraded inputs paired with their original clear references. The resulting supervision is examined through intrinsic data analysis, downstream restoration performance, human perceptual evaluation, and real-world validation.
Fig. 2: Appearance distributions of natural references, synthetic underwater images, and real underwater images across 5 descriptors, with Wasserstein-1 ( W1 ) distances between real and synthetic distributions indicated. The bottom-right panel shows the joint distribution of real and synthetic images in a PCA space fitted to real underwater data.
Pair structure
Depth association
Metric
Median
Attribute
Median ρ
Grayscale correlation
0.9426
Chroma shift
−0.1551
Edge correlation
0.8602
∣ΔL∣
−0.0752
Edge SSIM
0.7133
Log contrast ratio
0.0675
Log detail ratio
0.0698
TABLE I: Structure and depth statistics of UWNature.
Fig. 3: Spatial variation of synthetic degradation across relative scene depth. 3 representative cases show reference and synthetic images, depth estimates from DA2 and MoGe-2, and their corresponding depth-binned chroma shifts and log-detail ratios. The bottom panels summarize the trends across all pairs using DA2, with median values and 95% confidence intervals.
Variant
Chroma ρ
SWD ↓
Precision ↑
Coverage ↑
Edge SSIM ↑
O
−0.1551
0.2044
0.8070
0.6056
0.7382
R
0.0180
0.2074
0.7932
0.6028
0.7423
V
−0.1555
0.1857
0.8089
0.6042
0.7271
B
−0.1545
0.2004
0.7997
0.6000
0.7326
RV
0.0115
0.1885
0.7997
0.6088
0.7331
RB
0.0180
0.2034
0.7905
0.6025
0.7375
TABLE II: Intrinsic comparison of 8 modulation variants.
Fig. 4: Visual comparison of the original synthesis and depth modulation variants across 3 representative scenes. Each 2 column group shows the reference image, DA2 depth map, and eight synthesis variants: O, R, V, B, RV, RB, VB, and RVB.
Fig. 5: Qualitative comparison of restoration and enhancement results on the U45 [ 43 ] test set. Models trained on the proposed synthetic datasets (Syn1: UWImgNetSD; Syn2: UWNature; 4th, 5th, 9th, and 10th rows) show more vivid colors and visually consistent restoration than several models trained on public datasets. Columns (a)–(f) correspond to UWCNN [ 6 ] , UIEC 2 -Net [ 41 ] , WaterNet [ 9 ] , U-shape [ 7 ] , DMunderwater [ 42 ] , and WF-Diff [ 8 ] , respectively. The same conventions apply to Figs. 6 and 7 .
Fig. 6: Qualitative restoration results on the RUIE [ 44 ] test set. Models trained on the proposed synthetic datasets (4th, 5th, 9th, and 10th rows) correct underwater color casts and recover scene details across different images. The model and training-source layout follows Fig. 5 .
Fig. 7: Qualitative comparison on the LSUI&EUVP test set. Models trained on the proposed synthetic datasets (4th, 5th, 9th, and 10th rows) show clear textures and improved color appearance across different underwater scenes. The model and training-source layout follows Fig. 5 .
Model
Training Set
PAUQA ↑
MUSIQ ↑
TOPIQ-NR ↑
LIQE ↑
CLIP-IQA+ ↑
U45
RUIE
LSUI
U45
RUIE
LSUI
U45
RUIE
LSUI
U45
RUIE
LSUI
U45
RUIE
LSUI
&EUVP
&EUVP
&EUVP
&EUVP
&EUVP
UWCNN [ 6 ]
UIEB [ 9 ]
41.68
35.03
42.57
48.98
41.02
46.11
0.514
0.386
0.441
2.255
1.930
2.411
0.518
0.441
0.519
LSUI [ 7 ]
34.33
37.38
43.77
49.78
40.49
48.88
0.478
0.385
0.467
2.205
1.999
2.707
0.515
0.480
0.577
EUVP [ 15 ]
33.71
33.68
42.72
47.64
42.78
49.09
0.448
0.406
0.477
2.022
1.887
2.766
0.481
0.448
0.596
UWImgNetSD
30.87
32.47
40.43
50.82
41.89
50.27
0.473
0.414
0.478
2.372
2.150
2.809
0.507
0.461
0.599
TABLE III: No-reference image quality comparison using five learned metrics.
Model
Training Set
UIQM ↑
UCIQE ↑
U45
RUIE
LSUI
U45
RUIE
LSUI
&EUVP
&EUVP
UWCNN [ 6 ]
UIEB [ 9 ]
0.688
0.347
0.795
0.5507
0.5115
0.5732
LSUI [ 7 ]
0.646
0.746
0.980
0.5390
0.5113
0.5773
EUVP [ 15 ]
0.454
0.334
0.746
0.5369
0.5177
0.5745
UWImgNetSD
0.639
0.547
0.913
0.5612
0.5493
0.5979
TABLE IV: Underwater image quality comparison using UIQM and UCIQE.
Training Data
UIEB
LSUI
EUVP
UWNature
UWImgNetSD
MOS ↑
3.321
3.469
3.284
3.673
3.784
95% CI
[3.17,3.48]
[3.29,3.64]
[3.10,3.47]
[3.52,3.83]
[3.59,3.97]
TABLE V: Mean opinion scores (MOS) and 95% confidence intervals for different training datasets.
Fig. 8: Qualitative comparison of restoration results using the original UWNature synthesis ( O ) and its depth-aware modulation ( RV ) on U45, RUIE, and LSUI&EUVP. Each test set contains the same input image restored by models trained with O and RV . Columns (a)–(f) correspond to UWCNN, UIEC 2 -Net, WaterNet, U-shape, DMunderwater, and WF-Diff, respectively.
Fig. 9: Real-world underwater restoration examples using DMunderwater trained on UWNature. The figure presents the camera setup and restoration results for color card, coral reefs, and divers in a swimming pool.
Fig. 10: Structural consistency in 6 UWNature pairs, ordered from higher to lower edge SSIM with two pairs per row. Each pair shows the reference, synthetic input, edge overlay (white: shared; red: reference-only; cyan: synthetic-only), and local edge-SSIM map (blue–violet: lower; orange–red: higher similarity). Main object geometry persists in the upper examples, while local detail loss becomes more prominent in the lower examples.
Measure
P5
P25
Median
P75
P95
Gray correlation
0.794
0.905
0.943
0.967
0.988
Edge correlation
0.704
0.803
0.860
0.908
0.954
Edge SSIM
0.474
0.632
0.713
0.789
0.884
TABLE VI: Quantiles of structural consistency over 1,523 UWNature pairs.
Attribute
Shifted depth
Positive control
Chroma shift
0.0005
0.9875
Absolute ΔL
0.0011
0.8964
Log contrast ratio
−0.0012
−0.7209
Log detail ratio
−0.0020
−0.8635
TABLE VII: Depth association controls using DA2 relative depth, reported as median Spearman correlations.
Fig. 11: Image-level depth associations for four degradation attributes using DA2 relative-depth estimates. Boxes show the original synthesis (O), the within-image median of 50 shifted-depth comparisons (Null), and the imposed positive control (PC). The near-zero null distributions and correctly signed control correlations distinguish measured spatial trends from random correspondence.
Model
Training Set
PSNR ↑
SSIM ↑
LPIPS ↓
UWCNN [ 6 ]
UIEB [ 9 ]
19.90
0.8143
0.2760
LSUI [ 7 ]
19.56
0.8229
0.2600
EUVP [ 15 ]
21.50
0.8449
0.2418
UWImgNetSD
19.82
0.7999
0.2770
UWNature (O)
20.30
0.8175
0.2598
UWNature (RV)
20.46
0.8267
0.2766
TABLE VIII: Full-reference restoration quality on LSUI&EUVP.
Model
Training Set
UCIQE ↑
σc
CL
μs
U45
RUIE
LSUI
U45
RUIE
LSUI
U45
RUIE
LSUI
U45
RUIE
LSUI
&EUVP
&EUVP
&EUVP
&EUVP
UWCNN [ 6 ]
UIEB [ 9 ]
0.5507
0.5115
0.5732
0.2607
0.1844
0.2628
0.8026
0.7584
0.8596
0.8090
0.8425
0.8317
LSUI [ 7 ]
0.5390
0.5113
0.5773
0.2489
0.1944
0.2692
0.7843
0.7844
0.8677
0.8045
0.7957
0.8273
EUVP [ 15 ]
0.5369
0.5177
0.5745
0.2666
0.2128
0.2769
0.7380
0.7355
0.8324
0.8135
0.8394
0.8404
UWImgNetSD
0.5612
0.5493
0.5979
0.2957
0.2555
0.3053
0.7641
0.7689
0.8557
0.8269
0.8488
0.8544
TABLE IX: UCIQE and its components for the 6 restoration architectures on U45, RUIE, and LSUI&EUVP.
Model
Training Set
UIQM ↑
UICM
UISM
UIConM
U45
RUIE
LSUI
U45
RUIE
LSUI
U45
RUIE
LSUI
U45
RUIE
LSUI
&EUVP
&EUVP
&EUVP
&EUVP
UWCNN [ 6 ]
UIEB [ 9 ]
0.688
0.347
0.795
6.730
3.354
5.893
4.765
3.875
5.029
-0.2542
-0.2495
-0.2394
LSUI [ 7 ]
0.646
0.746
0.980
6.479
10.186
10.196
4.452
4.445
5.175
-0.2383
-0.2389
-0.2339
EUVP [ 15 ]
0.454
0.334
0.746
4.260
4.705
7.957
3.955
3.498
4.721
-0.2334
-0.2326
-0.2441
UWImgNetSD
0.639
0.547
0.913
5.765
9.098
9.549
4.416
3.884
4.871
-0.2315
-0.2396
-0.2222
TABLE X: UIQM and its components for the 6 restoration architectures on U45, RUIE, and LSUI&EUVP.
Target
UCIQE ↑
σc
CL
μs
a1
0.5159
0.1927
0.7429
0.8610
a2
0.5170
0.2495
0.6671
0.8429
a3
0.4895
0.1801
0.7518
0.7718
a4
0.4914
0.1955
0.7121
0.7938
a5
0.5465
0.2602
0.7641
0.8348
a6
0.5260
0.1890
0.6861
0.8547
TABLE XI: UCIQE and its components for DMunderwater on RUIE using individual I2IT appearance targets and the datasets constructed from all 6 targets.
Fig. 12: DMunderwater outputs on RUIE after training with individual appearance targets ( a1 – a6 ) or combined-target datasets. The reddish–violet cast from a5 and yellow–orange cast from a2 contrast with the more balanced outputs from mixed-target supervision.
Model
Test set
UIEB
LSUI
EUVP
Nature
ImgNetSD
UWCNN
U45
3.125
3.312
3.062
3.438
3.562
RUIE
3.400
3.800
4.200
4.200
3.900
LSUI&EUVP
3.464
3.607
3.321
3.679
3.786
U-shape
U45
2.750
2.312
2.125
3.062
3.312
RUIE
3.800
3.800
3.900
4.000
4.400
LSUI&EUVP
2.964
3.250
3.000
3.750
3.179
TABLE XII: MOS grouped jointly by restoration architecture and test set, with five training sources compared on the same scenes.
Model
Test set
RV
Unclear
O
Total
UWCNN
U45
0
1
6
7
RUIE
1
0
6
7
LSUI&EUVP
2
0
8
10
UIEC 2 -Net
U45
1
4
2
7
RUIE
0
1
6
7
LSUI&EUVP
5
1
4
10
TABLE XIII: Outcomes of the paired O/RV comparison for each architecture and test set.
Test set
Local K=5
Global
Random
Shuffled
U45
0.870
0.830
0.786
0.777
RUIE
0.958
0.897
0.849
0.906
LSUI&EUVP
0.902
0.855
0.814
0.792
Macro mean
0.910
0.861
0.816
0.825
TABLE XIV: Mean cosine similarity between estimated inverse degradation and restoration directions for U-shape trained on UWNature (O).
Descriptor subset
U45
RUIE
LSUI&EUVP
All five
0.870
0.958
0.902
Color only ( a∗,b∗ )
0.885
0.981
0.963
Without a∗,b∗
0.873
0.937
0.669
Without L∗
0.893
0.969
0.935
TABLE XV: Mean directional cosine under different selections of appearance attributes for the same U-shape model.
Fig. 13: Paired-bootstrap comparison of UWNature (O) and depth-modulated UWNature (RV) across 6 architectures and 144 evaluation conditions. Within each test-set block, R indicates RV favored, O indicates original favored, and a dot indicates no clear difference under the 95% confidence interval. The distinct model- and test-set patterns show the nonuniform effects of depth modulation.
Existing paired clean-degraded datasets have advanced learning-based underwater image enhancement, but they remain either limited in scale or difficult to acquire, and still fail to adequately cover the diverse degradations encountered in real-world underwater environments. Such scarce and insufficient data ultimately restrict the generalization of underwater image enhancement models, highlighting the need for realistic and controllable degradation modeling to better capture real underwater conditions and improve enhancement robustness. To this end, we propose a unified degradation-aware training framework for the underwater image enhancement task. A Text-Conditioned Degradation Augmentation Pipeline uses a text-guided diffusion model to synthesize diverse and semantically consistent degraded samples, providing paired data to support the subsequent curriculum learning strategy. The proposed Progressive Degradation Curriculum organizes training across multiple attributes and difficulty levels to promote balanced learning, and a Transmission-Guided One-Step Diffusion Model leverages this progressive process to better address the wide range of degradation severity and restoration ambiguity in underwater images. Extensive experiments on multiple underwater benchmarks demonstrate that our proposed method achieves superior performance. Our code and models will be released upon the publication of this work.
Hongri Tian, Wenxue Li, Tian Ye +7
The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China. · China Geological Survey, Beijing, China. · Director of Applied Science, Amazon.com, Inc., U.S.. +1
Underwater computer vision tasks, such as detection, restoration, and segmentation, are limited by the scarcity of large-scale and diverse training data. We introduce WaterGen, a method for generating large-scale, realistic, and diverse underwater images that provides independent control of the scene and water medium conditions. Our approach treats underwater image generation as the decoupled control of two factors: realistic and diverse scene content (what is in the image), and accurate and controllable water medium effects (what the water does to the image). Existing methods generally achieve only part of this objective: they either provide controllability with limited realism or diversity, or generate realistic scenes without accurately and independently modeling water-medium effects. Our key insight, that allows us to avoid this compromise, is that scene generation and medium modeling can be decoupled within a latent diffusion framework, enabling diverse scene generation together with accurate and controllable underwater appearance. To do this, we decompose underwater image synthesis into two stages. First, we fine-tune the latent diffusion U-Net using degradation-free underwater images so that it learns to generate diverse and realistic latent embeddings of underwater scene content without medium-induced degradation. Second, we formulate the physically accurate medium degradation synthesis as a conditional decoding process applied to these latent embeddings. This decoupled design allows our model to generate diverse scenes with full control of underwater appearance. We leverage WaterGen to build large-scale synthetic underwater datasets that are diverse in scene structures and accurate in water effects and pseudo-labels. We demonstrate that our synthetic data consistently improve downstream performance in underwater restoration and semantic segmentation.
Jiayi Wu, Tianfu Wang, Tianyi Xiong +6
University of Maryland · University of South Florida · University of Florida
Underwater image enhancement aims to correct color and recover detail degraded by wavelength-dependent attenuation, scattering, and noise. Paired-reference supervision constrains output similarity, but does not by itself specify the color mapping, the strength of local detail corrections, or consistency with the degraded observation. We propose MCL-UIE, a paired-supervised framework that addresses these requirements by constraining restoration operations and the training objective. Monotonic Chromatic Transform (MCT) restricts global color mapping in log-LMS space through positive diagonal entries and bounded cross-channel coefficients. Uncertainty-Gated Wavelet Reconstruction (UGWR) modulates high-frequency residuals with a learned confidence gate calibrated against normalized wavelet reconstruction errors. During training, Reference-Anchored Degradation Replay (RADR) renders the enhanced image and its paired reference using shared latent degradation proxies, encouraging both to reproduce the input. The three components share a restoration network that combines structured color correction, locally controlled detail recovery, and reference-anchored replay. MCL-UIE ranks first in thirteen of sixteen full-reference dataset-metric comparisons and produces competitive visual results on no-reference benchmarks. Objective ablations indicate that confidence calibration is most effective when combined with replay supervision. RADR is used only during training and adds no inference-time computation. Our code is available at https://github.com/Wangtao-Bao/MCL-UIE.
Wangtao Bao, Liping Xie, Kanjian Zhang +1
Key Laboratory of Measurement and Control of Complex Systems of Engineering, Ministry of Education, School of Automation, Southeast University, Nanjing 210096, China · Advanced Ocean Institute of Southeast University, Nantong 226010, China