Authors: Yiqiang Zhou, Yifan Chen, Zhe Sun, Jijun Lu, Ye Zheng, Xuelong Li
Organizations: Institute of Artificial Intelligence (TeleAI), China Telecom, P. R. China · College of Computer and Control Engineering, Northeast Forestry University, Harbin, 150040, China · College of Future Information Technology, Fudan University, Shanghai 200433, China · School of Artificial Intelligence, Optics and Electronics (iOPEN), Northwestern Polytechnical University, Xi’an 710072, China · Fujian Ocean Innovation Center, Xiamen 361102, China
Underwater image enhancement plays a crucial role in providing reliable visual information for underwater platforms, since strong absorption and scattering in water-related environments generally lead to image quality degradation. Existing high-performance methods often rely on complex architectures, which hinder deployment on underwater devices. Lightweight methods often sacrifice quality for speed and struggle to handle severely degraded underwater images. To address this limitation, we present a real-time underwater image enhancement framework with accurate color restoration. First, an Adaptive Weighted Channel Compensation module is introduced to achieve dynamic color recovery of the red and blue channels using the green channel as a reference anchor. Second, we design a Multi-branch Re-parameterized Dilated Convolution that employs multi-branch fusion during training and structural re-parameterization during inference, enabling large receptive field representation with low computational overhead. Finally, a Statistical Global Color Adjustment module is employed to optimize overall color performance based on statistical priors. Extensive experiments on eight datasets demonstrate that the proposed method achieves state-of-the-art performance across seven evaluation metrics. The model contains only 3,880 inference parameters and achieves an inference speed of 409 FPS. Our method improves the UCIQE score by 29.7% under diverse environmental conditions, and the deployment on ROV platforms and performance gains in downstream tasks further validate its superiority for real-time underwater missions.
Figures & tables
Fig. 1: Comparison of different methods in the FLOPs–PSNR space on UIEB dataset. The x-axis denotes FLOPs (G, 256 × 256 inputs), y-axis indicates PSNR (dB), and circle size is used to indicate parameter count.
Fig. 2: Overview of our proposed UIE framework. The pipeline begins with the AWCC module, which adaptively compensates degraded color channels and stabilizes luminance through secondary correction. The enhanced image is then processed by a CNN backbone incorporating MRDConv to capture multi-directional features with efficient inference. Finally, the SGCA module leverages global statistical priors to refine color temperature, tint, and saturation, producing high-fidelity enhancement with low computational overhead.
Fig. 3: Visualization of intermediate values for AWCC and the SGCA modules: The red and blue channels are visualized using the AUTUMN and WINTER colormaps, respectively.
Fig. 4: Architecture of the MRDConv and the Re-parameterization process.
Method
Venue
UIEB
LSUI
EUVP-S
PSNR ↑
SSIM ↑
LPIPS ↓
UIQM ↑
PSNR ↑
SSIM ↑
LPIPS ↓
UIQM ↑
PSNR ↑
SSIM ↑
LPIPS ↓
UIQM ↑
U-Shape [ 32 ]
IEEETIP’23
20.518
0.782
0.215
3.280
24.492
0.861
0.148
3.119
23.503
0.827
0.187
3.130
FiveA+ [ 47 ]
BMVC’23
21.566
0.892
0.134
3.092
23.883
0.901
0.164
2.993
25.968
0.887
0.159
2.932
Boths [ 50 ]
IEEEGRSL’23
22.230
0.901
0.156
3.273
21.221
0.851
0.222
3.083
23.487
0.838
0.268
3.038
SFGNet [ 48 ]
ICASSP’24
22.100
0.882
0.140
3.010
24.641
0.899
0.162
2.967
25.815
0.866
0.178
3.037
Lite [ 49 ]
ESWA’24
22.227
0.894
0.130
3.129
20.596
0.825
0.237
2.937
25.139
0.842
0.190
2.946
TABLE I: Quantitative comparison of performance and computational efficiency on UIEB, LSUI, EUVP-S, EUVP-D and EUVP-I datasets. Best results are in bold .
Fig. 5: Comparison of visual performance across different methods on five full-reference benchmark datasets.
Method
UIEB
LSUI
U45
RUIE
U45
RUIE
WaterNet [ 29 ]
0.5480
0.5154
0.5315
0.5116
U-shape [ 32 ]
0.5964
0.5529
0.5654
0.5489
UIE-DM [ 16 ]
0.5854
0.5430
0.5827
0.5700
SFGNet [ 48 ]
0.5937
0.5546
0.5799
0.5564
Lite [ 49 ]
0.5748
0.5396
0.5531
0.5424
TABLE II: UCIQE ↑ performance across different training (UIEB&LSUI) and validation (U45&RUIE) sets. Best results are in bold .
Camera
(A)
(B)
(C)
(D)
(E)
(F)
Ours
Canon D10
9.96
9.62
7.41
7.01
6.43
6.37
5.89
Fujifilm Z33
8.23
12.01
14.03
8.00
6.25
10.89
10.77
Olympus T6000
7.71
10.49
11.12
7.40
12.97
10.10
7.62
Olympus T8000
11.14
6.90
8.48
10.43
9.53
8.34
9.03
Panasonic TS1
9.81
12.01
10.89
17.52
14.41
12.29
13.95
Pentax W60
7.87
8.19
7.10
8.63
7.15
6.70
6.50
TABLE III: CIEDE2000 ↓ values across different camera models. Best results are in bold . (A) U-Shape [ 32 ] , (B) Lite [ 49 ] , (C) Boths [ 50 ] , (D) Phaseformer [ 34 ] , (E) SFGNet [ 48 ] , (F) MobileIE [ 43 ] .
Fig. 6: Visual comparison results on U45 and RUIE. From top to bottom, the rows represent inference configurations (Training → Testing): (1) UIEB → U45, (2) LSUI → U45, (3) UIEB → RUIE, and (4) LSUI → RUIE. Method: (A) U-shape, (B) WaterNet, (C) SFGNet, (D) Lite, and (E) MobileIE (please zoom in for better visualization).
Fig. 7: Qualitative results on diverse benchmarks. The bottom-right corners display the reference color checker for Colorcheck7 dataset.
Method
AWCC
MRDConv
SGCA
PSNR ↑
SSIM ↑
Baseline
22.227
0.891
(A)
✓
✓
22.518
0.881
(B)
✓
✓
23.686
0.886
(C)
✓
✓
23.370
0.894
Full
✓
✓
✓
24.055
0.902
TABLE IV: Ablation analysis of the three components.
Fig. 8: Visual ablation results of the proposed modules. The bottom row displays the corresponding sRGB color space point cloud distributions, illustrating the progressive correction of color degradation.
Method
μc
σc
vbright
vdark
PSNR ↑
SSIM ↑
LPIPS ↓
UIQM ↑
(A)
✓
✓
✓
24.034
0.898
0.123
3.055
(B)
✓
✓
✓
24.007
0.899
0.122
3.078
(C)
✓
✓
✓
23.721
0.894
0.128
3.090
(D)
✓
✓
✓
23.411
0.891
0.128
3.029
Full
✓
✓
✓
✓
24.055
0.902
0.121
3.045
TABLE V: Ablation study of different components in SGCA.
Fig. 9: Visual comparison of different components in the SGCA module.
Method
Lchar
Lpsnr
Lvgg
Lcolor
PSNR ↑
SSIM ↑
Baseline
✓
23.133
0.891
(A)
✓
✓
✓
23.522
0.890
(B)
✓
✓
✓
23.961
0.890
(C)
✓
✓
✓
24.014
0.886
Full
✓
✓
✓
✓
24.055
0.902
TABLE VI: Ablation study of different loss components.
Fig. 10: Visual ablation results of different loss functions (please zoom in for better visualization).
Method
Params
FLOPs
FPS
PSNR ↑
SSIM ↑
(K)
(G)
640×480
Rep-scale=4
3.88
0.145
409.09
24.055
0.902
Rep-scale=8
10.87
0.686
326.15
24.406
0.904
TABLE VII: Efficiency and performance analysis for MRDConv.
Fig. 11: Instance segmentation results on the UIIS dataset. Compared with other methods, our model not only enhances overall visual quality, but also preserves sharper boundaries and finer structural details, producing more accurate mask predictions for categories such as Human Divers and Aquatic Plants.
Method
Fish
Reefs
Aquatic Plants
Wrecks& Ruins
Human Divers
Robots
Sea-floor
mAP ↑
WaterMask [ 57 ]
0.442
0.260
0.076
0.184
0.473
0.075
0.066
0.225
Boths [ 50 ]
0.438
0.262
0.074
0.215
0.442
0.134
0.065
0.233
FiveA+ [ 47 ]
0.434
0.256
0.078
0.192
0.478
0.143
0.054
0.234
Lite [ 49 ]
0.432
0.261
0.074
0.192
0.458
0.120
0.063
0.229
MobileIE [ 43 ]
0.438
0.254
0.072
0.184
0.472
0.138
0.060
0.231
Ours
0.440
0.263
0.079
0.193
0.478
0.135
0.067
0.237
TABLE VIII: Instance segmentation performance (mAP) across different enhancement methods on the UIIS dataset.
NTU
Illum.
Raw
Enhanced
Δ%
σc
conl
μs
UCIQE ↑
σc
conl
μs
UCIQE ↑
0.55
I
0.189
0.624
0.836
0.475
0.287
0.926
0.842
0.605
27.3
II
0.187
0.628
0.833
0.474
0.270
0.949
0.841
0.604
27.3
0.99
I
0.162
0.545
0.815
0.435
0.288
0.926
0.852
0.608
39.7
II
0.161
0.557
0.814
0.438
0.262
0.941
0.849
0.600
36.9
2.41
I
0.183
0.537
0.782
0.435
0.279
0.949
0.839
0.607
39.7
TABLE IX: Quantitative quality assessment (UCIQE) under varying turbidity and illumination intensity.
Fig. 12: Visual results from controlled laboratory tank experiments under varying turbidity levels and illumination intensities, demonstrating the robustness of our method in diverse turbid and lighting conditions.
Fig. 13: Downstream task performance. After enhancement, the model enables precise background segmentation of pool boundaries and successful detection of the toy submarine, which are obscured in the raw frames. For the pool experiments, the stereo camera core is enclosed in housings designed for deep-sea environments.
Fig. 14: Real-world deployment. ROV operation resuspends bottom sediment, causing turbidity and reduced visibility (about 15 m depth, this condition does not correspond to the clear-water state of 2-3 NTU). For the Real-world experiments, the same stereo camera core is used without the deep-sea housings. The SIFT analysis shows that our method significantly increases the number of stable keypoint correspondences between consecutive frames.
Real-time underwater image enhancement (UIE) is crucial for mobile underwater photography and autonomous robotic systems, where practical deployment typically requires low latency and compact models under constrained computational resources. Recent ultra-lightweight CNNs based on structural re-parameterization meet these constraints but operate purely in the spatial domain, ignoring the frequency-sensitive nature of underwater degradation. To address this, we propose a lightweight UIE framework that integrates two key components: a Multi-Branch Reparameterizable Convolution with Fixed DCT Priors (MBRConv-DCT) that injects structured directional frequency priors during training, and a Frequency-Guided Dual-Path Attention (FGDPA) module that fuses spatial and spectral cues via a dual-path design for adaptive feature modulation. Both components are fully compatible with structural re-parameterization: the convolution branch introduces zero additional inference cost after re-parameterization, while the attention module incurs only a minimal computational overhead. Experiments show our model achieves state-of-the-art performance with only 4.23K parameters and 600+ FPS, outperforming much larger methods in both quantitative metrics and visual quality. Code is available at https://github.com/LethyZhang/FGDPA.
Leshen Zhang, Ao Li, Ce Zhu
Glasgow College, UESTC, Chengdu, China · University of Glasgow, Glasgow, UK · School of Information and Communication Engineering, UESTC, Chengdu, China
Due to wavelength dependent light absorption and scattering, underwater images usually suffer from color distortion and blurred details, which limits underwater object detection performance. Existing underwater image enhancement methods mainly focus on visual quality improvement, while it is still difficult to balance enhancement quality, processing efficiency, and downstream detection performance. Therefore, this paper proposes an efficient dual-branch underwater image enhancement framework for object detection. The detail enhancement branch improves brightness and local contrast to recover texture details in dark regions. The color restoration branch uses adaptive compensation to reduce color distortion and improve color gradation. By combining the complementary outputs of the two branches, the proposed framework provides clearer and more informative images for object detection. On the UIEB and EUVP datasets, the proposed method achieves UIQM scores of 2.249 and 2.576. When applied to the YOLOv8 detection task on the URPC dataset, the proposed method improves mAP50 by 2.1% compared with the baseline. Extensive experiments show that our method improves object detection in complex underwater scenes, while balancing enhancement quality and processing efficiency.
Liyuan Cao, Zheng Liu, Guanghao Liao +2
School of Electronic and Information Engineering, University of Science and Technology Liaoning, No.189 Qianshan Middle Road, Lishan District, Anshan, 114051, Liaoning, China
Underwater image enhancement is crucial for improving visual perception in marine applications. Existing underwater image enhancement studies mainly focus on enhancement quality and visual fidelity, while rarely considering real-time deployment capability, which is essential for resource-constrained underwater robots. To this end, we introduce a physics-guided spectral distillation (PSD) method, which reduces model capacity for real-time applications while maintaining the high performance of underwater image enhancement models. To decompose the outputs of teacher and student models, PSD adopts a multilevel Haar discrete wavelet transform. It transfers low-frequency color and illumination information as well as high-frequency structural details through band-specific objectives. Moreover, the distillation process of PSD is degradation-aware. We estimate degradation-aware weights through a physical head and combine them with ground-truth-guided reliability masks to selectively retain valuable teacher guidance. Experiments on the UIEB, LSUI, and EUVP datasets validate the effectiveness of the proposed method. Furthermore, we demonstrate the benefits of enhanced images for downstream perception tasks, including object detection. Deployment on a self-developed ROV further demonstrates its practical applicability in real-world underwater scenarios.
Yifan Chen, Kai He, Ye Zheng +3
College of Future Information Technology, Fudan University, Shanghai 200433, China · Institute of Artificial Intelligence (TeleAI), China Telecom, China · School of Computer Science and Technology, Harbin Institute of Technology, Weihai 264209, China