Organizations: Stazione Zoologica Anton Dohrn, Villa Comunale, 80121 Napoli, Italy. · School of Computer Science and Engineering, The University of Aizu, Aizuwakamatsu, Japan.
Underwater images often suffer from severe color distortion, low contrast, and a hazy appearance due to wavelength-dependent light absorption and scattering. Simultaneously, existing deep learning models exhibit high computational complexity, which limits their practical deployment for real-time underwater applications. To address these challenges, this paper presents a novel underwater image enhancement model, called Adaptive Frequency Fusion and Illumination Aware Network (AQUA-Net). It integrates a residual encoder-decoder with dual auxiliary branches, which operate in the frequency and illumination domains. The frequency fusion encoder enriches spatial representations with frequency cues from the Fourier domain and preserves fine textures and structural details. Inspired by Retinex, the illumination-aware decoder performs adaptive exposure correction through a learned illumination map that separates reflectance from lighting effects. This joint spatial, frequency, and illumination design enables the model to restore color balance, visual contrast, and perceptual realism under diverse underwater conditions. Additionally, we present a high-resolution, real-world underwater video-derived dataset from the Mediterranean Sea, which captures challenging deep-sea conditions with realistic visual degradations to enable robust evaluation and development of deep learning models. Extensive experiments on multiple benchmark datasets show that AQUA-Net performs on par with SOTA in both qualitative and quantitative evaluations while using less number of parameters. Ablation studies further confirm that the frequency and illumination branches provide complementary contributions that improve visibility and color representation. Overall, the proposed model shows strong generalization capability and robustness, and it provides an effective solution for real-world underwater imaging applications.
Figures & tables
Figure 1: Component-wise evaluation of AQUA-Net on an underwater image: (a) raw input, (b) encoder–decoder baseline, (c) encoder–decoder with an added frequency branch, and (d) full AQUA-Net combining global correction, frequency-guided enhancement, and illumination estimation.
Figure 2: AQUA-Net architecture. The network combines a frequency enhancement block, an illumination branch, and a multi-level encoder–decoder backbone with Residual Enhancement Modules (REMs). The frequency block output fuses with the encoder input. Skip connections integrate illumination information to enhance feature refinement and restore underwater image quality.
Figure 3: Residual Enhancement Module (REM) structure. The REM consists of two depthwise separable convolution blocks followed by a point-wise convolution with Leaky ReLU activation only in the first stage. Each block performs element-wise addition with its input to form a residual connection.
Methods
PSNR ↑
SSIM ↑
UIQM ↑
UCIQE ↑
Raw
16.134
0.748
2.346
0.362
Fusion(TIP’17) Ancuti et al. (2017)
18.033
0.861
2.684
0.406
SMBL(TB’20) Song et al. (2020)
16.513
0.781
2.167
0.455
MLLE(TIP’22) Zhang et al. (2022)
18.727
0.790
2.305
0.468
UWCNN(PR’20) Li et al. (2020)
18.147
0.847
2.878
0.357
WaterNet(TIP’19) Li et al. (2019)
19.914
0.859
2.846
0.410
Table 1: Quantitative evaluation on the UEIB dataset using PSNR, SSIM, UIQM, and UCIQE metrics. The top three results are shown in red , green , and blue .
Methods
UIEB-C60
EUVP-T515
RUIE-T78
DeapSea-T80
UIQM ↑
UCIQE ↑
UIQM ↑
UCIQE ↑
UIQM ↑
UCIQE ↑
UIQM ↑
UCIQE ↑
Raw
1.856
0.359
2.217
0.417
2.437
0.321
2.035
0.247
Fusion Ancuti et al. (2017)
2.163
0.378
2.636
0.430
2.772
0.366
2.698
0.393
SMBL Song et al. (2020)
1.724
0.439
1.857
0.513
2.459
0.431
2.407
0.382
MLLE Zhang et al. (2022)
1.956
0.464
2.354
0.461
2.798
0.441
2.843
0.436
WWPF Zhang et al. (2023)
2.1304
0.3943
2.3941
0.4695
2.7388
0.3726
2.485
0.386
Table 2: Quantitative comparison on UIEB-C60, EUVP-T515, RUIE-T78, and DeepSea-T80 in terms of UIQM and UCIQE. The top three scores are marked in red , green , and blue .
Methods
FLOPs (G) ↓
Parameters (M) ↓
Fusion (TIP’18) Ancuti et al. (2017)
–
–
SMBL (TB’20) Song et al. (2020)
–
–
MLLE (TIP’22) Zhang et al. (2022)
–
–
UWCNN (PR’20) Li et al. (2020)
11.36
0.04
WaterNet (TIP’19) Li et al. (2019)
310.82
1.09
PUIE (ECCV’22) Fu et al. (2022)
2073.2
0.83
Table 3: Comparison of different UIE models in terms of FLOPs (G) and parameter counts (M). The three lowest results are highlighted in red , green , and blue , respectively.
Figure 4: Comparison on the UEIB-T90 dataset showing two samples with greenish and bluish color casts. AQUA-Net effectively corrects the color casts while preserving structural details and natural appearance.
Figure 5: Comparison of the UEIB C60 challenging UIWs. AQUA-Net restores colors and maintains visual consistency.
Figure 6: Comparison on the RUIE-T78 underwater image subset with strong greenish degradation. AQUA-Net removes the greenish color cast and clearly restores object visibility, including the black sea urchin.
Figure 7: Qualitative comparison on the proposed DeepSea-T80 real deep-sea dataset. Four representative images with different color casts are enhanced by classical and DL-based UIE methods, as well as the proposed AQUA-Net. In these examples, AQUA-Net better preserves color balance and illumination and reduces reddish artifacts, resulting in a clearer separation between foreground structures and the background.
Modules
UIEB-T90
UIEB-C60
PSNR ↑
SSIM ↑
UIQM ↑
UCIQE ↑
UIQM ↑
UCIQE ↑
Raw
16.134
0.748
2.346
0.362
1.856
0.359
Base
18.473
0.832
2.872
0.377
1.602
0.418
Base + Frequency
20.614
0.872
3.089
0.415
2.233
0.436
Base + Illumination
20.730
0.879
3.086
0.391
2.123
0.414
Full Model
21.482
0.890
3.250
0.397
2.313
0.427
Table 4: Quantitative results of the ablation study. The table shows the contribution of each component in the AQUA-Net model. The top scores are marked in red .
Figure 8: Visual results of the ablation study. The frequency block improves object visibility, and the illumination block enhances overall scene clarity in the AQUA-Net model.
Figure 9: Qualitative comparison on the LSUI-400 test set using different underwater image enhancement methods. The comparison includes the raw input image, reference image, and enhanced results obtained using Fusion, WWPF, MLLE, WFAC, WaterNet, LMF, OUNet-JL, and AQUA-Net.
Methods
PSNR ↑
SSIM ↑
Fusion Ancuti et al. (2017)
20.13
0.8294
MLLE Zhang et al. (2022)
16.63
0.6491
WFAC Zhang et al. (2025a)
15.85
0.6665
WWPF Zhang et al. (2023)
16.97
0.6622
WaterNet Li et al. (2019)
22.75
0.8583
LMF Zhou et al. (2025)
21.44
0.8547
Table 5: Comparison with SOTA methods on the LSUI-400 test dataset. The top three results are shown in red , green , and blue , respectively.
Table 6: Downstream object detection comparison of different underwater image enhancement methods on the DUO dataset using YOLO12s.
Figure 10: Assessment of underwater object detection on the DUO dataset using different enhancement methods with the YOLO12s detection algorithm. The visual results show that AQUA-Net-enhanced images support reliable underwater object detection compared with alternative enhancement approaches.
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.
Yiqiang Zhou, Yifan Chen, Zhe Sun +3
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 +2
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
Underwater images often suffer from severe degradation, such as color distortion, low contrast, and blurred details, due to light absorption and scattering in water. While learning-based methods like CNNs and Transformers have shown promise, they face critical limitations: CNNs struggle to model the long-range dependencies needed for non-uniform degradation, and Transformers incur quadratic computational complexity, making them inefficient for high-resolution images. To address these challenges, we propose Hero-Mamba, a novel Mamba-based network that achieves efficient dual-domain learning for underwater image enhancement. Our approach uniquely processes information from both the spatial domain (RGB image) and the spectral domain (FFT components) in parallel. This dual-domain input allows the network to decouple degradation factors, separating color/brightness information from texture/noise. The core of our network utilizes Mamba-based SS2D blocks to capture global receptive fields and long-range dependencies with linear complexity, overcoming the limitations of both CNNs and Transformers. Furthermore, we introduce a ColorFusion block, guided by a background light prior, to restore color information with high fidelity. Extensive experiments on the LSUI and UIEB benchmark datasets demonstrate that Hero-Mamba outperforms state-of-the-art methods. Notably, our model achieves a PSNR of 25.802 and an SSIM of 0.913 on LSUI, validating its superior performance and generalization capabilities.
Tejeswar Pokuri, Shivarth Rai
Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal - 576104, India