Back

An Underwater Image Enhancement Benchmark Dataset and Beyond

Chongyi LiChunle GuoWenqi RenRunmin CongJunhui HouSam KwongDacheng Tao

article2019IEEE Transactions on Image Processing2,131 citationsESI Highly Cited Paper (2020-2026), ESI Hot Paper (2022)

Establishes a standardized real-world benchmark dataset containing 950 underwater images alongside a convolutional baseline network, Water-Net, to rigorously evaluate and advance underwater image restoration.

Listen

Underwater imaging plays a critical role in marine biology, archaeology, environmental monitoring, and aquatic robotics. However, underwater photographs frequently suffer from severe color casts, reduced contrast, and heavy haze caused by light absorption and scattering in water. While numerous enhancement algorithms have been proposed in recent years, their real-world effectiveness has remained difficult to gauge because existing evaluations rely heavily on synthetic datasets or small, unrepresentative image samples. The article set out to address this gap by establishing a comprehensive, real-world benchmark to systematically evaluate existing enhancement techniques and to train an effective deep learning model for real-world underwater enhancement.

To achieve this, the article introduces a benchmark consisting of 950 real underwater images captured under diverse lighting and environmental conditions. High-quality reference counterparts were established for 890 of these images through extensive pairwise visual comparisons conducted by 50 human observers across twelve enhancement techniques, while the remaining 60 images without satisfactory outputs were designated as a challenging test set. Using this benchmark, the article evaluated state-of-the-art physical, non-physical, and data-driven methods across visual quality, full-reference metrics, non-reference metrics, and computational runtime. Additionally, the article developed Water-Net, a baseline convolutional neural network trained on this dataset using a gated fusion strategy and a perceptual loss function to combine white balancing, histogram equalization, and gamma correction.

Key findings show that no single existing enhancement technique consistently succeeds across all underwater scenarios. Fusion-based and commercial enhancement tools produced the most visually pleasing references, while physical-model methods frequently failed due to inaccurate light attenuation assumptions. Crucially, the analysis revealed that commonly used non-reference underwater quality metrics often disagree with human visual judgment, rewarding artificial contrast and color shifts that human observers rated poorly. In quantitative and qualitative tests on unseen images, the proposed Water-Net model outperformed both traditional and generative deep learning methods, delivering superior contrast, natural colors, and fast processing speeds of roughly eight frames per second.

These findings indicate that future research and operational deployments must move away from simplified optical models and unreliable non-reference metrics. Instead, engineering teams and researchers should adopt data-driven fusion architectures and physically accurate imaging formulations, while developing better quality metrics aligned with human perception. Although the benchmark significantly advances the field, the authors note limitations: far-distance backscatter remains difficult to fully eliminate, and some human observers may overlook subtle physical scattering artifacts. Moving forward, incorporating depth maps, expanding datasets to include video sequences, and refining reference selection with advanced optical physics will be essential next steps before deploying these systems in highly automated aquatic missions.

Beijing Jiaotong UniversityCity University of Hong KongInstitute of Information Engineering, Chinese Academy of SciencesTianjin UniversityUniversity of Sydneyconvolutional neural networksimage enhancementreference imageunderwater image enhancement benchmark datasetsingle underwater image enhancement algorithmsunderwater color image quality evaluation
arXiv: 1901.05495Li-Chongyi/Water-Net_Code
  • Paper: Benchmarking Single-Image Dehazing and Beyond, Boyi Li et al. (2017). Its systematic dehazing benchmark provides a useful precedent for understanding this paper’s evaluation across algorithms, objective metrics, and human judgments.
Cover for An Underwater Image Enhancement Benchmark Dataset and Beyond

Abstract

Underwater image enhancement has been attracting much attention due to its significance in marine engineering and aquatic robotics. Numerous underwater image enhancement algorithms have been proposed in the last few years. However, these algorithms are mainly evaluated using either synthetic datasets or few selected real-world images. It is thus unclear how these algorithms would perform on images acquired in the wild and how we could gauge the progress in the field. To bridge this gap, we present the first comprehensive perceptual study and analysis of underwater image enhancement using large-scale real-world images. In this paper, we construct an Underwater Image Enhancement Benchmark (UIEB) including 950 real-world underwater images, 890 of which have the corresponding reference images. We treat the rest 60 underwater images which cannot obtain satisfactory reference images as challenging data. Using this dataset, we conduct a comprehensive study of the state-of-the-art underwater image enhancement algorithms qualitatively and quantitatively. In addition, we propose an underwater image enhancement network (called Water-Net) trained on this benchmark as a baseline, which indicates the generalization of the proposed UIEB for training Convolutional Neural Networks (CNNs). The benchmark evaluations and the proposed Water-Net demonstrate the performance and limitations of state-of-the-art algorithms, which shed light on future research in underwater image enhancement. The dataset and code are available at this https URL.

Topics

convolutional neural networksimage enhancementreference imageroboticsunderwater imageunderwater sensor networks

Keywords

underwater image enhancement benchmark datasetsingle underwater image enhancement algorithmsunderwater color image quality evaluationnon-reference underwater image qualityfull-reference image quality evaluationimage quality assessmentimage restorationimage dehazingcontrast enhancement

Table of Contents

  • I Introduction
  • II Existing Methodology, Evaluation Metric, and Dataset: An Overview
  • II-A Underwater Image Enhancement Method
  • II-B Underwater Image Quality Evaluation
  • II-C Underwater Image Datasets
  • III Proposed Benchmark Dataset
  • III-A Data Collection
  • III-B Reference Image Generation
  • IV Evaluation and Discussion
  • IV-A Qualitative Evaluation
  • IV-B Quantitative Evaluation
  • IV-B1 Full-reference Evaluation
  • IV-B2 Non-reference Evaluation
  • IV-B3 Runtime Evaluation
  • V Proposed Model
  • V-A Input Generation
  • V-B Network Architecture
  • V-C Implementations
  • V-D Experiments
  • VI Conclusion, Limitations, and Future Work
  • References

Knowls

  1. Knowl 1 — Underwater Image Enhancement Benchmark (UIEB) Dataset Composition

    definition

    The Underwater Image Enhancement Benchmark (UIEB) is a real-world dataset comprising 950950 real underwater images divided into two subsets:

    1. Reference Subset (890890 image pairs): 890890 raw underwater images paired with high-quality reference images selected via a rigorous subjective pairwise voting process.
    2. Challenging Subset (6060 images): 6060 raw underwater images for which none of the candidate enhancement algorithms produced a satisfactory result (greater than 50%50\% dissatisfaction votes by human observers).

    The dataset covers a diverse range of image resolutions, underwater degradation patterns (such as severe greenish and bluish color casts, low and high backscatter, forward-looking and downward-looking orientations), and scene contents (including fringing reefs, barrier reefs, marine life such as turtles and sharks, shipwrecks, and underwater sculptures) captured under natural, artificial, or mixed lighting conditions.

  2. Knowl 2 — Reference Image Selection via Subjective Pairwise Voting Protocol

    model/method

    Due to the physical impossibility of simultaneously capturing real degraded underwater scenes and pristine ground truth images, reference images in the UIEB dataset are established through a multi-algorithm candidate pool and human pairwise evaluation:

    1. Candidate Pool Generation: For each raw underwater image, 1212 candidate enhanced images are generated using 99 underwater enhancement methods (fusion-based, two-step-based, Retinex-based, Underwater Dark Channel Prior [UDCP], regression-based, Generalized Dark Channel Prior [GDCP], Red Channel, histogram prior, and blurriness-based), 22 single-image dehazing methods (Dark Channel Prior [DCP] and Multi-Scale Convolutional Neural Network [MSCNN]), and 11 specialized commercial underwater application (dive+).
    2. Pairwise Voting: 5050 human evaluators (2525 with image processing domain experience and 2525 non-experts) perform 1111 rounds of randomized pairwise comparisons per image on a standardized monitor to select the visually superior result.
    3. Quality Verification and Thresholding: Evaluators label the chosen best result as satisfactory or unsatisfactory. If more than 50%50\% of participants judge the winning candidate unsatisfactory, the candidate is discarded and the raw image is assigned to the challenging subset (6060 images). Otherwise, the winning image becomes the official reference (890890 images).

    The distribution of selected reference images across the candidate generation methods is as follows:

    Method Percentage (%)
    dive+ 43.93
    fusion-based 24.72
    histogram prior 13.37
    two-step-based 7.30
    blurriness-based 3.48
    DCP 2.58
    regression-based 1.80
    MSCNN 0.90
    Red Channel 0.90
    GDCP 0.34
    retinex-based 0.22
    UDCP 0.00

  3. Knowl 3 — Water-Net Gated Fusion Architecture

    model/method

    Water-Net is a gated fusion convolutional neural network designed for underwater image enhancement. Given a raw input image IRAWI_{\text{RAW}}, the framework derives three distinct pre-processed inputs to address specific underwater degradation characteristics:

    1. White Balance (IWBI_{\text{WB}}): Removes non-uniform color casts via color constancy adjustment.
    2. Histogram Equalization (IHEI_{\text{HE}}): Improves contrast by applying Contrast Limited Adaptive Histogram Equalization (CLAHE via adapthisteq) on the LL luminance channel in the CIELAB color space and converting back to RGB.
    3. Gamma Correction (IGCI_{\text{GC}}): Lightens under-exposed and dark regions using power-law gamma correction with γ=0.7\gamma = 0.7.

    Each pre-processed image Ik∈{IWB,IHE,IGC}I_k \in \{I_{\text{WB}}, I_{\text{HE}}, I_{\text{GC}}\} is concatenated with IRAWI_{\text{RAW}} and passed through an individual Feature Transformation Unit (FTU) consisting of three convolutional layers (7×7×327\times 7\times 32, 5×5×325\times 5\times 32, and 3×3×33\times 3\times 3, each with ReLU activation) to suppress artifacts and color distortions introduced during pre-processing, yielding refined inputs RWB,RHE,RGCR_{\text{WB}}, R_{\text{HE}}, R_{\text{GC}}.

    Simultaneously, the concatenation of IRAW,IWB,IHE,IGCI_{\text{RAW}}, I_{\text{WB}}, I_{\text{HE}}, I_{\text{GC}} (1212 channels) is processed by a confidence prediction CNN subnet consisting of eight convolutional layers (filter sizes 7×7×1287\times 7\times 128, 5×5×1285\times 5\times 128, 3×3×1283\times 3\times 128, 1×1×641\times 1\times 64, 7×7×647\times 7\times 64, 5×5×645\times 5\times 64, 3×3×643\times 3\times 64 with ReLUs, and 3×3×33\times 3\times 3 with a Sigmoid activation) to produce three spatial confidence maps CWB,CHE,CGCC_{\text{WB}}, C_{\text{HE}}, C_{\text{GC}}.

  4. Knowl 4 — Water-Net Fusion Formulation and Perceptual Loss

    equation

    The final enhanced image IenI_{\text{en}} produced by Water-Net is obtained by element-wise linear blending of the refined inputs weighted by their corresponding learned confidence maps:

    Ien=RWB⊙CWB+RHE⊙CHE+RGC⊙CGCI_{\text{en}} = R_{\text{WB}} \odot C_{\text{WB}} + R_{\text{HE}} \odot C_{\text{HE}} + R_{\text{GC}} \odot C_{\text{GC}}

    where ⊙\odot represents the Hadamard (element-wise) matrix product; RWB,RHE,RGC∈RH×W×3R_{\text{WB}}, R_{\text{HE}}, R_{\text{GC}} \in \mathbb{R}^{H \times W \times 3} are the refined versions of the white-balanced, histogram-equalized, and gamma-corrected inputs generated by the Feature Transformation Units; and CWB,CHE,CGC∈[0,1]H×W×3C_{\text{WB}}, C_{\text{HE}}, C_{\text{GC}} \in [0, 1]^{H \times W \times 3} are the predicted confidence maps.

    The network parameters are optimized by minimizing a perceptual loss defined on the feature space of the pre-trained 19-layer VGG network ϕ\phi at the relu5_4 layer:

    Ljϕ=1CjHjWj∑i=1N∥ϕj(Ien(i))−ϕj(Igt(i))∥1\mathcal{L}_j^{\phi} = \frac{1}{C_j H_j W_j} \sum_{i=1}^{N} \|\phi_j(I_{\text{en}}^{(i)}) - \phi_j(I_{\text{gt}}^{(i)})\|_1

    where NN is the batch size, ϕj(⋅)\phi_j(\cdot) denotes the feature map output of the jj-th layer (relu5_4) of VGG-19 pre-trained on ImageNet, Igt(i)I_{\text{gt}}^{(i)} is the ii-th target reference image, and Cj,Hj,WjC_j, H_j, W_j denote the channel depth, height, and width of the feature maps in layer jj.

  5. Knowl 5 — Quantitative Benchmark of Classical and Physical Underwater Enhancement Algorithms

    data/table

    A quantitative evaluation of state-of-the-art single underwater image enhancement algorithms on the 890890 paired images of the UIEB dataset reveals substantial differences between full-reference metrics (MSE, PSNR, SSIM) and underwater no-reference metrics (UCIQE, UIQM):

    Method MSE (×103\times 10^3) ↓\downarrow PSNR (dB) ↑\uparrow SSIM ↑\uparrow UCIQE ↑\uparrow UIQM ↑\uparrow
    dive+ 0.5358 20.8408 0.8705 0.6227 1.3410
    fusion-based 0.8679 18.7461 0.8162 0.6414 1.5310
    two-step-based 1.1146 17.6596 0.7199 0.5776 1.4002
    regression-based 1.1365 17.5751 0.6543 0.5971 1.2996
    retinex-based 1.3531 16.8757 0.6233 0.6062 1.4338
    blurriness-based 1.5826 16.1371 0.6582 0.6001 1.3757
    histogram prior 1.6282 16.0137 0.5888 0.6778 1.5440
    Red Channel 2.1073 14.8935 0.5973 0.5421 1.2147
    GDCP 3.6345 12.5264 0.5503 0.5993 1.4301
    UDCP 5.1300 11.0296 0.4999 0.5852 1.6297

    In full-reference assessment, dive+ and the fusion-based method attain the highest scores due to faithful color restoration and structural fidelity. Physical model-based methods (UDCP, GDCP, Red Channel) perform poorly in full-reference metrics because outdoor haze physical models fail to account for wavelength-dependent attenuation coefficients.

  6. Knowl 6 — Discrepancy Between Underwater Non-Reference Metrics and Human Visual Perception

    empirical result

    Standard no-reference image quality metrics specifically designed for underwater images—such as the Underwater Color Image Quality Evaluation (UCIQE) and the Underwater Image Quality Measure (UIQM)—exhibit a systematic discrepancy with human subjective visual assessment and pairwise preference voting:

    1. High Scores for Over-Enhanced and Distorted Results: Methods that introduce heavy reddish color shifts, over-saturation, and high local contrast (such as the histogram prior method and UDCP) achieve the highest scores on UCIQE (0.67780.6778) and UIQM (1.62971.6297).
    2. Lower Scores for Visually Superior Results: Methods with superior subjective visual quality and color balance (such as dive+ and fusion-based methods) receive lower UCIQE (0.62270.6227) and UIQM (1.34101.3410) scores despite matching human preferences in user studies.
    3. Failure Mode: UCIQE and UIQM evaluate linear combinations of chroma, saturation, contrast, and sharpness, but fail to penalize unnatural color deviations, artifacts, and noise across the full image. Consequently, optimizing for or relying on UCIQE and UIQM can lead to degraded, unnatural underwater image restoration.

  7. Knowl 7 — Quantitative Performance Comparison of Water-Net on UIEB Testing Set

    data/table

    The performance of Water-Net was evaluated against deep learning models and conventional enhancement algorithms on the UIEB test set consisting of 9090 paired real-world underwater images (800800 pairs used for training):

    Method MSE (×103\times 10^3) ↓\downarrow PSNR (dB) ↑\uparrow SSIM ↑\uparrow
    Water-Net 0.7976 19.1130 0.7971
    fusion-based 1.1280 17.6077 0.7721
    Dense GAN 1.2152 17.2843 0.4426
    retinex-based 1.2924 17.0168 0.6071
    histogram prior 1.7019 15.8215 0.5396
    Water CycleGAN 1.7298 15.7508 0.5210
    blurriness-based 1.9111 15.3180 0.6029
    GDCP 4.0160 12.0929 0.5121

    Water-Net achieves the lowest MSE (0.7976×1030.7976 \times 10^3) and the highest PSNR (19.1130 dB19.1130\text{ dB}) and SSIM (0.79710.7971), outperforming generative adversarial network baselines (Water CycleGAN and Dense GAN) as well as classical hand-crafted prior and fusion methods.

  8. Knowl 8 — Subjective User Study on Challenging Underwater Images

    data/table

    To evaluate performance on the 6060 challenging underwater images from the UIEB dataset where no reference could be agreed upon, a subjective user study was conducted with 5050 participants using a 5-point Mean Opinion Score (MOS) scale (55: Excellent, 44: Good, 33: Fair, 22: Poor, 11: Bad):

    Method Average Score ↑\uparrow Standard Deviation ↓\downarrow
    Water-Net 2.57 0.7280
    fusion-based 2.28 0.8475
    retinex-based 2.23 0.8720
    histogram prior 2.08 0.7897
    blurriness-based 2.02 0.7762
    GDCP 1.90 0.8099

    Water-Net achieves the highest average score (2.572.57) and lowest variance (standard deviation 0.72800.7280), demonstrating greater robustness across severe degradation conditions compared to conventional priors that introduce heavy color casts and over-enhancement.

  9. Knowl 9 — Computational Runtime Across Image Resolutions

    data/table

    Average computational runtimes across different input resolutions (500×500500\times 500, 640×480640\times 480, and 1280×7201280\times 720 pixels) evaluated in MATLAB R2014b on an Intel Core i7-6700 CPU with 32 GB RAM:

    Method 500×500500 \times 500 640×480640 \times 480 1280×7201280 \times 720
    two-step-based 0.2978 s 0.4391 s 1.0361 s
    fusion-based 0.6044 s 0.6798 s 1.8431 s
    retinex-based 0.6975 s 0.8829 s 2.1089 s
    UDCP 2.2688 s 3.3185 s 9.9019 s
    Red Channel 2.7523 s 3.2503 s 9.7447 s
    GDCP 3.2676 s 3.8974 s 9.5934 s
    histogram prior 4.6284 s 5.8289 s 16.9229 s
    blurriness-based 37.0018 s 47.2538 s 146.0233 s
    regression-based 138.6138 s 167.1711 s 415.4935 s

    The two-step method is the fastest classical approach across all resolutions. In comparison, Water-Net implemented in TensorFlow processes a 640×480640\times 480 image in 0.128 s0.128\text{ s} ( ≈8 frames per second\,\approx 8\text{ frames per second}) on an Nvidia GeForce GTX 1080Ti GPU.

  10. Knowl 10 — Limitations in Distant Backscatter Removal and Reference Image Annotation

    limitation

    The construction of the UIEB dataset and training of deep models thereon have two primary limitations:

    1. Persistent Distant Backscatter: Current enhancement algorithms assume uniform, simplified medium transmission models that fail to capture the exponentially increasing physical backscatter at larger scene distances. As a result, candidate enhancement algorithms cannot fully eliminate dense backscatter in distant scene regions, causing the selected reference images to retain residual backscatter.
    2. Human Annotation Bias: Human evaluators participating in the reference selection voting process often prioritize high-contrast foreground objects and salient colors (such as divers, fish, or corals) and fail to penalize physical inaccuracies or incomplete backscatter suppression in the background regions.

Coverage note — None was omitted; all primary contributed dataset details, reference selection procedures, network architecture, training equations, empirical evaluations, user study results, and stated limitations are fully represented.

References

  1. 1.J. Jaffe, ‘‘Underwater optical imaging: The past, the present, and the prospects,’’ IEEE J. Oceanic Eng., vol. 40, no. 3, pp. 683-700, 2015.
  2. 2.M. Sheinin and Y. Schechner, ‘‘The next best underwater view,’’ in Proc. of IEEE Int. Conf. Comput. Vis. Pattern Rec. (CVPR), 2016, pp. 3764-3773.
  3. 3.B. McGlamery, ‘‘A computer model for underwater camera systems,’’ in Ocean Optics VI, pp. 221-231, 1980.
  4. 4.J. Jaffe, ‘‘Computer modeling and the design of optimal underwater imaging systems,’’ IEEE J. Oceanic Eng., vol. 15, no. 2, pp. 101-111, 1990.
  5. 5.W. Hou, S. Woods, E. Jarosz, et al., ‘‘Optical turbulence on underwater image degradation in natural environments,’’ Appl. Opt., vol. 15, no. 14, pp. 2678-2686, 2012.
  6. 6.D. Akkaynak and T. Treibitz, ‘‘A revised underwater image formation model,’’ in Proc. of IEEE Int. Conf. Comput. Vis. Pattern Rec. (CVPR), 2018, pp. 6723-6732.
  7. 7.D. Akkaynak, T. Treibitz, T. Shlesinger, et al., ‘‘What is the space of attenuation coefficients in underwater computer vision,’’ in Proc. of IEEE Int. Conf. Comput. Vis. Pattern Rec. (CVPR), 2017, pp. 568-577.
  8. 8.M. Ludvigsen, B. Sortland, G. Johnsen, et al., ‘‘Applications of geo-referenced underwater photo mosaics in marine biology and archaeology,’’ J. Oceanography, vol. 20, no. 4, pp. 140-149, 2007.
  9. 9.N. Strachan, ‘‘Recognition of fish species by colour and shape,’’ J. Image Vis. Comput., vol. 11, no. 1, pp. 2-10, 1993.
  10. 10.M. Yang, J. Hu, C. Li, et al., ‘‘An in-depth survey of underwater image enhancement and restoration,’’ IEEE Access, vol. 7, pp. 123638-123657, 2019.
  11. 11.C. Li, C. Guo, J. Guo, et al., ‘‘PDR-Net: Perception-inspired single image dehazing network with refinement,’’ IEEE Trans. Multimedia, 2019.
  12. 12.W. Wang, Q. Lai, H. Fu, et al., ‘‘Salient object detection in the deep learning era: An in-depth survey,’’ in arXiv preprint arXiv: 1904.09146, 2019.
  13. 13.C. Li, R. Cong, J. Hou, et al., ‘‘Nested network with two-stream pyramid for salient object detection in optical remote sensing images,’’ IEEE Trans. Geosci. Remote Sens., vol. 57, no. 11, pp. 9156-9166, 2019.
  14. 14.R. Cong, J. Lei, H. Fu, et al., ‘‘An iterative co-saliency framework for RGBD images,’’ IEEE Trans. Cybernetics., no. 49, no. 1, pp. 233-249, 2019.
  15. 15.C. Guo, C. Li, J. Guo, et al., ‘‘Hierarchical features driven residual learning for depth map super-resolution,’’ IEEE Trans. Image Process., vol. 28, no. 5, pp. 2545-2557, 2019.
  16. 16.R. Schettini and S. Corchs, ‘‘Underwater image processing: State of the art of restoration and image enhancement methods,’’ J. Advan. Signal Process., vol. 2010, pp. 1-14, 2010.
  17. 17.M. Han, Z. Lyu, and T. Qiu, ‘‘A review on intelligence dehazing and color restoration for underwater images,’’ IEEE Trans. Syst., Man, Cybern., Syst., pp. 1-13, 2018 (Early Access).
  18. 18.R. Cui, L. Chen, C. Yang, et al., ‘‘Extended state observer-based integral sliding mode control for an underwater robot with unknown disturbances and uncertain nonlinearities,’’ IEEE Trans. Ind. Electron., vol. 64, no. 8, pp. 6785-6795, 2017.
  19. 19.S. Narasimhan and S. Nayar, ‘‘Contrast restoration of weather degraded images,’’ IEEE Trans. Pattern Anal. Mach. Intell., vol. 25 no. 6, pp. 713-724, 2003.
  20. 20.N. Shashar, R. Hanlon, and A. deM. Petz, ‘‘Polarization vision helps detect transparent prey,’’ Nature, vol. 393 no. 6682, pp. 222-223, 1998.
  21. 21.Y. Schechner and N. Karpel, ‘‘Clear underwater vision,’’ in Proc. of IEEE Int. Conf. Comput. Vis. Pattern Rec. (CVPR), 2004, pp. 536-543.
  22. 22.Y. Schechner and N. Karpel, ‘‘Recovery of underwater visibility and structure by polarization analysis,’’ IEEE J. Oceanic Eng., vol. 30, no. 3, pp. 570-587, 2005.
  23. 23.T. Treibitz and Y. Schechner, ‘‘Active polarization descattering,’’ IEEE Trans. Pattern Anal. Mach. Intell., vol. 31, no. 3, pp. 385-399, 2009.
  24. 24.C. Tan, G. Sluzek, and D. He, ‘‘A novel applications of range-gated underwater laser imaging system in near target turbid medium,’’ Opticas and Lasers Eng., vol. 43, no. 9, pp. 995-1009, 2005.
  25. 25.H. Li, X. Wang, T. Bai, et al., ‘‘Spoeckle noise suppression of range gated underwater imaging system,’’ Applid Optics, vol. 38, no. 18, pp. 3937-3944, 2009.
  26. 26.T. Treibitz and Y. Schechner, ‘‘Turbid scene enhancement using multi-directional illumination fusion,’’ IEEE Trans. Image Process., vol. 21, no. 11, pp. 4662-4667, 2012.
  27. 27.Z. Murez, T. Treibitz, R. Ramamoorthi, et al., ‘‘Photometric stereo in a scattering medium,’’ in Proc. of IEEE Int. Conf. Comput. Vis. Pattern Rec. (CVPR), 2015, pp. 3415-3423.
  28. 28.K. Iqbal, M. Odetayo, and A. James, ‘‘Enhancing the low quality images using unsupervised colour correction method,’’ in Proc. of IEEE Int. Conf. Syst., Man, Cybern., 2010, pp. 1703-1709.
  29. 29.A. Ghani and N. Isa, ‘‘Underwater image quality enhancement through integrated color model with Rayleigh distribution,’’ Appl. Soft Comput., vol. 27, pp. 219-230, 2015.
  30. 30.A. Ghani and N. Isa, ‘‘Enhancement of low quality underwater image through integrated global and local contrast correction,’’ Appl. Soft Comput., vol. 37, pp. 332-344, 2015.
  31. 31.C. Ancuti, C. O. Ancuti, and P. Bekaert, ‘‘Enhancing underwater images and videos by fusion,’’ in Proc. of IEEE Int. Conf. Comput. Vis. Pattern Rec. (CVPR), 2012, pp. 81-88.
  32. 32.X. Fu, Z. Fan, and M. Ling, ‘‘Two-step approach for single underwater image enhancement,’’ in Symposium. of IEEE Intell. Signal Process. Commun. Syst., 2017, pp. 789-794.
  33. 33.X. Fu, P. Zhang, Y. Huang, et al., ‘‘A retinex-based enhancing approach for single underwater image,’’ in Proc. of IEEE Int. Conf. Image Process. (ICIP), 2014, pp. 4572-4576.
  34. 34.S. Zhang, T. Wang, J. Dong, et al., ‘‘Underwater image enhancement via extended multi-scale Retinex,’’ Neurocomput., vol. 245, no. 5, pp. 1-9, 2017.
  35. 35.K. He, J. Sun, and X. Tang, ‘‘Single image haze removal using dark channel prior,’’ IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 12, pp. 2341-2353, 2011.
  36. 36.J. Chiang and Y. Chen, ‘‘Underwater image enhancement by wavelength compensation and dehazing,’’ IEEE Trans. Image Process., vol. 21, no. 4, pp. 1756-1769, 2012.
  37. 37.P. Drews-Jr, E. Nascimento, S. Botelho, et al., ‘‘Underwater depth estimation and image restoration based on single images,’’ IEEE Comput. Graph. Appl., vol. 36, no. 2, pp. 24-35, 2016.
  38. 38.H. Liu and L. Chau, ‘‘Underwater image restoration based on contrast enhancement,’’ in Proc. of IEEE Int. Digit. Signal Process., 2016, pp. 584-588.
  39. 39.C. Li, J. Guo, C. Guo, et al., ‘‘A hybrid method for underwater image correction,’’ Pattern Rec. Lett., vol. 94, pp. 62-67, 2017.
  40. 40.Y. Peng, T. Cao, and P. Cosman, ‘‘Generalization of the dark channel prior for single image restoration,’’ IEEE Trans. Image Process., vol. 27, no. 6, pp. 2856-2868, 2018.
  41. 41.N. Carlevaris-Bianco, A. Mohan, and R. Eustice, ‘‘Initial results in underwater single image dehazing,’’ in Proc. of IEEE Int. Conf. Oceans, 2010, pp. 1-8.
  42. 42.A. Galdran, D. Pardo, and A. Picn, ‘‘Automatic Red-Channel underwater image restoration,’’ J. Vis. Commu. and Image Repre., vol. 26, pp. 132-145, 2015.
  43. 43.X. Zhao, T. Jin, and S. Qu, ‘‘Deriving inherent optical properites from background color and underwater image enhancement,’’ Ocean Eng., vol. 94, pp. 163-172, 2015.
  44. 44.C. Li, J. Guo, S. Chen, et al., ‘‘Underwater image restoration based on minimum information loss principle and optical properties of underwater imaging,’’ in Proc. of IEEE Int. Conf. Image Process. (ICIP), 2016, pp. 1993-1997.
  45. 45.C. Li, J. Guo, R. Cong, et al., ‘‘Underwater image enhancement by dehazing with minimum information loss and histogram distribution prior,’’ IEEE Trans. Image Process., vol. 25, no. 12, pp. 5664-5677, 2016.
  46. 46.Y. Peng and P. Cosman, ‘‘Underwater image restoration based on image blurriness and light absorption,’’ IEEE Trans. Image Process., vol. 26, no. 4, pp. 1579-1594, 2017.
  47. 47.D. Berman, T. Treibitz, and S. Avidan, ‘‘Diving into haze-lines: Color restoration of underwater images,’’ in Proc. of Briti. Mach. Vis. Conf. (BMVC), 2017, pp. 1-12.
  48. 48.Y. Wang, H. Liu, and L. Chau, ‘‘Single underwater image restoration using adaptive attenuation-curve prior,’’ IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 65, no. 3, pp. 992-1002, 2018.
  49. 49.D. Akkaynak and T. Treibitz, ‘‘ Sea-thru: A method for removing water from underwater images,’’ in Proc. of IEEE Int. Conf. Comput. Vis. Pattern Rec. (CVPR), 2019, pp. 1682-1691.
  50. 50.S. Anwar and C. Li, ‘‘ Diving deeper into underwater image enhancement: A survey,’’ in arXiv preprint arXiv: 1907.07863, 2019.
  51. 51.J. Li, K. Sinner, R. Eustice, et al., ‘‘WaterGAN: Unsupervised generative network to enable real-time color correction of monocular underwater images,’’ IEEE Robot. Autom. Lett., vol. 3, no. 1, pp. 387-394, 2018.
  52. 52.C. Li, S. Anwar, and F. Porikli, ‘‘Underwater scene prior inspired deep underwater image and video enhancement,’’ Pattern Rec., vol. 98, pp. 107038-107049, 2020.
  53. 53.C. Li, J. Guo, and C. Guo, ‘‘Emerging from water: Underwater image color correction based on weakly supervised colro transfer,’’ IEEE Signal Process. Lett., vol. 25, no. 3, pp. 323-327, 2018.
  54. 54.Y. Zhu, T. Par, P. Isola, et al., ‘‘Unpaired image-to-image translation using cycle-coinsistent adversarial networks,’’ in Proc. of IEEE Int. Conf. Comput. Vis.(ICCV), 2017, pp. 2242-2251.
  55. 55.Y. Guo, H. Li, and P. Zhuang, ‘‘Underwater image enhancement using a multiscale dense generative adversarial network,’’ IEEE J. Oceanic. Eng., pp. 1-9, 2019.
  56. 56.Z. Wang, A. Bovik, H. Sheikh, et al., ‘‘Image quality assessment: From error visibility to structural similarity,’’ IEEE Trans. Image Process., vol. 13, no. 4, pp. 600-612, 2004.
  57. 57.N. Hautiere, J. Tarel, D. Aubert, et al., ‘‘Blind contrast enhancement assessment by gradient ratioing at visible edges,’’ Image Anal. Stereo., vol. 27, no. 2, pp. 87-95, 2011.
  58. 58.T. Aydin, R. Mantiuk, K. Myszkowski, et al., ‘‘Dynamic range indepedetn image quality assessment,’’ ACM Trans. Graphics, vol. 27, no. 3, pp. 69-79, 2008.
  59. 59.M. Yang and A. Sowmya, ‘‘An underwater color image quality evaluation metric,’’ IEEE Trans. Image Process., vol. 24, no. 12, pp. 6062-6071, 2015.
  60. 60.K. Panetta, C. Gao, and S. Agaian, ‘‘Human-visual-system-inspired underwater image quality measures,’’ IEEE J. Ocean Eng., vol. 41, no. 3, pp. 541-551, 2016.
  61. 61.J. Xiao, J. Hays, K. Ehinger, et al., ‘‘Sun database: Large-scale scene recgonition from abbey to zoo,’’ in Proc. of IEEE Int. Conf. Comput. Vis. Pattern Rec. (CVPR), 2010, pp. 3485-3492.
  62. 62.D. Berman, D. Levy, S. Avidan, et al., ‘‘Underwater single image color restoration using haze-lines and a new quantitative dataset,’’ in arXiv preprint arXiv:1811.01343, 2018.
  63. 63.A. Duarte, F. Codevilla, J. Gaya, et al., ‘‘A dataset to evaluate underwater image restoration methods,’’ in Proc. of IEEE Int. Conf. Oceans, 2016, pp. 1-6.
  64. 64.H. Blasinski, T. Lian, and E. Farrell, ‘‘Underwater image systems simulation,’’ in Imaging and Applied Optics, 2017, pp. 1-2.
  65. 65.H. Blasinski and E. Farrell, ‘‘A three parameter underwater image formation model,’’ Electron. Imag., vol. 18, pp. 1-8, 2016.
  66. 66.W. Ren, S. Liu, H. Zhang, et al., ‘‘Single image dehazing via multi-scale convolutional neural networks,’’ in Proc. of Eur. Conf. Comput. Vis. (ECCV), 2016, pp. 154-169.
  67. 67.K. Zuiderveld, ‘‘Contrast limited adaptive histogram equalization,’’ Graphics Gems, pp. 474-485, 1994.
  68. 68.O. Ronneberger, P. Fischer, and T. Brox, ‘‘U-Net: Convolutional networks for biomedical image segmentation,’’ in Proc. of Med. Image Comput. Comput. Assit. Inter. (MICCAI), 2015, pp. 234-241.
  69. 69.K. He, X. Zhang, S. Ren, et al., ‘‘Deep residual learning for image recognition,’’ in Proc. of IEEE Int. Conf. Comput. Vis. Pattern Rec. (CVPR), 2016, pp. 770-778.
  70. 70.H. Chen and Y. Li, ‘‘Progressively complementarity-aware fusion network for RGB-D salient object detection,’’ in Proc. of IEEE Int. Conf. Comput. Vis. Pattern Rec. (CVPR), 2018, pp. 3051-3060.
  71. 71.C. Dong, C. Loy, K. He, et al., ‘‘Learning a deep convolutional network for image super-resolution,’’ in Proc. of Eur. Conf. Comput. Vis. (ECCV), 2014, pp. 184-199.
  72. 72.J. Johnson, A. Alahi, and L. Fei-Fei, ‘‘Perceptual losses for real-time style transfer and super-resolution,’’ in Proc. of Eur. Conf. Comput. Vis. (ECCV), 2016, pp. 694-711.
  73. 73.Q. Chen and V. Koltun, ‘‘Photographic image synthesis with cascaded refinement networks,’’ in Proc. of IEEE Int. Conf. Comput. Vis. (ICCV), 2017, pp. 1511-1520.
  74. 74.K. Simonyan and A. Zisserman, ‘‘Very deep convolutional networks for large-scale image recognition,’’ in Proc. of Int. Conf. Machine Learn. (ICML), 2015, pp. 1-14.
  75. 75.J. Deng, W. Dong, R. Socher, et al., ‘‘Imagenet: A large-scale hierarchical image database,’’ in Proc. of IEEE Int. Conf. Comput. Vis. Pattern Rec. (CVPR), 2009, pp. 248-255.

Citation

MLA
Li, C., et al. “An Underwater Image Enhancement Benchmark Dataset and Beyond”. IEEE TRANSACTIONS ON IMAGE PROCESSING 2019, 2019, http://arxiv.org/abs/1901.05495v2.
APA
Li, C., Guo, C., Ren, W., Cong, R., Hou, J., Kwong, S., & Tao, D. (2019). An Underwater Image Enhancement Benchmark Dataset and Beyond. IEEE TRANSACTIONS ON IMAGE PROCESSING 2019. http://arxiv.org/abs/1901.05495v2
Chicago
Li, C., C. Guo, W. Ren, et al. 2019. “An Underwater Image Enhancement Benchmark Dataset and Beyond”. IEEE TRANSACTIONS ON IMAGE PROCESSING 2019. http://arxiv.org/abs/1901.05495v2.
Harvard
Li, C. et al. (2019) “An Underwater Image Enhancement Benchmark Dataset and Beyond”, IEEE TRANSACTIONS ON IMAGE PROCESSING 2019 [Preprint]. Available at: http://arxiv.org/abs/1901.05495v2.
Vancouver
1. Li C, Guo C, Ren W, Cong R, Hou J, Kwong S, Tao D (2019) An Underwater Image Enhancement Benchmark Dataset and Beyond. IEEE TRANSACTIONS ON IMAGE PROCESSING 2019

BibTeX

@article{li2019underwater,
  title = {An Underwater Image Enhancement Benchmark Dataset and Beyond},
  author = {Li, Chongyi and Guo, Chunle and Ren, Wenqi and Cong, Runmin and Hou, Junhui and Kwong, Sam and Tao, Dacheng},
  year = {2019},
  journal = {IEEE TRANSACTIONS ON IMAGE PROCESSING 2019},
  url = {http://arxiv.org/abs/1901.05495v2},
  eprint = {1901.05495}
}
Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF