UWGAN: Underwater GAN for Real-world Underwater Color Restoration and Dehazing

1 Introduction

• Conventional enhancement algorithms do not model underwater imaging physics, making direct application to different underwater scenes challenging and ineffective.Examples include white balance, gray world, histogram equalization, and fusion methods.Source: §1 · p. 1
• Imaging-model-based traditional methods can be time-consuming because they calculate image characteristics.Source: §1 · p. 1
• CNN-based underwater enhancement is limited by scarce underwater datasets and the difficulty of obtaining clear images without water for real underwater scenes.The passage identifies synthetic underwater datasets as an important approach.Source: §1 · p. 1
• The proposed pipeline uses an improved underwater imaging model and GAN to synthesize realistic degraded images, then trains U-Net on synthetic data for enhancement.The synthetic images include color cast and haze effects.Source: §1 · p. 1
• The authors report recovery of underwater images while maintaining structural similarity, and compare U-Net loss functions to identify a suitable restoration loss.The method is validated on real underwater images and target-detection datasets for low- and high-level vision tasks.Source: §1 · p. 2

2 Our Proposed Method

• Synthetic underwater images train a U-Net restoration network, while UWGAN learns to synthesize underwater appearances from in-air RGB-D inputs and site-specific underwater samples.The synthetic pairs support real-time enhancement.Source: §2 · p. 2

2.1 Improved underwater imaging model

• The improved model adds a scattering-based haze term to wavelength-dependent attenuation, with α set to 1 for moderate, homogeneous haze.Figure 1 shows in-air samples and synthetic images for different water types.Source: §2.1 · p. 2, Figure 1 · p. 2

2.2 UWGAN for generating realistic underwater images

• UWGAN takes a color image and depth map, then learns through generative adversarial training to synthesize realistic underwater images using the underwater optical imaging model.Source: Figure 2 · p. 3

2.3 Underwater image restoration based on U-Net

• The U-Net encoder downsamples a degraded RGB image to a 32x32x256 latent vector, and its decoder upsamples while concatenating corresponding encoder features.The output feature maps return to three channels.Source: §2.3 · p. 3, Figure 3 · p. 3

2.4 Loss Functions

• The study compares L2, L1, SSIM, MS-SSIM, and gradient losses, including L1-based combinations, for underwater restoration.The gradient loss compares gradients of reconstructed and clear images; α is set to 0.8 after experiments.Source: §2.4 · pp. 3–4, Table 1 · p. 4

2.5 Dataset

• The in-air datasets contain 3,733 NYU Depth RGB images with corresponding depth maps; underwater RealA and RealB contain 2,069 and 2,173 images, while RealC is an 890-image testing set.RealA contains near-field green-hued images, while RealB contains blue-green-hued far-field scenes.Source: §2.5 · p. 4

3 Experimental Setup

• The reported training computer used an Intel i7 HQ 8700 processor, 16GB RAM, and an NVIDIA 1080Ti 12GB graphics card.Source: §3 · pp. 4–5
• UWGAN synthesis training ran for 30 epochs with Adam, a 0.0001 learning rate, momentum 0.5, batch size 64, and 256x256 outputs.Source: §3 · p. 5
• The U-Net enhancement network trained on synthetic pairs for 200 epochs with Adam, a 0.0001 learning rate, batch size 32, and 256x256 outputs.Source: §3 · p. 5

4 Result and Discussion

• Real-world assessment uses non-reference UIQM because ground-truth scenes are unavailable, while synthetic assessment uses MSE, PSNR, and SSIM.Source: §4 · p. 6
• Across RealA, RealB, and RealC, UWGAN recovers color while maintaining proper brightness and contrast, whereas UGAN and WaterGAN leave color errors and artifacts that damage structure.Source: Figure 5 · p. 5, §4 · p. 6, Figure 6 · p. 6
• UWGAN has the fastest processing speed among compared methods and the fewest parameters and FLOPs among deep-learning methods.The reported timing configuration uses an Intel i7-8750H CPU, 16GB RAM, and GTX1060 6G GPU.Source: §4 · p. 8, Table 6 · p. 8
• On synthetic data, UWGAN achieves the best MSE, PSNR, and SSIM results; on RealA, RealB, and RealC, it achieves the best color-restoration scores.Table 5 reports the synthetic full-reference comparison; Table 4 summarizes average scores on the real-world datasets.Source: §4 · p. 7, Table 4 · p. 7, Table 5 · p. 8
• YOLO v3 detection performs better on model-enhanced than degraded images; Figure 7 marks scallops red, sea cucumbers blue, and sea urchins green.Source: §4 · p. 8, Figure 7 · p. 8
• Ablations show L1/L2 produce stripe artifacts, SSIM/MSSSIM uneven color, and GDL natural but undersaturated output; combined losses score better on synthetic metrics.The visual comparison includes RealC samples, while MSE, PSNR, and SSIM are reported for the synthetic dataset.Source: §4 · p. 9, Table 7 · p. 9

• UWGAN uses an improved underwater imaging model to generate realistic underwater images.Source: §5 · p. 9
• U-Net with combined loss functions is used to enhance degraded underwater images.Source: §5 · p. 9
• Validation on low- and high-level underwater vision tasks demonstrates the model’s effectiveness and robustness.Source: §5 · p. 9
• An ablation study examines the effects of different loss functions on underwater image restoration and enhancement, supporting future research.Source: §5 · p. 9