[论文解读] Deep Underwater Image Enhancement

Saeed Anwar, Chongyi Li+1 more|arXiv (Cornell University)|Jul 10, 2018

Image Enhancement Techniques参考文献 34被引用 125

一句话总结

一个基于数据驱动的CNN（UWCNN）通过学习降级图像与清晰图像之间的残差来提升水下图像，使用合成水下数据训练多种模型以适应不同水型，并加入后处理步骤（UWCNN+）以改善对比度。它在合成和真实水下图像上均优于现有方法。

ABSTRACT

In an underwater scene, wavelength-dependent light absorption and scattering degrade the visibility of images, causing low contrast and distorted color casts. To address this problem, we propose a convolutional neural network based image enhancement model, i.e., UWCNN, which is trained efficiently using a synthetic underwater image database. Unlike the existing works that require the parameters of underwater imaging model estimation or impose inflexible frameworks applicable only for specific scenes, our model directly reconstructs the clear latent underwater image by leveraging on an automatic end-to-end and data-driven training mechanism. Compliant with underwater imaging models and optical properties of underwater scenes, we first synthesize ten different marine image databases. Then, we separately train multiple UWCNN models for each underwater image formation type. Experimental results on real-world and synthetic underwater images demonstrate that the presented method generalizes well on different underwater scenes and outperforms the existing methods both qualitatively and quantitatively. Besides, we conduct an ablation study to demonstrate the effect of each component in our network.

研究动机与目标

• 在波长相关的衰减和散射下实现鲁棒的水下图像增强。
• 提出一个数据驱动的端到端CNN（UWCNN），直接从降级的水下图像重建潜在的清晰图像。
• 综合合成的水下数据集以训练适用于不同水型和退化水平的模型。
• 证明UWCNN对真实世界和合成图像具有泛化能力，并通过消融研究分析各组件的贡献。

提出的方法

• 将水下图像增强表述为图像修复任务，直接预测潜在图像，而不进行显式的B和T估计。
• 使用残差学习方法，网络学习降级图像与清晰图像之间的差异并将其加回输入（I = U + Δ(U, θ)）。
• 采用三块和密集连接的全卷积架构，端到端训练，不使用池化或批量归一化。
• 训练对应十种水下图像形成类型的十个模型，使用 NYU-v2 RGB-D 的合成数据并对不同水型建模衰减系数。
• 将 MSE（重建误差）与 SSIM 作为损失（L = L_MSE + L_SSIM），并用 ADAM 优化（学习率 0.0002）。
• 包含一个后处理步骤（UWCNN+），在 HSV/HSI 空间归一化饱和度和强度以提升真实图像的对比度。

实验结果

研究问题

• RQ1 Can a data-driven CNN trained on synthetic underwater data generalize to real underwater scenes with varying color casts and visibility?
• RQ2 Does residual learning with a densely connected architecture improve restoration quality over direct mapping approaches?
• RQ3 How do synthesized water-type datasets impact robustness across open-ocean and coastal turbidities?
• RQ4 What is the impact of adding a post-processing stage on perceptual quality and contrast in real underwater images?

主要发现

Types	MSE_RAW	MSE_RED	MSE_UDCP	MSE_ODM	MSE_UIBLA	MSE_UWCNN	PSNR_RAW	PSNR_RED	PSNR_UDCP	PSNR_ODM	PSNR_UIBLA	PSNR_UWCNN	SSIM_RAW	SSIM_RED	SSIM_UDCP	SSIM_ODM	SSIM_UIBLA	SSIM_UWCNN
Type-1	2367.3	3489.7	2062.3	2508.6	2812.6	587.70	15.535	15.596	15.757	16.085	15.079	21.790	0.7065	0.7406	0.7629	0.7240	0.6957	0.8558
Type-3	2676.5	4953.2	3380.6	3130.1	3490.1	747.50	14.688	12.789	14.474	14.282	13.442	20.251	0.5788	0.6639	0.6614	0.6765	0.5765	0.7951
Type-5	4851.2	8385.8	6708.9	3488.9	4563.7	1295.1	12.142	11.123	10.862	14.123	12.611	17.517	0.4219	0.5934	0.4269	0.6441	0.4748	0.7266
Type-7	7381.1	9809.8	8591.6	5337.1	6737.9	2974.1	10.171	9.991	9.467	12.266	10.753	14.219	0.1794	0.3192	0.1624	0.4178	0.2202	0.4920
Type-9	9060.6	5952.3	9500.1	10634.0	8433.1	4121.5	9.502	11.620	9.317	9.302	10.090	13.232	0.1794	0.3192	0.1624	0.4178	0.2202	0.4920
Water-Type I	1449.0	936.9	1020.7	1272.0	1492.2	209.70	17.356	19.545	18.816	18.095	17.488	25.927	0.8621	0.8816	0.8264	0.8172	0.7449	0.9376
Water-Type II	941.9	851.3	1466.0	1401.9	1141.4	251.60	20.595	20.791	17.204	17.610	18.064	24.817	0.8716	0.8837	0.8387	0.8251	0.8017	0.9236
Water-Type III	1851.0	2240.0	2337.6	1701.1	1697.8	456.40	16.556	16.690	14.924	16.710	17.100	22.633	0.7526	0.7911	0.7587	0.7546	0.7655	0.8795

• UWCNN achieves superior quantitative metrics (MSE, PSNR, SSIM) across ten water-type datasets compared to RAW, RED, UDCP, ODM, and UIBLA baselines.
• On synthetic test sets, UWCNN variants substantially reduce MSE and increase PSNR and SSIM relative to competing methods (e.g., Type-1: MSE 587.70 vs higher baselines; PSNR 21.790 for UWCNN vs ~15–16 for others; SSIM 0.8558 vs ~0.70–0.76).
• UWCNN+ (post-processed version) further improves colorfulness and contrast on real images without introducing color artifacts.
• Qualitative results show UWCNN removing greenish tones and recovering natural textures, outperforming ODM and UIBLA on real-world images.
• The paper provides an ablation study indicating the contribution of residual learning, dense concatenation, and end-to-end training to overall performance.