[Paper Review] UWGAN: Underwater GAN for Real-world Underwater Color Restoration and Dehazing

Nan Wang, Yabin Zhou+3 more|arXiv (Cornell University)|Dec 21, 2019

Image Enhancement Techniques16 references107 citations

TL;DR

UWGAN introduces an unsupervised GAN to simulate real underwater effects from in-air images and depth maps and uses a U-Net autoencoder for end-to-end underwater color restoration and dehazing, achieving up to 125 FPS on a single NVIDIA 1060 GPU.

ABSTRACT

In real-world underwater environment, exploration of seabed resources, underwater archaeology, and underwater fishing rely on a variety of sensors, vision sensor is the most important one due to its high information content, non-intrusive, and passive nature. However, wavelength-dependent light attenuation and back-scattering result in color distortion and haze effect, which degrade the visibility of images. To address this problem, firstly, we proposed an unsupervised generative adversarial network (GAN) for generating realistic underwater images (color distortion and haze effect) from in-air image and depth map pairs based on improved underwater imaging model. Secondly, U-Net, which is trained efficiently using synthetic underwater dataset, is adopted for color restoration and dehazing. Our model directly reconstructs underwater clear images using end-to-end autoencoder networks, while maintaining scene content structural similarity. The results obtained by our method were compared with existing methods qualitatively and quantitatively. Experimental results obtained by the proposed model demonstrate well performance on open real-world underwater datasets, and the processing speed can reach up to 125FPS running on one NVIDIA 1060 GPU. Source code, sample datasets are made publicly available at https://github.com/infrontofme/UWGAN_UIE.

Motivation & Objective

• Motivate real-world underwater image restoration and dehazing using a data-efficient, unsupervised GAN framework.
• Propose an improved underwater imaging model to generate realistic underwater images from in-air image and depth map pairs.
• Adopt a U-Net based color restoration and dehazing network trained on synthetic underwater data with preserved scene structure.
• Demonstrate qualitative and quantitative improvements over existing methods on real underwater datasets.
• Provide publicly available code and sample datasets to facilitate further research.

Proposed method

• Propose an unsupervised GAN that generates realistic underwater color distortion and haze from in-air images and depth maps using an improved underwater imaging model.
• Use a U-Net architecture to perform color restoration and dehazing in an end-to-end autoencoder framework.
• Train efficiently on synthetic underwater datasets to preserve scene content and structural similarity.
• Demonstrate real-world applicability by evaluating on open underwater datasets and reporting processing speed (up to 125 FPS on a NVIDIA 1060 GPU).
• Release source code and sample datasets publicly for reproducibility.

Experimental results

Research questions

• RQ1 Can an unsupervised GAN generate realistic underwater distortion and haze from in-air images and depth maps using an improved imaging model?
• RQ2 Can a U-Net based autoencoder effectively restore color and dehaze underwater scenes while preserving structural content?
• RQ3 How does the proposed UWGAN perform on real-world underwater datasets in qualitative and quantitative comparisons?
• RQ4 What is the practical processing speed of UWGAN on common hardware (e.g., NVIDIA GPUs) for real-time applications?

Key findings

• The proposed unsupervised GAN can generate realistic underwater color distortion and haze from input in-air images and depth maps.
• The U-Net restoration network preserves scene structure while performing color restoration and dehazing.
• Qualitative and quantitative evaluations show competitive performance on real-world underwater datasets.
• The method achieves processing speeds up to 125 FPS on a single NVIDIA 1060 GPU.
• Source code and sample datasets are publicly available for reproducibility.