Enhancing Underwater Imagery using Generative Adversarial Networks (1801.04011v1)

Abstract: Autonomous underwater vehicles (AUVs) rely on a variety of sensors - acoustic, inertial and visual - for intelligent decision making. Due to its non-intrusive, passive nature, and high information content, vision is an attractive sensing modality, particularly at shallower depths. However, factors such as light refraction and absorption, suspended particles in the water, and color distortion affect the quality of visual data, resulting in noisy and distorted images. AUVs that rely on visual sensing thus face difficult challenges, and consequently exhibit poor performance on vision-driven tasks. This paper proposes a method to improve the quality of visual underwater scenes using Generative Adversarial Networks (GANs), with the goal of improving input to vision-driven behaviors further down the autonomy pipeline. Furthermore, we show how recently proposed methods are able to generate a dataset for the purpose of such underwater image restoration. For any visually-guided underwater robots, this improvement can result in increased safety and reliability through robust visual perception. To that effect, we present quantitative and qualitative data which demonstrates that images corrected through the proposed approach generate more visually appealing images, and also provide increased accuracy for a diver tracking algorithm.

• The paper introduces UGAN, a GAN-based restoration model that uses CycleGAN to create paired datasets for underwater image correction.
• It employs a WGAN-GP framework to stabilize training and reduce mode collapse by minimizing both Wasserstein and L1 losses.
• Enhanced imagery led to a 350% improvement in diver tracking, illustrating significant practical benefits for autonomous underwater vehicles.

Enhancing Underwater Imagery using Generative Adversarial Networks

This paper tackles the complex challenge of improving the quality of underwater images for autonomous underwater vehicles (AUVs) using Generative Adversarial Networks (GANs). Underwater images are typically degraded by various factors such as light absorption, scattering by suspended particles, and color distortion, all of which impact the efficacy of vision-driven tasks performed by AUVs, such as tracking and classification.

Methodology

The authors propose a solution using CycleGAN, an unpaired image-to-image translation model, to generate a paired dataset for training a GAN-based restoration model, referred to as Underwater GAN (UGAN). CycleGAN is leveraged to create a mapping from undistorted underwater images to distorted ones, thus simulating realistic underwater conditions. This simulated dataset serves as the basis for training UGAN to convert distorted back to ‘corrected’ underwater images.

Adversarial Framework

UGAN employs a WGAN-GP (Wasserstein GAN with Gradient Penalty) framework to ensure stable training and reduce the risk of mode collapse common in GANs. The generator within this framework aims to improve visual fidelity by minimizing both Wasserstein and L1 losses, creating plausibly restored imagery without the need for paired datasets.

Results and Implications

The proposed method shows significant improvements in visual quality and reduced noise in comparison to traditional methods like CycleGAN, demonstrated through edge detection and localized patch analysis across images. Figures depicting qualitative results reveal the network's ability to retain or restore color richness and clarity in underwater scenes. The application of the UGAN outputs also demonstrates a pronounced enhancement in a diver tracking algorithm, with a 350% increase in correct positive detections over unaltered imagery. This suggests robust improvements in algorithmic tasks reliant on visual data integrity.

Future Directions

The paper hints at several future directions, including expanding the dataset to cover a broader range of underwater conditions and introducing additional elements such as particle and lighting effects to enhance robustness. By doing so, the network's adaptability across diverse underwater environments could improve.

Conclusion

Overall, this research presents a compelling advancement in the field of underwater image enhancement. By leveraging GANs and unpaired image datasets, the authors address critical challenges inherent to underwater robotics and highlight the potential for widespread application in deep-sea exploration and other domains reliant on underwater visual data. The paper’s implications extend beyond immediate practical improvements, suggesting a viable path toward more capable and reliable AUV operations in visually degraded environments.