Self-Supervised Underwater Image Generation for Underwater Domain Pre-Training

The rapid progress in computer vision has presented new opportunities for enhancing the visual capabilities of underwater robots. However, most deep learning-based visual perception algorithms often underperform due to the scarcity of underwater datasets. To address this issue, we propose an underwater image synthesis method for pre-training in the underwater domain. By leveraging self-supervised learning, we simulate the physical imaging process of underwater scenes, allowing for style transfer from in-air images to underwater images using a reduced amount of underwater data. Furthermore, we propose a pre-training strategy that utilizes synthetic underwater images to enhance underwater visual perception. Finally, abundant experiments are conducted, including quantitative and qualitative comparisons. The results validate the effectiveness and superiority of the proposed underwater image synthesis method, highlighting the substantial improvement in underwater environment perception achieved through the underwater domain pre-training (UDP) strategy.

I. Introduction

In recent years, underwater robots have been widely used in underwater measurement, manipulation, and observation, with the deepening of ocean exploration [1]. However, to achieve autonomous operation, a crucial requirement is to enhance the perception capabilities of underwater robots in the underwater environment. In the field of computer vision, most visual perception algorithms heavily rely on extensive datasets, especially with the emergence of deep learning-based methods in recent years. Unfortunately, due to the unique characteristics of the underwater environment, collecting underwater data is costly, resulting in a limited availability of effective underwater datasets. Consequently, visual perception of the underwater environment remains a challenging problem that requires further attention and research.