When underwater imagery analysis meets deep learning: A solution at the age of big visual data

Abstract

Underwater imagery processing is in great demand, while the research is far from enough. The unrestricted natural environment makes it a challenging task. On the other hand, prior to the advent of cabled observatories, the majority of deep-sea video data was acquired by remotely operated vehicles (ROVs), and was analyzed and annotated manually. In contrast, seafloor cabled observatories such as the NEPTUNE and VENUS observatories offer a 24/7 presence, resulting in unprecedented volumes of visual data. The analysis of underwater imagery imposes a series of unique challenges, which need to be tackled by the computer vision community in collaboration with biologists and ocean scientists. In this paper, we introduce how deep learning, the state-of-the-art machine learning technique, can benefit underwater imagery understanding at the age of big data.