Abstract

Computational vision is the ability of a machine to process images or videos. Many technologies are used to develop this ability. Deep learning is a technique inside the machine learning area, where a neural network is used but with a more profound complexity than the usual neural networks. Deep learning is changing the scenario for computational vision since it is more efficient than the others and is becoming more popular. Given this popularity, deep learning for computational vision is applied for multiple fields of study. In the maritime industry, this technique is being used by different researchers to help them identify and classify ships automatically through images and videos. Also, it is being applied in subsea inspections to simplify the task of detecting objects in the seabed. This paper reviews the current state of the art of using deep learning techniques for computational vision within the Maritime Industry. The target for each study is analyzed. A comparison of the dataset used and the type of deep learning are made. The most common target is to identify ships using surveillance cameras or satellite images. Subsea equipment, corals, underwater mines, marine organisms, and oil spills are other targets. The most used deep learning technique is the convolutional networks. This result is not only observed in the maritime domain but for any computational vision problem. The neighbors' data influence the convolutional network result, and the pixel of an image is very similar to its neighbor pixel, evidencing its advantage. There are at least ten different types of convolutional networks being used in the reviewed papers, making it clear that there is space for innovation in this matter. This review and comparison can help future research, giving information on which deep learning technique to choose and how to evaluate its target.