A Visual Inspection Proposal to Identify Corrosion Levels in Marine Vessels Using a Deep Neural Network

• Main Authors: Soares, Luciane, Botelho, Silvia, Nagel, Ricardo, Drews, Paulo Lilles
• Format: Conference Proceeding
• Language: English
• Subjects: Convolutional neural networks; Corrosion; Degradation; Inspection; Software; Unmanned vehicles; Visualization
• ISSN: 2643-685X
• DOI: 10.1109/LARS/SBR/WRE54079.2021.9605400

The increase in oil production in Brazil, consequently, resulted in the concern with the integrity of the platforms, with the objective of avoiding biological and human society disasters. Most inspections are carried out by divers or performing the practice of docking the platform, however, a docking requires the emptying of the platform's tanks, causing it to stop producing what is economically unfavorable. The use of remotely operated vehicles (ROVs) is increasingly common. The operator is in a safe environment while the ROV is able to perform a detailed pass of videos and images of the hulls of the floating units. Because with the action of the sea and the installation time, the tendency is that these units suffer degradation over the years, corrosion is one of the most common degradations. Thus, the main contribution of this work is to present a convolutional neural network that classifies underwater images into four corrosion levels (high, medium, low and no corrosion), with the goal of embedding classification software in an ROV in the future. It is also presented a synthetic dataset produced, simulating turbidity, through a function of degradation and alteration of the variable gamma of the images.