Database of Submarine Risers for Corrosion Detection

Abstract

Developments in Machine Learning fundamentals have been paramount in product innovations and disruptive technologies. This progress lies in artificial neural networks and their applications, such as in computer vision, allowing autonomous systems to be able to identify complex visual patterns in a similar way to human perception, at the cost of heavy dependence and availability of extensive data sets. With this in mind, this work aims to build a unique set of annotated underwater data to be applied to the detection of external corrosion on risers. The methodology consisted of collecting images from frames of YouTube videos, standardizing and processing this data to classify images with damage and without damage, and storing them in the cloud. To achieve this goal, the YOLOv8 nano version architecture was used. During the training and validation stages, the model showed a rapid decrease in error rates and an increase in accuracy and recall, reflecting effective adaptation to the data. In addition, validation using a new data set confirmed the reliability of the detector in practical contexts, correctly identifying the 'corroded' and 'clear' states with high confidence. However, some overlap was observed in individual images, due to the sensitivity of the model.