Pipeline tracking and event classification for an automatic inspection vision system

Automatic inspection of underwater pipelines has been a task of growing importance for the detection of a variety of events, such as inner coating exposure and twists, which might indicate risks for future leakages. Most often performed by autonomous underwater vehicles, the video inspections might benefit from image processing, estimation and machine learning techniques in order to accurately detect and classify such occurrences. This article describes algorithms designed to segment the pipelines and to classify some important events. In order to obtain a robust technique for pipeline border detection in a diversity of scenarios, several image processing techniques and tracking strategies are employed. A deep convolutional neural network algorithm and a wavelet-based multilayer perceptron are developed for the classification of four types of events. The convolutional neural network technique outperforms the perceptron algorithm, for different event classes and without requiring manual feature extraction, when a large amount of training samples is used.