Classification of underwater pipeline events using deep convolutional neural networks

Abstract

Automatic inspection of underwater pipelines has been a task of growing importance for the detection of four different types of events: inner coating exposure, presence of algae, flanges and concrete blankets. Such inspections might benefit of machine learning techniques in order to accurately classify such occurrences. In this work, we present a deep convolutional neural network algorithm for the classification of underwater pipeline events. The neural network architecture and parameters that result in optimal classifier performance are selected. The convolutional neural network technique outperforms the perceptron algorithm preceded by wavelet feature extraction for different event classes, reaching on average 93.2% classification accuracy, while the accuracy achieved by the perceptron is 91.2%. Besides the results obtained in the test set, accuracy and cross entropy curves obtained in the validation set during training are analyzed, so that the performances of each method and for each event class are compared. Visualizations of the convolutional neural network intermediate layer outputs are also provided. These visualizations are interpreted and associated to the results obtained.