Hand-Crafted Feature Based Classification against Convolutional Neural Networks for False Alarm Reduction on Active Diver Detection Sonar Data

This paper provides a method for false alarm reduction on active diver detection sonar data by means of contact classification. Two different classification techniques are compared, on the one hand a classical hand-crafted feature based method and on the other hand a deep learning approach which inherently includes an automated feature extraction. For the classical method 53 hand-crafted features are extracted from the sonar contacts and used as inputs for a feed forward neural network (FNN). The inputs for the convolutional neural networks (CNNs) are two-dimensional sonar images (level vs. bearing and time/range) of the contacts. The performance of the classifiers is compared to that of the standard active signal processing by means of Receiver-Operating-Characteristic (ROC) curves. It is shown that both classification techniques outperform the standard active signal processing in which the detector threshold represents the only adjustable parameter.