ACOUSTIC-COLOR-BASED CONVOLUTIONAL NEURAL NETWORKS FOR UXO CLASSIFICATION WITH LOW-FREQUENCY SONAR

In this work, we contribute a new target classification approach for low-frequency sonar data. More specifically, we illustrate the feasibility of using convolutional neural networks (CNNs) trained on acoustic-color data, a representation that expresses target strength as a function of object aspect and frequency. We show that it is possible, using only limited amounts of this sonar data, to design and train efficient networks with low capacity that avoid overfitting and generalize robustly, even to new objects not seen during training. We demonstrate this in the context of an unexploded ordnance (UXO) classification task using real, measured sonar data collected at sea.