Exploiting Auxiliary Information for Improved Underwater Target Classification with Convolutional Neural Networks

This work deals with the classification of objects as targets or clutter in synthetic aperture sonar (SAS) imagery using convolutional neural networks (CNNs). First, a new image-annotation tool is developed that allows extra auxiliary information (beyond the basic binary label) to be easily recorded about a given input image. The additional information consists of an estimate of the image quality; the local background environment; and for targets, the specific object shape, orientation, and length. The architecture of the CNNs - specifically the final dense layer and output layer - is then modified so that these extra quantities are additional outputs to be predicted simultaneously. As such, the task of the augmented CNNs becomes to provide a richer representation of an image beyond the binary label. This more complete operational picture can then better inform subsequent mine countermeasures (MCM) decisions. Experiments on a set of real, measured SAS data collected at sea demonstrate that tiny CNNs can accurately predict the additional auxiliary qualities without suffering a significant drop in binary classification performance.