Deep Learning for Clutter Reduction in Multi-static Coherent Active Sonar Systems

Enhanced contact screening for MAC missions will be accomplished by adapting deep learning techniques developed by researchers for image processing to perform pattern recognition on the Bayesian target state estimator for MAC. The state estimate, a probability mass function encapsulating the existence of a target and its potential position and velocity, contains 225,000 data points for every ping. This information will serve as the input for a deep convolutional neural network which will label regions of the field as target, bottom clutter, ephemeral clutter, or empty. This labeled map will be used to reduce clutter by improving snippet scores, and increasing the speed with which persistent clutter is identified by MAC. Other improvements include an enhanced operator display and better ping control algorithms. Anticipated Benefits/Commercialization: clutter suppression, enhanced operator displays, improved ping control for Multistatic sonar with applications to image, speech and hyperspectral data.