Feature selection, adaptive detection/classification, and beam forming for mine avoidance sonar

New techniques are proposed for clutter and multipath suppression, reliable adaptive detection/classification, and beam forming. Generalized (multivariate) prediction-subtraction is proposed for clutter reduction. Innovative blind deconvolution algorithms are proposed for multipath suppression. Adaptive maximum likelihood beam forming, beam deconvolution, and parametric sonar are considered for effective beam narrowing. Proposed detection/classification algorithms use "feature-grams," a generalization of time-frequency distributions that includes many range-varying parameter representations, such as incomplete synthetic aperture feature images from forward-looking echo data. A novel, totally automatic method discovers the best features for a reduced-dimensional feature space from observations of design set feature-grams. The resulting feature space data representation is more meaningful to a sonar operator than the usual principal component representation. New feature-grams and associated classification algorithms are demonstrated via single-echo classification of wideband, wide-beam sonar data from targets in clutter. Segmented feature-grams are proposed for extended hidden Markov models that adaptively classify single-echo or multi-echo data. An adaptive classifier learns to improve its performance using unlabeled echoes from mine-like objects and clutter that are not included in the original design set. Reliable adaptive feature extraction and classification are needed for real-world operation against environments and objects that cannot all be included in an initial design set.