Adaptive classification of ensonified underwater objects

The biological sonar system of bottlenose dolphins is adapted for cluttered, high noise, and extremely reverberant shallow-water environments such as bays, estuaries, and near-shore waterways. Echolocation behavior and signals are plastic, modified in-stride during encounters with novel targets and in novel surroundings. Moreover, biosonar target discrimination performance provides an existence proof for biological sonar-based recognition of various kinds of targets. For the dolphin this is accomplished in part by auditory neural computations that are not yet fully understood. A model of adaptive target classification processes was developed using heuristics based on dolphin psychophysics and echolocation behavior. Echoes were recorded while a dolphin ensonified several aspect-dependent targets during a target discrimination task. A computational model of dolphin ear filters was adapted to include filter gain recursively scaled to acoustical characteristics of local ambient noise and reverberation. Stochastic noise was attenuated by multi-ping fusion within an Integrator Gateway processing scheme. Output from the spectral filters served as inputs for a neural net target classifier. The system accommodated changes in local ambient noise and produced classification signficantly better than did traditional classifiers that lacked adaptive filtering processes. [Work funded by ONR 321US.]