Signal processing for target recognition in biosonar

Some properties of biological sonar systems are reviewed along with corresponding signal processing techniques. Relevant target representations appear to involve time-frequency or range-Doppler distributions and their projections as well as images derived from echoes observed at different positions. Target recognition can occur via association of echo features derived from such representations, e.g. by using spatially registered feature maps. Informative features can be obtained from sequential estimation of a vision-like acoustic image, i.e. a high-resolution representation of target reflectivity as a function of azimuth, elevation, and range. Such an image can be obtained with a rotated wavelet (line segment) transform, which is a generalization of wavelet and Radon transforms. The rotated wavelet transform permits recursive delay-and-sum beamforming with a sparsely sampled synthetic aperture constructed with a moving multibeam sonar system. The basis functions for the transform are rotated space-time transmission patterns similar to those used by echolocating dolphins.