Classification of Underwater Objects Via Impulse Excitation

A technique for classifying objects based on modeling the transient characteristics of their impulse response is developed and tested. A set of targets identical in geometry and differing in shell and filler material were constructed. The targets were manually struck exciting an impulse response which was sampled and recorded. The impulse response of each target was decomposed via windowed short-time Fourier transform into a set of feature vectors. The feature vectors were quantized via the LBG VQ algorithm, and the sets of quantized vectors were used to estimate the parameters of a discrete-output hidden Markov model (HMM) for each class of object. A blind test set was evaluated against the trained HMMs and the results are presented along with a discussion of the generalization ability of the individual classifiers