Thomas J. Hayward
Naval Res. Lab., Washington, DC 20375-5350
A multiresolution, likelihood-based statistical approach is presented for characterizing labeled classes of samples (e.g., measured time series or time-frequency distributions of acoustic transients) and for classifying new samples based on this statistical characterization. The labeled classes are characterized by a histogram associated with a multiresolution decomposition of the data in each class. Classification of a new sample is then performed by calculating, for each labeled class, the conditional probability of the sample given the statistics of that class. These conditional probabilities, which are interpreted as relative likelihoods that the sample belongs to each of the classes, are calculated in a recursive computation that proceeds from coarse to fine resolution. A simple, efficient computer implementation using associative arrays is presented. Successful classification of both time series and time-frequency distributions of marine-mammal vocalizations is demonstrated using relatively small numbers of labeled samples (~10 per class). [Work supported by ONR.]