Recognition of underwater transient patterns

摘要

Underwater transients present an unusual but unique recognition problem in that pattern characteristics and categories are not well defined, and the patterns are highly nonstationary. This paper examines the characterization, segmentation and classification of the transient events. A comparison with speech processing and recognition is also made. Although it is difficult to fully characterize the transient data, it is shown that by extracting event portions of the transient waveform through segmentation, a low order autoregressive model can provide an effective feature set for cluster analysis and event classification. Both the segmentation procedure and the recognition experiments with real data are presented in detail.