Automatic classification of acoustic sequences by multiresolution image processing and neural networks

The sounds emanating from whales and other marine mammals in the ocean offer a wide variety of acoustic signals which can be interpreted as both an image in generalized time-frequency space and a sequence of images over time. If one is interested in detecting and classifying these signals in the cluttered environment of the ocean, methods must be developed to characterize and categorize these images. The authors concentrate on initial results of exploiting the multiresolution nature of this problem. The concept of multi-dimensional wavelets is most significant as a characterization of the sequential evolution of the image features of these signals. The use of neural networks to classifying underwater acoustic waveforms is not new. The authors take a first step toward the development and application of multiresolution neural networks to this image processing and classification problem. The fundamental neural network will be a bank of three neural networks, each tuned to a different scale of time-frequency resolution. The representations and the networks provide a strong vision analogy to the zoom of visual/recognition acuity