Interactive Sonar Feature Extraction.

Abstract : The problem considered is where n samples of data are to be analyzed and classified. A particular sample of the data is, in general, two dimensional. For example, a sample may consist of d digitized time functions, each waveform the output from a hydrophone transducer. If the hydrophones are arranged along a straight line which is called the space dimension, then a sonar sample may be considered a discretized space-time waveform. Our objective is to use the n samples to construct a mapping from the data observation space (the digitized data is assumed to consist of Ks data points were s indicates sample s) to what will be called the class space. There are M classes in the class space and M may be an unknown quantity. Mathematically, the M classes may be considered as points on the real line. Historically, statistical learning theory and pattern recognition have produced certain operations from which the above mapping from the observation space to the real line is constructed. Before considering these operations, lets try to construct a single mapping from the raw data space to the class space.