Data driven underwater transient detection based on time-frequency distributions

The complexity of real-life transients, coupled with the incomplete (or absent) knowledge of their statistical structure or defining features has motivated the interest on the use of blind, data driven detection schemes. One such scheme, proposed by Jones and Sayeed (1995), uses time-frequency distributions to implement sub-optimal quadratic detectors which, under certain conditions, approach the performance of optimal quadratic detectors. However, their use of Fisher's discriminants to obtain class separation has some drawbacks, which we solve by using a simple perceptron to obtain the discriminant. Also, more often than not, we will have a multiclass situation, implying the use of different time-frequency distributions, each one of them tuned for a given class of transients. The different nature of these distributions (bias, type of cross-terms, time-frequency resolution, etc.) will hamper the performance of the algorithm, forcing the need for experimental validation of its heuristical aspects. The algorithm will be applied to real data, and its performance investigated.