Nonstationary signal classification using time-frequency optimization

We explore in this paper the use of pairwise Fisher criterion and weighted pairwise Fisher criterion as the objective functions for time-frequency based classification. The approach uses optimisation algorithms to alter and test the time-frequency kernel parameters based on the Fisher criterion objective function. For parameterised time-frequency representations (TFRs) kernels the determination of the optimal kernel parameters reduces to a maximization of the objective function. The classification process is based on joint optimization of parametric TFRs and distance measures. The optimal parameters realized from the classifier training and testing are used to classify novel whale songs. A classification error rate of 6.6% was achieved with the minimum distance classifier.