Automatic feature-finding for time-frequency distributions
Les Atlas, L. Owsley, J. McLaughlin, Gary D. Bernard · 2002
Given the detailed time and frequency resolution of time-frequency distributions, trainable automatic classifiers can easily be overwhelmed by the complexity of this input representation. This problem becomes even more severe as more advanced and higher resolution time-frequency distributions come into use. Our research is directed to making a better match to automatic classification by automatically finding a set of lower-dimensionality features within time-frequency distributions. We show the efficacy and generality of this approach to a wide variety of time-frequency distributions. A connection is also made to hidden Markov model-based classification and a comparative study is shown for this type of classifier for conventional and more advanced proper time-frequency distributions. We conclude that, when used within the context of hidden Markov model-based classification, the proper time-frequency distribution offers the best ability to reserve classes representing changes in constituents of short acoustic transients. We have developed a vector quantization technique which is a modified version of Kohonen's (1990) self-organizing feature map and then applied it to conventional time-frequency representations (the magnitude of the short-time Fourier transform), more advanced time-frequency representations (the minimum cross-entropy (MCE) proper and positive distribution), and to a proper-distribution derived measure.