Bispectral Gammatone Cepstral Coefficient based Neural Network Classifier

The estimation of the power spectrum of discrete-time signals is one of the most fundamental and useful tools in signal processing. However, there are practical situations where one needs to look beyond the power spectrum, especially to extract information regarding the phase relations and deviations from Gaussianity. This has created considerable interest in the use of higher order spectra such as bispectrum, for the analysis of signals, particularly in the presence of additive Gaussian noise. This paper examines the use of Gammatone Cepstral Coefficients computed from the spectrum reconstructed from the bispectrum of the signal as a feature set for underwater target classification. A prototype Neural Network classifier with back propagation algorithm has been trained with the proposed feature set and the performance has been evaluated, which has yielded acceptable classification results.