Harnessing Gabor Wavelet Feature Localization with Diverse Machine Learning Techniques for Sonar Target Classification

We examine and quantify the impact of three Gabor wavelet orientations convolved across acoustic color magnitude spectra prior to a feature extraction and feature enumeration technique. Algorithm parameters are selected through performance metrics generated from a support vector machine trained and tested with a subset of the public domain field data. Once parameter selection is complete, three machine learning classifiers are trained and tested on a second subset of data. Comparisons between the three Gabor wavelet orientations are presented quantitatively via overall classification accuracy results and qualitatively through examination of the extracted features of a common target for all classifiers. The Gabor wavelet at two orientations were found to have an overall classification accuracy of 10% to 20% larger when compared to the remaining Gabor wavelet orientation. This superior classification accuracy is attributed to the compact feature localization when compared to the unfiltered and latter Gabor orientated acoustic color spectrograms.