Automatic feature extraction from wavelet coefficients using genetic algorithms
Shubhankar Ray, Albert Yick-Po Chan · 2002
Deciding what features can be effective for a signal classification problem is often a nontrivial task. We present a method that can be used for automatic extraction of high quality features from wavelet coefficients without a priori knowledge of features. Preprocessing of the wavelet coefficients is necessary to obtain a measurable set of features. The preprocessing is suitable for the Morlet wavelet. Genetic algorithms are used in combination with learning vector quantization neural networks to select the relevant features from the processed wavelet coefficients. A simple variation of the traditional feature selection genetic algorithms is used as it applies to this method. The method has been applied on different signals for classification and has shown high classification rates with a small number of features. Results from different signal classification problems are also presented.