Feature extraction based on subspace energy of local discriminant basis

For non-stationary and time-varying underwater sound signals,wavelet packet transform is used. The character istic of every wavelet packet basis is different,which can express the main feature of a signal. The local discriminant basis (LDB) can be calculated based on the distance criterion. A feature extraction method is proposed. The feature vector,which expresses the energy of sub-space in LDB,is obtained. Feature choice is done using Fisher criterion. A neural network target classifier is designed. And the classification experiment for three different classes of targets has been done. The results of experi-ment show that the feature extraction and choice method is useful.