This thesis investigates the application of wavelet decompositions to classification applications. Two feature extraction tools are considered: Local Discriminant Bases scheme (LDB) and Power method. Several dimension reduction schemes including a newly proposed one called the Mean Separator neural network (MS NN) are discussed. Two types of classifiers are investigated and compared: Classification Trees (CT) and Back-propagation neural network (BP NN). Classification experiments conducted on synthetic and real-world underwater signals show that: 1) the Power feature extraction method is more robust to time synchronization issues than the LDB scheme is; 2) the MS NN scheme is a successful dimension reduction scheme that may be used with both LDB and Power feature extraction methods; and 3) the BP NN is a more powerful classifier than CT as it has fewer constraints than CT in partitioning the feature input space.

Back to homepage.