MODIFICATION OF KARHUNEN-LOEVE TRANSFORM FOR PATTERN-RECOGNITION

Application of the Karhunen-Loeve transform in pattern recognition pro blems necessitates knowledge of the mean vectors for the training sets of different classes for construction of the respective covariance ma trices. This modality poses problems for recognizing an unknown signal in a multi-class environment and for its on-line implementation. This is because a priori it is unknown which mean vector of the several cl asses is to be subtracted from an unknown input signal. To remove this difficulty a global mean approach has been proposed. The proposed met hod has been applied to synthetic and experimentally observed acoustic signals successfully and its performance with respect to pattern reco gnition and data compression has been compared with the conventional m ethod.