Extraction of optimal time-frequency plane features for classification

A method based on Local Discriminant Bases is developed to extract discriminanting features for classification from time-frequency pattern of one dimensional signals. Acoustic signals from two classes are first divided into segments along the time axis according to their discrimination power. The signals in time segments are then decomposed into subbands in binary tree structure by using undecimated wavelet transform. The subband tree is then premed by assessing the discirimination power of the nodes. The resulting time-frequency map indicates the location of the best features for classification. This map is then used to extract features to be used for classification. It is observed that the extracted features increase the classification accuracy compared to various features previously used for the same problem.