Optimized two-dimensional local discriminant basis algorithm

Local discriminant basis algorithm (LDB) is a supervised scheme for feature extraction and nonstationary signal classification. Due to its fast computational time, O(n log n), and excellent time-frequency localization, it is a promising method for nonstationary signal analysis. An optimized version of local discriminant basis (OLDB) has been recently proposed that emphasizes certain regions of interest in different classes using initial LDB features. The optimization process is particularly useful when background structures show high correlation with desired features in signal or image space as in mammograms. In this paper the performance of OLDB algorithm is studied in a texture classification problem. Classification into more than two classes of signals or images is a challenging problem and OLDB is capable of obtaining 85% accuracy classifying grayscale 64×64 textured images into three classes using only the 280 top LDB features as studied in this paper.