Modified Eigen Vectors Arrangement for Better Representation of Images in Low Dimension

Dimension reduction techniques, PCA and LDA give preference to eigenvectors corresponding to higher Eigen values. This theory is not appropriate for all types of applications. In this work a new way of arranging the eigenvectors is explored. The proposed method combines the concept of correlation with a variability measure ‘range’ to rank dimensions and hence the eigenvectors. PCA and LDA are modified to incorporate the proposed dimension ranking method. Experiments are conducted with WANG and ZuBuD databases, and the performance is evaluated using precision values of a developed CBIR system for traditional as well as modified versions of PCA and LDA.