A simple feature extraction for high dimensional image representations

Abstract

We investigate a method to find local clusters in low dimensional subspaces of high dimensional data, e.g. in high dimensional image descriptions. Using cluster centers instead of the full set of data will speed up the performance of learning algorithms for object recognition, and might also improve performance because overfilling is avoided. Using ihe Graz01 dalabase, our melhod outperforms a current standard method for feature extraction from high dimensional image representations. © Springer-Verlag Berlin Heidelberg 2006.