A linear maximum variance unfolding algorithm and its application in image recognition

A novel linear dimensionality reduction algorithm, called linear maximum variance unfolding (LMVU), is proposed. Both in the training data or the new test data (out of samples) can be mapped to a new low-dimension subspace by a transformation matrix obtained by LMVU. This linear transformation optimally preserves local neighborhood information in a certain sense. Comprehensive comparisons and several experiments show that LMVU can discover faithful low dimensional representations of high-dimension images, and achieves much higher recognition rates than a few competing methods.