Abstract: An adaptive features extraction method is proposed. It defines an adaptive objective
function based on the projective space derived by using PCA method. Then the projection space of
the individual sample is computed. The distribution characteristics of each sample are well considered.
In order to make the algorithm applicable to the classification problem, a similarity measurement is
proposed to calculate the similarity between individual samples in different projection spaces.
Compared with the Euclidean metric, the proposed measurement is proved that can represent the
geodesic distance relationship between the samples better, so that the proposed method can learn the
manifold data effectually. The classification and reconstruction experiments on the different databases
indicate that the new method can obtain features more effectively and robustly.

Key words: feature extraction, principal component analysis, adaptive feature extraction, face recognition