TY - GEN

T1 - Class distance weighted locality preserving projection for automatic age estimation

AU - Ueki, Kazuya

AU - Miya, Masakazu

AU - Ogawa, Tetsuji

AU - Kobayashi, Tetsunori

PY - 2008/12/1

Y1 - 2008/12/1

N2 - We have developed new dimensionality reduction methods, extended from locality preserving projection (LPP), to estimate age using facial images. LPP seeks a linear transformation matrix such that optimally preserves the neighborhood structure of the data. Our focus has been on expanding LPP by making use of class label information. Specifically, one of our ideas is to assign weights only to the data with close class labels. A local scaling method is used for each class to compute the LPP affinity matrix. Another idea is to assign large weights to two samples with close class labels, i.e., close ages. By doing this, class label information for original data (i.e., age information) can be preserved. We thus call this "class distance weighted linear preserving projection" (CDLPP). Experimental results on a large database showed that CDLPP has more discriminative power than conventional methods such as PCA and LPP.

AB - We have developed new dimensionality reduction methods, extended from locality preserving projection (LPP), to estimate age using facial images. LPP seeks a linear transformation matrix such that optimally preserves the neighborhood structure of the data. Our focus has been on expanding LPP by making use of class label information. Specifically, one of our ideas is to assign weights only to the data with close class labels. A local scaling method is used for each class to compute the LPP affinity matrix. Another idea is to assign large weights to two samples with close class labels, i.e., close ages. By doing this, class label information for original data (i.e., age information) can be preserved. We thus call this "class distance weighted linear preserving projection" (CDLPP). Experimental results on a large database showed that CDLPP has more discriminative power than conventional methods such as PCA and LPP.

UR - http://www.scopus.com/inward/record.url?scp=67649088888&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=67649088888&partnerID=8YFLogxK

U2 - 10.1109/BTAS.2008.4699380

DO - 10.1109/BTAS.2008.4699380

M3 - Conference contribution

AN - SCOPUS:67649088888

SN - 9781424427307

T3 - BTAS 2008 - IEEE 2nd International Conference on Biometrics: Theory, Applications and Systems

BT - BTAS 2008 - IEEE 2nd International Conference on Biometrics

T2 - BTAS 2008 - IEEE 2nd International Conference on Biometrics: Theory, Applications and Systems

Y2 - 29 September 2008 through 1 October 2008

ER -