Abstract
Original language  Undefined 

Awarding Institution 

Supervisors/Advisors 

Thesis sponsors  
Award date  1 Jun 2012 
Place of Publication  Enschede 
Publisher  
Print ISBNs  9789036533676 
DOIs  
Publication status  Published  1 Jun 2012 
Keywords
 EWI22962
 PrincipleComponent Analysis
 Face Recognition
 IR80426
 METIS290374
Cite this
}
SOS, lost in a high dimensional space. / Hendrikse, A.J.
Enschede : University of Twente, 2012. 158 p.Research output: Thesis › PhD Thesis  Research UT, graduation UT › Academic
TY  THES
T1  SOS, lost in a high dimensional space
AU  Hendrikse, A.J.
PY  2012/6/1
Y1  2012/6/1
N2  The trend in facial biometrics has been to use ever increasing image resolution, with the purpose of increasing the recognition performance by exploiting the added information. One category of biometric systems expected to benefit from the increased image resolution consists of systems based on secondorder statistics (SOS) estimates, such as those based on principle component analysis (PCA). Increasing the image resolution without sufficiently increasing the number of training samples has several effects on the SOS estimates, such as a bias in the eigenvalue estimates and errors in the eigenvector estimates. We analyze how the increasing ratio of the dimensionality over the number of samples affects biometric systems, in particular those based on secondorder statistics in combination with a – theoretically optimal – loglikelihood ratio classifier. We show that the classical solution to the singularity problem, PCA dimensionality reduction, is far from optimal and fails completely for very high dimensionalities and we present several solutions to adjust the SOS estimates in order to achieve close to optimal performance, such as the eigenwise correction using fixedpoint eigenvalue correction, and the variance correction. Although the presented solutions are clearly superior if synthetic data is used, for real facial data they turned out to be outperformed by PCA dimensionality reduction. We found that this can be explained by the assumed underlying model of fixed position intensity sources, which cannot efficiently describe variations occurring in faces caused by moving features. We show that if facial data contains such moving features, then traditional solution to the singularity problem by dimensionality reduction based on PCA reduces the disruptive effect of these moving features on verification rates while our proposed bias correction methods actually increase this effect. This provides an explanation why PCA dimensionality outperforms the correction methods if real facial data is used.
AB  The trend in facial biometrics has been to use ever increasing image resolution, with the purpose of increasing the recognition performance by exploiting the added information. One category of biometric systems expected to benefit from the increased image resolution consists of systems based on secondorder statistics (SOS) estimates, such as those based on principle component analysis (PCA). Increasing the image resolution without sufficiently increasing the number of training samples has several effects on the SOS estimates, such as a bias in the eigenvalue estimates and errors in the eigenvector estimates. We analyze how the increasing ratio of the dimensionality over the number of samples affects biometric systems, in particular those based on secondorder statistics in combination with a – theoretically optimal – loglikelihood ratio classifier. We show that the classical solution to the singularity problem, PCA dimensionality reduction, is far from optimal and fails completely for very high dimensionalities and we present several solutions to adjust the SOS estimates in order to achieve close to optimal performance, such as the eigenwise correction using fixedpoint eigenvalue correction, and the variance correction. Although the presented solutions are clearly superior if synthetic data is used, for real facial data they turned out to be outperformed by PCA dimensionality reduction. We found that this can be explained by the assumed underlying model of fixed position intensity sources, which cannot efficiently describe variations occurring in faces caused by moving features. We show that if facial data contains such moving features, then traditional solution to the singularity problem by dimensionality reduction based on PCA reduces the disruptive effect of these moving features on verification rates while our proposed bias correction methods actually increase this effect. This provides an explanation why PCA dimensionality outperforms the correction methods if real facial data is used.
KW  EWI22962
KW  PrincipleComponent Analysis
KW  Face Recognition
KW  IR80426
KW  METIS290374
U2  10.3990/1.9789036533676
DO  10.3990/1.9789036533676
M3  PhD Thesis  Research UT, graduation UT
SN  9789036533676
PB  University of Twente
CY  Enschede
ER 