Craniofacial Aging Impacts on the Eigenface Biometric

1
Craniofacial Aging Impacts on the Eigenface
Biometric
K. Ricanek Jr., E. Patterson, and E.
Boone Departments of Computer Science and
Mathematics and Statistics University of North
Carolina Wilmington 601 S. College Rd.,
Wilmington, NC Presenter Contact
ricanekk_at_uncw.edu
2
Outline

• Craniofacial Aging
• Craniofacial Morphologic Data Corpus
• Face Biometrics
• Algorithms
• PCA
• PCALDA
• Bayesian ML
• Bayesian MAP
• Evaluation
• Results
• Conclusions

3
Craniofacial Aging

• Craniofacial changes continue throughout
  adulthood.
• Boney structure bone remodeling, bone
  thickening, cranial expansion, prognathism
  (protrusion of the jaw), etc.
• Soft tissue hyperdynamic expressions, loss of
  elasticity in dermis layer, etc.
• Various behaviors can affect aging.
• Smoking
• Sun Exposure (Photo-Aging)
• Drug and Alcohol Abuse
• Variable rate of morphology
• People age at varying rates gender, ethnicity,
  and behavior.

4
Craniofacial Aging
Male 16 to 44
Female 32 to 42
5
Craniofacial Morphologic Data Corpus

• Developed specifically for research into adult
  craniofacial aging.
• Largest number of subjects of the three publicly
  available corpuses
• FERET (J. Phillips)
• (Dup dataset) age spans less than 5 yrs.
• FG-NET Aging Database (A. Lanitis)
• Childhood through Adulthood of 82 subjects
• MORPH
• Album 1 515 subjects across adulthood
• Album 2 3000 subjects

6
MORPH Album1
7
Face Biometrics

• Automated techniques for identification of a
  person using photographs/images of the face.

8
Algorithms

• A variety of algorithms have been developed for
  face biometrics which utilizes intensity values
  of pixels for recognition.
• These algorithms attempt to determine which
  pixels contain the most variation and can help
  correctly match the probe image to the gallery
  images.

9
Eigenface Biometric

• Principle Components Analysis (PCA).
• Determine which pixel intensities are correlated
  with each other using the covariance matrix ST
• The eigenvalue projection of the images via Wopt
  creates comparison space (Face space), which can
  be dimensionally reduced.

10
Fisherface Biometric

• Combines PCA with Linear Discriminate Analysis
  (LDA).
• Once in the projection space created by PCA
  determine which dimensions best discriminate
  between people.
• The PCALDA space provides a comparison space for
  images.

11
Bayesian ML and MAP

• Define images to Intraclass WI and WE.
• Given an image difference D we compare images to
  determine P(WI D) via Bayes rule.
• P(WI D) is often called a similarity measure.
• The subject with the highest similarity score is
  considered the match, maximum a posteriori

12
Bayesian MAP

• Determining P(DW) can be approximated by
• SMAPP(WID) is the maximum aposteriori
  similarity measure.
• Computationally intensive.

13
Bayesian ML

• A maximum likelihood approach.
• Less computationally intensive.
• SML is the similarity measure from this approach.

14
Aging of the Face
15
Evaluation

• Each image is either recognized or not hence a
  binary response.
• A generalized linear mixed model, specifically
  mixed logistic regression with a random effect
  for subject.
• Age difference is used as a predictor.
• The resulting odds ratio can be interpreted as
  for each one year increase in time difference we
  expect an OddsRatio times increase in recognition
  rate.
• This was performed for rank one and rank five
  recognition.

16
Results
17
Graph of Results
18
Conclusions

• These basic methods have a strong relationship
  between time difference and recognition rate.
• Any method utilizing these methods should be
  studied further with respect to aging in order to
  ascertain the methods true recognition rate.