Imtnan QAZI

1
Gender Classification based on Facial
information

• Imtnan QAZI
• Alina oprea
• Katerine diaz
• InayatUllah khan
• 16th Summer School on Image Processing
• July 15, 2008.

2
The project team
3
Layout

• Problem statement.
• State of the Art.
• The system overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

4
Gender classifier based on Facial information

• Men and Women Same
• Species, Different Planets
• Mathematical/Image processing viewpoint
• Binary classification provided constrained prior
  information and an elevated difficulty level for
  probability distribution modeling of the test
  data.

• Problem statement.
• State of the Art.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

Gender Classifier
5
Gender classifier based on Facial information

• Local features
• Skin colour, shape size of the face,
• amount of hairs, shape colour of the lips
• Higher difficulty level.
• Classification accuracies are mediocre.
• Global features
• Whole facial signature considered as a complete
  feature set.
• Useful training sequences required.
• Higher classification accuracies.
• Subspace methods Statistical Learners
• Principal Component Analysis (PCA).
• Fisher Linear Discriminator (FLD).
• Common Vectors (CV).
• Support Vector Machines (SVM).
• ..

• Problem statement.
• State of the Art.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

6
Gender classifier based on Facial information

• Problem statement.
• State of the Art.
• The system overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

7
Gender classifier based on Facial information

• Karhunen-Loeve Transform (KLT).
• Maps vectors from an M-d space
• to a n-d space n ltlt M.
• Computes eigenvectors of the covariance matrices
  for normal distributions.
• ,
• Other distances can also be used.
• Optimal linear dimensionality reducer.

• Problem statement.
• State of the Art.
• System overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

8
Gender classifier based on Facial information

• Supervised method.
• Label information considered.
• Inter-class Intra class scatter matrices
  proportional to covariance matrices.
• ,
• Generalized eigenvalue problem.
• Choice of suitable eigenvalue eigenvector for
  the solution.
• Largest eigenvalue is chosen.

• Problem statement.
• State of the Art.
• System overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

9
Gender classifier based on Facial information

• Feature space is divided in two orthogonal
  subspaces.
• Each sample in training sequence
• Difference subspace is equal to the rank of
  scatter matrix for each class.
• Minimizes the criterion
• which takes the form

• Problem statement.
• State of the Art.
• System overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

10
Gender classifier based on Facial information

• Optimal separating hyper plane.
• Function that predicts best
• response from some training functions.
• Given, observation-label pairs
• Minimizes the criterion
• 
  ,
• Kernel function

• Problem statement.
• State of the Art.
• System overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

11
Gender classifier based on Facial information

• Problem statement.
• State of the Art.
• System overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

Stability of PCA
12
Gender classifier based on Facial information

• Problem statement.
• State of the Art.
• System overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

Stability of FLD
13
Gender classifier based on Facial information

• Problem statement.
• State of the Art.
• System overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

Stability of CV
14
Gender classifier based on Facial information

• Problem statement.
• State of the Art.
• System overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

Stability of SVM
15
Gender classifier based on Facial information

• Problem statement.
• State of the Art.
• System overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

Stability of PCA SVM
16
Gender classifier based on Facial information

• Problem statement.
• State of the Art.
• System overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

17
Gender classifier based on Facial information

• Problem statement.
• State of the Art.
• System overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

18
Gender classifier based on Facial information

• Thank GOD!! My mind can
• recognize female faces easily. ?
• Stability of different methods depends on number
  of training sequences.
• SVM proves to be stable and reliable global
  classifier with acceptable accuracy.
• Using PCA as dimension reducer and SVM as a
  classifier can produce better results, if more
  training sequences can be used.

• Problem statement.
• State of the Art.
• System overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.

19
Gender classifier based on Facial information

• Tests with larger databases.
• In depth stability analysis of different global
  classifiers.
• Other techniques like Neural Networks may be used
  to validate different conclusions drawn.
• To find a funding source to attend next summer
  school. ?

• Problem statement.
• State of the Art.
• System overview.
• Principal Component Analysis.
• Fisher Linear Discriminator.
• Common Vector method.
• Support Vector Machines.
• Simulations Results.
• Conclusion.
• Future Perspectives.