Hausdorff Distance Based Face Recognition

1
Hausdorff Distance Based Face Recognition

• CS-676
• Term Project
• Final Presentation
• Work Under Dr. Amitabha Mukherjee

Submitted by Ashwini Shukla (Y4112)
2
Work Basis Motivation

• To improvise on the recent methods of face
  recognition which mainly constitute handling edge
  images.
• Can be used in a variety of areas like-
• Personnel identification for intelligence from
  surveillance videos/security cameras,
• Automatic scorer identification in case of
  automaed scoring system.
• Inefficiency of other methods to cope with huge
  expression/pose variations ( as will be show
  later).
• Main reference of the work taken from- Robust
  Hausdorff distance measure for face recognition,
  Vivek/Sudha(2006) 1

3
Methodology

• Original Preprocessed
  Error Images

4
Methodology contd

• Extraction of
• Slopes.
• 
  a(7) a(6) a(5) a(4) a(3) a(2) a(1)
  a(0)
• Mapping of the Slope vector to RGB Value

5
Methodology contd

• For images in question calculate the H value as -
  
• 
  
• if v(a) v(b)
• 
  if v(a) v(b)
• A variant of this classical definition has been
  used where the max is replaced by the median as
  suggested by P.Guha in the proposal presentation.

6
Methodology contd

• Use of efficient Linked Lists based Algorithm to
  search nearest matching pixel.
• All pixels with same H value stored in same array
  index
• i
• i

7
Methodology contd

• Hence, we have now defined a similarity measure
  between two images.
• The image with the least H value is considered to
  be the best match, the one with next larger
  value, the second best match and so on.
• Results have been tabulated for two standard
  Face Databases viz. the ORL Face Database 3 and
  the YALE Database 2

8
Comparison with other existing methods

• Edge Images vs fully transformed image
• Loss of data in non edge regions
• Lam/Lin/Siu (2002) (spatially weighted
  hausdorff)
• Lam/Lin/Siu (2002) (eigen-weighted hausdorff)

9
Comparisons

• Comparison against Expression Variation

Comparison against Pose Variation
10
Results
11
What did not work

• As intitially proposed, to give higher weightage
  to eyes/nose ect.
• Reason In case of expression variations, these
  are the areas receiving the maximum change.
  Giving higher weightage thus increases the True
  Negatives.

12
References

• Robust Hausdorff distance measure for face
  recognition , E.P Vivek N.Sudha (2006)
• YALE University Face Database http//cvc.yale.edu
  /projects/yalefaces/yalefaces.html
• The ORL Database of Faceshttp//www.cl.cam.ac.uk/
  research/dtg/attarchive/facedatabase.html
• Spatially eigenweighted Hausdorff distances for
  human face recognition, Lam /Lin/Siu (2004)
• Human face recognition based on spatially
  weighted Hausdorff distance, Lam/Lin/Siu (2004)
• Locating and extracting the eye in human face
  images, Kin Man Lam Hong Yan(1995)

13
Thanks