Multiple Screening Face Detection

1
Multiple Screening Face Detection

• Noppadol Pringvanich
• Pamornpol Jinachitra
• Stanford University
• May 29th, 2002

2
Organization

• System overview
• Preparation and training
• Multiple screening process on testing image
• Results

3
System Overview
4
Preparation and Training

• Template
• Averaged representative faces with mean removed
• Size 40x34 pixels for closely center-cropped
  faces
• Representation of images by eigenimages
• Eigenfaces derived from 100 center-cropped faces
  using Sirovich-Kirby method
• Training face and non-face images projected onto
  the subspace spanned by 30 largest eigenfaces

5
Face Localization

• Template Matching
• 2-D convolution of flipped template with image
• Only peaks exceeding ¼ maximum output value
  are considered further

• Exact peak locations determined within a
  vicinity of twice the template size
• Only peaks which are highest in the vicinity
  pass as face candidates
• Input candidate locations into the
  classification unit

6
Distance Weighted k-Nearest Neighbor (DWNN)

• rj Euclidean distance between the projected
  target and its jth neighbor
• pj 1 if non-face
• 0 if face

7
Face/Non-face Classification

• Distance-weighted k-Nearest Neighbor
• Crop target symmetrically around candidate
  location
• Project targets onto the subspace of 30 largest
  eigenfaces
• Calculate the score from the sum of k-nearest
  neighbor, each term weighted by 1/distance
• Use k 11
• Effective when training faces will reappear!!

8
Skin-color criterion

• Faces have certain relationships of RGB
  components
• R gt B, R gt G and 0.8 lt G/B lt 1.7
• Vote face if gt 80 of the pixels satisfy the
  criterion
• Effective elimination of false detection such as
  jeans, tree, colored t-shirt.

9
Multiresolution Feedback

• Whole process repeated for different resolution
• Scaled template
• Patch already detected faces with green

10
Result (after template matching)
11
Result (after classification)
12
Final Result
13
Conclusion

• Performance
• Detection Rate 80.5
• False positive 16.4
• Multi-resolution
• Low run-time
• Scalability