Face Detection Using Color Thresholding and Eigenimage Template Matching

1
Face Detection Using Color Thresholding and
Eigenimage Template Matching

• Diederik Marius
• Sumita Pennathur
• Klint Rose

2
Approach
Input Image
Homogeneity/ Aspect Ratio Threshold
Size Threshold
Cross-covariance with Eigenimage
YCbCr Thresholding
Rejection
Individual blobs
Binary Image Processing
Match Face Location to Input Image
Duplicate Detection Removal
Separation of Blobs
3
Skin Segmentation

• Used distribution of Cr and Cb values for faces
  vs. background to threshold

Skin region 105ltCrlt135 140ltCblt165
4
Binary Image Processing

• Erosion and dilation with face-shaped
  segmentation element
• Removes small foreground and background objects
• Delineates between larger regions

5
Separation and Rejection

• Larger regions labeled and separated
• Series of Rejection Thresholding
• Size of regions (gt3000 total pixels)
• Aspect Ratio (0.5 ltARlt 1.8 )
• St. Dev. of image values (60ltslt100)

Coupled AR and St. Dev. based thresholding
Size based thresholding
6
Eigenimages

• Computed from set of good faces
• Employed Sirovich-Kirby method to calculate first
  10 eigenimages
• Eigenimage 2 most accurate
  location of center of face
• Used eigenimage 2 exclusively

7
Template Matching

• Found cross-covariance peak and marked as
  potential center
• Removed face sized area from image
• Found new highest peak - repeated 10x

8
False Detection Removal

• Thresholded detected peaks to obtain potential
  faces only
• Determined whether multiple peaks belong to the
  same person (neck, etc)
• Rejection criteria Near the same y-axis and
  within a predetermined vertical distance of a
  previous point

Rejected points
9
Typical Detection Result

• Detection problems rotated faces, small faces,
  hands, faces in lower 1/3

10
Results

• Average run time 100 seconds
• 95 of faces found
• 4.4 false positives

11
(No Transcript)
12
(No Transcript)