Automated Face Detection

1
Automated Face Detection

• Peter Brende
• David Black-Schaffer
• Veni Bourakov

2
Primary Challenges

• Scale differences
• Overlapping/obstructed faces
• Lighting variation

3
Implementation Overview

• Color-based skin separation
• Spatial analysis to generate candidate faces
• Morphological?
• or Template Matching?
• Eliminate false hits
• Hands arms, based on shape
• Roof, based on texture
• Neck, based on relative position

4
Color-based Skin Separation

• What color space to use?
• How to separate out the skin color?

5
Marginal Color Distributions
6
Parametric Separation

• Simple Fast (hgt0.98 or hlt0.01)
• Problems with non-linearity of HSV space in
  bright areas

7
Full Joint-Probability Distribution

• We have enough data, so why not?
• Provides most accurate per-pixel classification.
• Allows use to circumvent choosing a decision
  boundary. We can simply use Bayes rule.

8
Slices from 3D Joint Distribution
9
Skin-probability Image Obtained From Applying
Bayes Rule
10
Color-based Skin Separation

• What color space to use?
• HSV if separability of distributions is necessary
• How to separate out the skin color?
• Parametric is fast but loose in HSV
• Provides a binary mapping and requires choosing
  thresholds
• Full PDF is accurate in any color space
• Can be fast if done correctly (table lookup)
• No thresholds produces a pure probability map

11
Spatial Analysis Method

• Morphological
• Obtain binary mask through thresholding
• Perform morphological operations to separate and
  identify blobs corresponding to faces
• Difficult due to overlapping faces
• Template Match
• Search a scene for prototypical face image
• Need to decide which data to work with (luminance
  vs. skin probability)

12
Template Matching

• Using the skin-probability image
• Greatly simplifies information content
• Simple information ? simple algorithm
• Allows algorithm to focus on the single best
  facial clue oval-shaped skin regions
• Allows us to avoid creating a binary mask

13
Correlation of simple template with
Skin-probability image
14
Process of Inclusion/Elimination

• Iteratively pick strong regions of the
  skin-probability image as faces
• For each template search for matching face shapes
  (convolution peaks)
• For each detected face, subtract/erase the
  region from image to avoid duplicate detection
• Stop when no significant skin regions remaining

15
Positive Detection and Elimination
16
(No Transcript)
17
(No Transcript)
18
(No Transcript)
19
(No Transcript)
20
(No Transcript)
21
(No Transcript)
22
(No Transcript)
23
(No Transcript)
24
Template and Threshold Selection

• Critical step of our algorithm
• Potential problems
• Template matching face of a different size
• Small template leads to double hits later
• Large template leads to missed faces
• Bad thresholds ? low sensitivity or low
  specificity
• Solution
• Use templates of many sizes, going from largest
  to smallest
• Set threshold as high as possible without
  sacrificing sensitivity

25
Algorithm Implementation

• 1. Load probability and template data
• 2. Down-sample the image by factor 21
• 3. Calculate the face-probability image by color
• 4. Remove hands/arms
• 5. Template match with skin-probability image
• 6. Eliminate false positive hits on necks of
  large faces
• 7. Remove patterned hits

26
Overall Results
27
Conclusions

• Recognition of face-shaped blobs from the
  skin-probability map works excellently
• Requires that the skin colors be well known
• Requires that the general face sizes be well
  known
• Our set of images was relatively consistent in
  terms of these factors

28
(No Transcript)
29
(No Transcript)