Face Detection

1
Face Detection
EE 368 Group 10 Waqar Mohsin Noman
Ahmed Chung-Tse Mar
2
Overview

• Project goals
• Implementation
• Skin color segmentation
• Morphological processing
• Connected region analysis
• Template matching
• Female recognition
• Results

3
Project Goal

• Detect and locate human faces in color images
  similar to those from the training set
• Limited variations in zoom and lighting
• Similar scene conditions
• Many occlusions

4
Implementation Overview
5
Skin Color Segmentation
6
Skin Color Segmentation

• Non-skin color regions have been eliminated
• Arms, hands, and various regions still remain

7
Morphological Processing
8
Morphological Processing
After
Before
9
Connected Region Analysis (Geometry)

• Reject regions that are
• Narrow
• Short
• Narrow and tall
• Wide and short
• Thresholds derived from training image statistics

10
Connected Region Analysis (Euler Number)
Reject regions with Euler number greater or equal
to 0
11
Result of Connected Region Analysis
After
Before
12
Template Matching
13
Template Matching
14
Female Face Detection

• Find the face with the white scarf
• Draw a box around the face centroid and count the
  number of white pixels
• Find the face with the long hair
• Draw a box around the face centroid and count the
  number of black pixels
• The one with the largest number of black or white
  pixels is designated as a female face

15
Results
Training Image Total faces Detected False Positive Repeat Hit
1 21 21 0 0
2 24 23 1 0
3 25 24 0 0
4 24 24 0 0
5 24 22 0 0
6 24 24 0 0
7 22 22 0 0
TOTAL 164 160 1 0
Detects 160 out of 164 faces with 1 false
positive. Average run-time on a Pentium 4 1.8GHz
PC is 35 seconds.
16
Conclusion

• Face detection program with 97 accuracy over the
  training images
• Run-time under a minute
• Hardest part of project is separating the
  connected faces
• Solved with successive template matching and
  blacking out face regions
• Not very robust, sensitive to scene conditions in
  the images

17
(No Transcript)
18
(No Transcript)