Embed Size (px)
Citation preview
Face Detection Using Skin Color and Gabor Wavelet Representation
Information and Communication Theory GroupFaculty of Information Technology and System
Delft University of TechnologyDelft - Netherlands
Resmana Lim, Marcel J.T. Reinders
email: [email protected]
Agenda of Presentation
Introduction
Extraction of Face Candidate Regions
Face Detection in Face Candidate Regions
Graph Matching by Genetic Algorithm
Experiment Results
Conclusion
Introduction
Face detection and detecting facial landmarks (such as position of eyes, nose, mouth, etc.) play an important role in face recognition systems
This paper focuses on the robust and accurate detection of landmark points on the face
The approach first uses color information to detect face candidate regions and then uses a deformable graph matching to locate facial landmark points in these candidate regions
The method is made robust against lighting variantions and variations between people by representing the landmark points using Gabor filter responses
This paper introduce an alternative matching process by using a Genetic Algorithm (GA) to optimize the matching criteria in finding faces candidate regions and detection of the facial landmarks
Extraction of Face Candidate Regions
First step, to separate skin regions from non-skin regions based on color information
Main objective is to reduce search space for the faces drastically
The identification of the facial region is determined by utilizing a priori knowledge of the skin distribution in the HS color space
Extraction of Face Candidate Regions
Third step, to fill the holes in the face
Finally, small isolated regions that remain after this step are removed
Second step, to smoothen object silhouettes, and also to eliminate any isolated misclassified pixels that may appear as impulsive-type noise
Extraction of Face Candidate Regions
Flow Diagram of Skin Region Extraction
Extraction of Face Candidate Regions
Face Detection in Face Candidate Regions
We need identify for each face candidate region whether it is a face or not and if so what the position of the landmarks points are
For indentifying face candidate region is a face or not, we apply a graph matching procedure based on Genetic Algorithm
Matching the face candidate regions againts face model graph
facial landmarks in the face region is found by maximizing the simmilarity between the face model graph and the face region image
Each facial landmark is represented by the expected local Gabor filter responses in the image
Face Detection in Face Candidate Regions
iyx
yxyxyxf
kk
y
kk
x
kkk
)sincos(2exp.
)cossin()sincos(
2
1exp),,,(
2
2
2
2
2-D Gabor Filter Kernel
Where
nkknk ,..,2,1)1(
Face Detection in Face Candidate Regions
1
)(
1
)(
),,,(),(),,,(xN
xNu
yN
yNvkk vufvyuxIyxg
Single Gabor Filter Response
j=1,..,m )exp( jjj iaJ
Gabor Jet
This research uses eight different orientations (n=8) and four different wavelengths (=3,5,7,10) resulting in 32 filter response
Face Detection in Face Candidate Regions
Face model consisting of four landmark points (p1,…,p4) represented by their Gabor filter bank responses
Graph Matching by Genetic Algorithm
Objective function
l
j jjj
jjj
e
eeJl
Jl
Jl
Eaa
aa
JJS
)(
)(2
2
2'2
'
'
'
),(
),(max JJSJ
Where
Crossover rate = 1, mutation rate = 0.01
80 individuals at each generation
Graph Matching by Genetic Algorithm
Solution is represented by five parameters
x and y position (coordinate of the left eye)
scaling factor in x and y direction, rotation angle rotation angle
The threshold value of 0.7 is used to judge whether the candidate face region constitutes a face or not.
Experiment Result
Facial Landmark localization
Conclusion
We have proposed a detection scheme for locating facial landmarks based on color information and graph matching using a genetic algorithm as optimization strategy.
The performance of the proposed method was demonstrated on various color images containing single and multiple faces.
The results are quite promising for frontal pose faces with moderate rotation and tilting. From the results of the experiment, we conclude that the proposed method has a good prospect and should be considered in the design of face recognition systems
