Robodog Frontal Facial Recognition

1
Robodog Frontal Facial Recognition
AUTHORS GROUP 5 Jing Hu EE 05 Jessica
Pannequin EE 05 Chanatip Kitwiwattanachai EE 05
ADVISORS Dr. Daniel Lee Dr. Nabil Fahrat
Special Thanks to Paul Vernaza
Picture from Dog
University of Pennsylvania Department of
Electrical and Systems Engineering ABSTRACT Fac
ial Recognition has drawn a significant amount of
attention in the research area in the past few
years. There is an increasing interest in the
implementation of facial recognition systems
because of the emerging demands of more efficient
security systems. The ability to take into
account differences in lighting conditions,
facial orientation and background objects is
crucial for the implementation of a successful
system. Many different approaches of the problem
have been developed over the past two decades. So
far, each proposed method has different
comparative advantages and disadvantages. With
the chosen approach of this project, the face
region is first extracted from the original
picture using skin color analysis. The facial
features are then generated from the face
region. By doing so, the background noise can be
eliminated, thus increasing the recognition
accuracy and decrease the computation volume of
the system. The facial features are then fed
into a neural network to overcome image
distortion due to lighting condition, facial
expression and orientation of the face.
Finally, in order to enhance the role of
human-robot interaction for which recognition is
a crucial capability, the Sonys Aibo Dog is used
as the interface to the system.
Gray Scale Image First Crop
Skin Color Analysis
Rotation
Second Crop
Dog Reaction
Minimum Error
Feature Vector
Trained Neural Network
Recognition Algorithm
Feature Extraction
Graphical User Interface
Recognition Algorithm
The neural network outputs five errors, each
measuring the level of similarity with one of the
individuals trained by the network. The lower the
error, the closer the person is to one of the
trained individuals. Therefore, the smallest
error indicates who was recognized in a given
picture. In order to improve the robustness of
the system, the following algorithm is
implemented. Instead of determining recognition
based on a single picture,
Neural Network Training
For each person, 400 pictures are stored in a
database, with different facial orientations or
expressions. At each iteration, a picture is
taken at random from the database and passed once
through the neural network. The network is
trained over 150,000 iterations, which allows on
average each picture to be trained 75 times.
First Image Min Error1 lt 0.005?
Yes
Recognition
No
Second Image Min Error2 lt 0.005?
Yes
Recognition
Pass features through the Network once and adjust
the weights Wij and Wjk
No
a group of pictures is analyzed. At each stage, a
measure of the minimal error helps determine if
an immediate conclusion is possible or whether an
additional picture is required.
Yes
Min Error belongs to the same person as in 1st
Image?
Recognition
Neural Network Outline with three layers
Neural Network
Three most prominent SIFT Features of random
picture from database
No
Third Image Min Error3 lt 0.005?
Yes
After all iterations
Recognition
Recognition
Sample Set
Repeat for 150,000 iterations
No
Yes
Min Error1
No
Yes
Min Error belongs to the same person as in 1st
or 2nd Image?
Is Min Error of single Picture lt 0.05?
Recognition
Final Weights Wij and Wjk
Min Error2
Yes
No
No
Min Error3
Does Min belong to pair of pictures?
Take Minimum of 3 Errors
Convergence of Neural Network weights during
training
Start over with new group of pictures
Convergence of 1 element in the Wij matrix
Convergence of 1 element in the Wjk matrix
Recognition Results
Facial Database
95
88
85
86
84
Convergence of 1 element in the Delta Wjk matrix
Convergence of 1 element in the Delta Wij matrix
JP
JH
CH
FR
SG
86
88
95
85
84
DEMO TIMES Thursday, April 21st, 2005 930AM,
10AM, 1PM and 130PM
Note The Delta Wij and Delta Wjk matrices
correspond to the adjustment to the
Weights performed at each iteration. After
successful training, these numbers should
converge to zero, as shown on the above graph.
87.6 Rate of Success
2
(No Transcript)