2. BIRD EYE VIEW ABOUT THIS PAPER
Object Detection Framework(Viola-Jones Descriptor)
Down Sampled by Bessel Transform
Gabor Featrue Extraction Technique employed
Select numerous features using AdaBoost hypothesis
Neural Network Backpropogation algorithm use for
classification.
Tested on JAFEE and Yale Facial Expression
Database
Avergae Recogniton Rate is 96.83% and 92.2%.
Execution time for 100 Х 100 pixel is 14.5ms
3. THE PAPER
Title: A neural AdaBoost based facial expresson
recognition system
By: E.Owusu, Y. Zhan, Qi Rong Mao
From: Jiangsu University, China
Year: 2014
Citations: 12
Published: Expert System With Applications
Reference: http://www.sciencedirect.com/science/article/pii/S0957417413009615
4. INTRODUCTION
Facial expression involves application of AI.
It is related to Patten recognition and computer vision
Facial expression are seven prototypical ones, namely
Anger, fear, surprise, sad, disgust, happy, neutral
This technology is applied in various fields like robotics,
mobile applications, digital signs e.g.
AIBO Robot Biologically inspired robots
Some robots can display happiness feeling when detect face.
Databses: JAFFEE and Yale
5. SOME PREVIOUS WORK
Year Feature
Reduction
Feature
Extraction
Classification Performance
2001 PCA FFNN 84.5%
2012 PCA GFE NN 60-70%
2012 AMI FFNN 93.8%
2007 Sobel Filter Elmon N/W 84.7%
2009 PCA GFE NN 93.4%
In Most of the Studies:
Expression Classifier: Neural Network
Extracted Features: Gabor Filter
Feature Reduced: PCA
Displeasing is that the result is not encouragable.
6. PROPOSED TECHNIQUE
Data reduced by Bessel Transformation.
Extraction of the face by Gabor Methods
Feature Reduced by AdaBoost Feature Reduction Technique
Facial Expression Recognition using Bessel down
Sampling
Classifier is Multi layer feed forward neural network
using backpropogation
7. HOW PROPOSED TECHNIQUE WORKS
Face detection and image down-sampling
Gabor feature extraction
Feature selection
Multilayer feed forward neural network(MFNN)
8. FACE DETECTION AND DOWN-SAMPLING
Face Detection component was implemented by
Viola Jones.
Image is rescaled to 20 * 20px by Bessel Down
Sampling.
10. FEATURE SELECTION
Selection Algorithm
Initialize Sample Distribution
For the iteration t = 1, 2,..., T, where T is the final iteration
Normalize the Weight
Train a weak Clasifier
Select the hypothesis
Compute the weight
Update the weight distribution
Final Selection feature Hypothesis
12. TRAINING ALGORITHM
Process of Training Involves
Weight Initilization
Calculation of Activatin Function
Weight Adjustment
Weight Adaption
Testing for Convergence of N/W
13. TRAINING ALGORITHM
Training Algorithm Modeled as:
Activation Funciton of Hidden Units:
Activation Function of Output Units:
Network Error Function
14. HOW TO MINIMIZE THE ERROR
To minimize the error, each weight in the network
need to be computed.
Previous Weight Changes:
22. CONCLUSION
This study employs advance techniques
Improve recognition rate and execution time
Study Involves
Face Detection: Viola Jones Descriptor
Down Sampled: Bessel Transform
Extracted Feature: AdaBoost Algorithm
Select Feature: Gabour Wavelets
Selected Feature fed into MFFNN Classifier
Network trained by sample database JAFEE and
Yale
23. EXECUTION TIME AND RECOGNITION RATE OF
PROPOSED METHOD
Previous Performance
The execution time for a pixel of size 100 x 100 is 14.5
ms; the average recognition rate in JAFFE database is
96.83% and that in Yale is 92.22%.
Proposed Method
Study shows that
Automatic expression recognitions are very accurate in
surprise, disgusts and happy about 100%.
Mild expressions like sad, fear and neutral have lower
accuracies.