1. Face Recognition Method of OpenCV
Andrew stodghill

2. Overview
Introduction
Background
Theory
Experiment
Results
Conclusion

3. Introduction

4. Idea Facial recognition for a Security Robot
Addition to another project
Gain a better understanding of the subject
Derived from work last semester

5. Focus
Basic understand of OpenCV face recognition software and algorithms
Methods and Theory behind the EigenFace method for facial recognition
Implementation using Python in a Linux-based environment
Runs on a Raspberry Pi

6. Goal One half research One half implementation
General facial recognition methods
EigenFaces
OpenCV’s facial recognition
One half implementation
Create a system capable of facial recognition
Real-time
Able to run on a Raspberry Pi

7. Background

8. Different Facial Recognition Methods
Geometric
Eigenfaces
Fisherfaces
Local Binary Patterns
Active Appearance
3D Shape Models

9. Geometric First method of facial recognition Done by hand at first
Automation came later
Find the locations of key parts of the face
And the distances between them
Good initial method, but had flaws
Unable to handle multiple views
Required good initial guess

10. Eigenfaces Information theory approach
Codes and then decodes face images to gain recognition
Uses principal component analysis (PCA) to find the most important bits

11. Fisherfaces Same approach as Eigenface
Instead of PCA, uses linear discriminant analysis (LDA)
Better handles intrapersonal variability within images such as lighting

12. Local Binary Patterns Describes local features of an object
Comparison of each pixel to its neighbors
Histogram of image contains information about the destruction of the local micro patterns

13. Theory

14. Basic Idea
Let face image 𝐼(𝑥,𝑦) be a two-dimensional 𝑁 by 𝑁 array of (8-bit) intensity values
Can consider image an 𝑁 2 vector of dimensions
Image of 256 by 256 becomes a 65,536 vector of dimension
Or a point in 65,536-dimensional space

15. Basic Idea Images of faces will not differ too much
This allows a much smaller dimensional subspace to be used to classify them
PCA analysis finds the vectors that best define the distribution of images
These vectors are then
𝑁 2 long
Describe an 𝑁 by 𝑁 image
Linear combination of the original face images

16. Basic Idea These vectors are called eigenfaces
They are the eigenvectors of the covariance matrix
Resemble faces

17. Method Acquire initial training set of face images
Calculate eigenfaces
Keep only 𝑀 eigenfaces that correspond to the highest eigenvalues
These images now define the face space
Calculate corresponding distribution in 𝑀-dimensional weight space for each known individual

18. Method
Calculate weights for new image by projecting the input image onto each of the eigenfaces
Determine if face is known
Within some tolerance, close to face space
If within face space, classify weights as either known or unknown
(Optional) Update eigenfaces and weights
(Optional) If same face repeats, input into known faces

19. Classifying Four possibilities for an input image
Near face space, near face class
Known face
Near face space, not near face class
Unknown face
Not near face space, near face class
Not a face, but may look like one (false positive)
Not near face space, not near face class
Not a face

20. OpenCV and Theory
Beauty about OpenCV is a lot of this process is completely automated
Need:
Training images
Specify type of training
Number of eigenfaces
Threshold
Input Image

21. Experiment

22. Set-up OpenCV running on a Raspberry Pi Linux-based environment
Raspberry Pi Camera

23. Training Database of negatives
AT&T Laboratories Database of faces developed in the 90s

24. Training Captured Positives Used the camera to capture images
Images were then cropped and resized

25. Training Model was trained using positive and negative images
Creates training file that holds the 𝑀-dimensional face space
Now have a base to recognize from
model = cv2.createEigenFaceRecognizer() model.train(np.asarray(faces),np.asarray(labels))

26. Recognition Steps to recognizing face
Capture image
Detect face
Crop and resize around face
Project across all eigenvectors
Find face class that minimizes Euclidian distance
Return label from face class, and Euclidian distance
Euclidian distance also called Confidence level
model = cv2.createEigenFaceRecognizer()
model.load(config.TRAINING_FILE)
label, confidence = model.predict(image)

27. Test Created four different Test
First data set uses 24 positive training images
Almost no pose and lighting variation
Second data set uses 12 positive training images
Good pose variation, little lighting variation
Third data set uses 25 positive training images
Good pose and lighting variation
Fourth data set uses second and third data set but with Fisherface method

28. Results

29. Results Results from Data sets 1-3, each one from 20 input images
Confidence represents distance from known face class
Data Set
Mean Confidence
Max Confidence
Min Confidence 1
3462
3948
3040 2
2127
2568
1835 3
1709
2196
1217

30. Results Results from eigenface vs. fisherface comparison Algorithm
Data Set
# Training Images
Mean Confidence
Max Confidence
Min Confidence
Eigen 2 12
2127
2568
1835
Fisher
2029
2538
1468 3 25
1709
2196
1217
2017
2748
1530

31. Conclusion

32. Conclusion Theory behind eigenfaces
Face space
Training
Simple implementation of OpenCV’s eigenface recognizer
Compared different training models
Number of training images
Pose and lighting variations
Compared eigenfaces and fisherfaces

33. Conclusion Future work Further testing of different training models
Implement updating facial recognition

34. Questions?