DO I KNOW YOU? OPTIMIZING THE FACIAL RECOGNITION SYSTEM THROUGH - - PowerPoint PPT Presentation

DO I KNOW YOU? OPTIMIZING THE FACIAL RECOGNITION SYSTEM THROUGH DISTANCE METRICS

KEVIN PARK UNCW REU: STATISTICAL LEARNING AND DATA MINING 7/25/2017

ACKNOWLEDGEMENTS

 With all my heart, I express sincere gratitude and thankfulness to my REU family.  Thank

You to National Science Foundation and University of North Carolina Wilmington for sponsoring this summer experience.

2

OUTLINE

1

I. Introduction

2

II. Experimental Design

3

III. Results and Conclusion

3

• 1. INTRODUCTION

WHAT IS FACE RECOGNITION?

 Mechanism for a computer to determine

whether a given face image is identifiable or unknown

 Use machine learning and computer vision

techniques to match new, unseen face images to a given set of trained images computer has already seen

 Face Recognition has been developing interest

with its wide applications in security, biometrics, and industry

Image Known Unknown

Source: http://www.nec.com/en/global/solutions/safety/face_recognition/images/Face_Recognition_FR_Pic.png

5

PROJECTION ALGORITHMS AND DIMENSION REDUCTION

Eigenfaces (Principal Component Analysis)



“Transform a face image into a set of crucial features, called eigenfaces” (Turk)



These eigenfaces are put into an eigenvector which contains features that explain the variation between face images

Fisherfaces (Linear Discriminant Analysis)

Image Convert Image to Vector Too Many Dimensions!! Dimension Reduction



LDA obtains discriminative information by maximizing the ratio between group and within group variance which optimizes class separability



Face space is then obtained after this maximization

Source: http://courses.ee.sun.ac.za/Pattern_Recognition_813/lectures/lecture01/img33.png

Source: Turk, M.a., and A.p. Pentland. "Face Recognition Using Eigenfaces." IEEE Computer Society Conference on Computer Vision and Pattern Recognition (1991): 586-91. Web.

6

DISTANCE METRICS

Image One Image Two How to Compare Them? Distance Metrics Name Formula Euclidean City Block Cosine Distance Mahalanobis Cosine d = [(x − y)t S−1(x − y)]1/2 Name Formula Bray Curtis Canberra Correlation Chebyshev Di = max | xi – yi |

7

RESEARCH PROJECT GOALS

 Objective:  1) Determine which combination of dimension reduction techniques and distance

metrics will provide the greatest accuracy in the facial recognition task.

 2) Assess the impact of combining multiple distance metrics (decision fusion-like)

• n face recognition accuracy.

Distance Metric One Distance Metric T wo Distance Metric Three Distance Metric Four Distance Metric Five Eigenface Fisherface

8

• 1I. EXPERIMENTAL DESIGN

DATA USED

 The Morph II Database is a collection of

55,134 face images of 13,617 unique subjects.

 The 55,134 mugshots are from 2003 and 2008

and include images of individuals that were arrested once or multiple times.

 The size and longitudinal features of Morph II

make it a widely used data source for computer vision and machine learning research.

Sample Images Cleaned Images

10

SUBSET SCHEME

Morph II Subset

• The subset was developed to only contain

individuals with 10 or more face images in Morph II.

• 166 subjects with a total of 1,660 images
• 83 males and 83 females.
• 126 black individuals, 39 white subjects, and

1 Hispanic person.

• The age distribution of the subset ranges

from 16 to 61 years of age where 75% of the subjects are 41 years of age or younger.

11

EXPERIMENTAL SETUP

 Leave One Image Out Setup

For every subject, nine out of the ten face images are randomly assigned to training. The tenth image is then assigned to testing. Facial recognition accuracy is calculated

• nce the test images are matched to a

training image.

12

RESEARCH PROJECT PROCEDURE: FACIAL RECOGNITION

Image Eigenface (PCA) Fisherface (LDA) Euclidean Distance Distance Metrics Classification Method

• Nearest

Neighbors Facial Recognition

13

Projection Algorithms City Block Distance Cosine Distance

Correlation Distance

Bray Curtis Distance Canberra Distance Chebyshev Distance

• M. Cosine Distance

1,494 Training Images and 166 Testing Images 100 principal components

IMPORTANCE OF DISTANCE METRICS

 A new face image will be matched to an image by comparing the test image’s feature vector to every

training image’s feature vector.

 “Distances between two vectors can be treated as a measure of dissimilarity of the two biometric

samples” (Yassin)

Testing Feature Vector Training Feature Vector 1 Training Feature Vector 2 Training Feature Vector (n) Comparison Calculate Distance Metric Facial Recognition Classification

14

DECISION FUSION IDEA USING DISTANCE METRICS

Train Image Test Image Calculate distance with single distance metric Apply Nearest Neighbors Facial Recognition Clasification Distance Metric 1 Distance Metric 2 Distance Metric N Apply Nearest Neighbors Method Used Before Proposed Method Train Image Test Image Facial Recognition Clasification Re-Scale Distance Using minimum distance

15

• 1II. RESULTS AND CONCLUSION

DISTANCE METRICS RESULTS

Euclidean City Block Cosine Correlation Bray Curtis Canberra Chebyshev

Mahalanobis Cosine

Eigenfaces

69.27% 75.90% 68.07% 70.48% 78.31% 77.71% 46.98% 78.91%

Fisherfaces

89.15% 81.92% 95.18% 94.57% 93.97% 75.90% 75.90% 80.72%

17

Experiment Attributes:

• Leave One Out: 9 Images or Training and1Image for Testing
• 1,494 Training Images and 166 Testing Images
• 100 Principal Components
• Chebyshev distance metric resulted in the lowest

accuracy for face recognition.

• Face Recognition Accuracy under

Eigenfaces: 69% to 78% Fisherfaces: 75% to 95%.

DECISION FUSION PRELIMINARY RESULTS

Euclidean City Block Cosine Correlation Bray Curtis Canberra Chebyshev

Mahalanobis Cosine

Eigenfaces

69.27% 75.90% 68.07% 70.48% 78.31% 77.71% 46.98% 78.91%

18

CONCLUSION AND FURTHER DIRECTIONS

 The experiments illustrate that a particular distance metric with a corresponding projection technique

can play a major role in increasing face recognition accuracy.

 As the Cosine and Mahalanobis Cosine distance were the best distance metrics, there seems to be

evidence to suggest that studying the angle and orientation between two feature vectors is a crucial component in facial recognition.

 Studying the orientation rather than physical distance may be a more promising direction to explore in

face recognition.

 An interesting application to further explore would involve using distance metrics with elastic bunch

graphing method in face recognition.

 In regards to decision fusion, future work can try more involved weighting schemes and different

combinations of distance metrics.

19

REFERENCES

 Liu, Kuan-Hsien, ShuichengYan, and C.-C. Jay Kuo. "Age Estimation via Grouping and Decision Fusion."

IEEE Transactions on Information Forensics and Security 10.11 (2015): 2408-423. Web.

 Seo, Naotoshi. "Eigenfaces and Fisherfaces." ENEE633 Pattern Recognition, Jan. 2017. University of

Maryland.

 Turk, M.a., and A.p. Pentland. "Face Recognition Using Eigenfaces." IEEE Computer Society Conference

• n Computer

Vision and Pattern Recognition (1991): 586-91. Web.

 Yassin, Dk H. Phm, S. Hoque, and F. Deravi. "Age Sensitivity of Face Recognition Algorithms." 2013

Fourth International Conference on Emerging Security Technologies (2013): n. pag. Web.

20