face recognition on the morph-ii databasepeople.uncw.edu/chenc/stt592_deep learning/2017 nsf... ·...

26
Face Recognition on the MORPH-II Database Face Recognition on the MORPH-II Database Morgan Ferguson University of North Carolina at Wilmington July 25, 2017 1 / 26

Upload: others

Post on 01-Aug-2020

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Face Recognition on the MORPH-II Database

Morgan Ferguson

University of North Carolina at Wilmington

July 25, 2017

1 / 26

Page 2: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Overview

Introduction/ BackgroundMorph-II DatabaseFace Recognition

Our Work with Face RecognitionGoalsMethodsResultsAnalysis

ConclusionFuture ResearchReferences

2 / 26

Page 3: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Introduction/ Background

Morph-II Database

Morph-II Database

I Contains 55,134 mugshots of 13,617 individuals

I Collected over 5 years

I Ages range from 16 to 77 years

I Average of around 4 pictures per individual

I Provides race, gender, date of birth, date of arrest, age, agedifference since last picture, subject identifier, and picturenumber for each picture in the database

3 / 26

Page 4: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Introduction/ Background

Morph-II Database

Challenges and Improvements

I Original data yields poor accuracy ratesI Images have different sizes and faces are in different locationsI Pre-processed images!I Standardized data

Original

Modified

4 / 26

Page 5: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Introduction/ Background

Face Recognition

What is face recognition?

I Process of identifying a new face as a known individual orunknown individual

I Face recognition works by training some classifier on a set ofimages (training or gallery) and then matching new image(test or validation)

Figure: http://www.nec.com/en/global/solutions/safety/face_recognition/index.html

5 / 26

Page 6: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Introduction/ Background

Face Recognition

Eigenfaces

I Face images are projected onto a feature space (”face space”)I Face space is defined by ”eigenfaces”, or the eigenvectors of

the set of facesI Ability to learn to recognize new faces in unsupervised manner

Figure: M. Turk, A. Pentland, ”Eigenfaces for Recognition”, J. Cognitive Neuroscience, vol. 3, no. 1, 1991.

6 / 26

Page 7: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Introduction/ Background

Face Recognition

Fisherfaces

I Linear Discriminant Analysis (LDA)

I Maximizes distance between classes

I Supervised technique

Figure: http://www.scholarpedia.org/article/Fisherfaces

7 / 26

Page 8: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Introduction/ Background

Face Recognition

Local Binary Patterns (LBP) Review

I Different approach to obtaining vectors from an image

I Labels pixels of an image by analyzing the neighbors of eachpixel and considers the result as a binary number

I LBP’s have parameters such as block size and radius

Figure: http://www.scholarpedia.org/article/Local_Binary_Patterns

8 / 26

Page 9: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Introduction/ Background

Face Recognition

Classification and Distance Metrics

Support Vector Machine(SVM)

I Radial basis

Nearest Neighbor approach

I Euclidean distance

I Cosine distance

I Cityblock distance

I Bray-Curtis distance

I Canberra distance

I Mahalanobis Cosine distance

Figure: http://www.researchgate.net/figure/5423571_fig6_Figure-13-313-Illustration-of-distance-measures

9 / 26

Page 10: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Our Work with Face Recognition

Goals

Goals

Reproduce Results

I Compare benchmark results with previous research

New TrialsI Test out new subsets, distance metrics, feature vectors, etc.

Optimize

I Try out different methods to improve on initial results

Analyze

I Start creating tables and graphs to compare results

I Begin to make conclusions

10 / 26

Page 11: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Our Work with Face Recognition

Methods

Experimental Setup

1. Subset Morph-II

2. Begin with input data (pre-processed images from subsets orLBP feature vectors)

3. Break into training and testing data

4. Perform PCA (Eigenfaces) or LDA (Fisherfaces)

5. Use SVM or Nearest Neighbor

6. Classify test image as a subject from training images

11 / 26

Page 12: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Our Work with Face Recognition

Methods

Subset Scheme 1

I Take subjects with more than 10 images each

I Randomly select 10 images for each person

I Match the number of males with the number of females (83subjects each)

I Randomly select 1 image for each subject as testing imageand designate other 9 for training

I 166 subjects and 1660 total images

12 / 26

Page 13: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Our Work with Face Recognition

Methods

Subset Scheme 2

I Take subjects with more than 10 images each

I Randomly select 10 images for each person

I Randomly select 5 images for each subject as testing imageand designate other 5 for training

I 544 subjects and 5440 total images

13 / 26

Page 14: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Our Work with Face Recognition

Results

Initial Subset 1 Results

Table: Accuracy Rates and Run Times for Face Recognition Algorithms(before histogram equalization)

SVM-R Euclidean CityBlock Cosine

Eigenfaces(PCA)

Accuracy 74.1% 50.6% 66.9% 59.6%Run Time

(sec)177.3 5.2 3.9 8.6

Fisherfaces(LDA)

Accuracy 86.14% 87.3% 83.1% 95.8%Run Time

(sec)165.2 7.7 6.1 13.3

14 / 26

Page 15: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Our Work with Face Recognition

Results

More Subset 1 Results

Table: Accuracy Rates and Run Times for Face Recognition Algorithms(after histogram equalization)

SVM-R Euclidean CityBlock Cosine

Eigenfaces(PCA)

Accuracy 88.6% 69.9% 78.9% 71.1%Run Time

(sec)211.3 7.4 3.9 10.0

Fisherfaces(LDA)

Accuracy 89.2% 92.8% 89.2% 95.8%Run Time

(sec)179.4 10.8 7.3 12.8

15 / 26

Page 16: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Our Work with Face Recognition

Results

LBP Results (PCA)

16 / 26

Page 17: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Our Work with Face Recognition

Results

LBP Results (LDA)

17 / 26

Page 18: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Our Work with Face Recognition

Results

Overall Subset 1 Results

Table: Face Recognition Accuracy Rates on MorphII Subset 1: Train on9, Test on 1

SVM-R (%) Euclidean (%) CityBlock (%) Cosine (%)

Eigenfaces(PCA)

88.6 69.9 78.9 71.1

Fisherfaces(LDA)

89.2 92.8 89.2 95.8

LBP + PCA87.3

(s=10, r=1)71.7

(s=10, r=1)69.3

(s=12, r=1)76.5

(s=10, r=1)

LBP + LDA88.6

(s=10, r=2)84.9

(s=14, r=2)81.3

(s=14, r=2)89.2

(s=14, r=3)

18 / 26

Page 19: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Our Work with Face Recognition

Results

Back to the Data

Truth

Prediction

19 / 26

Page 20: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Our Work with Face Recognition

Results

Subset 2 Results

Table: Accuracy Rates and Run Times for Face Recognition Algorithmsusing 5 training images, 5 testing

Euclidean CityBlock Cosine BrayCurtis Canberra

Eigenfaces(PCA)

Accuracy (%) 54.0 63.1 56.0 69.7 66.0Run Time

(sec)139.2 78.8 268.5 121.6 223.9

Fisherfaces(LDA)

Accuracy (%) 62.9 55.9 78.2 71.7 51.9Run Time

(sec)163.7 115.3 281.0 146.1 255.8

20 / 26

Page 21: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Our Work with Face Recognition

Results

Subset 1 vs. Subset 2...What Happened?

SVM-R Euclidean CityBlock Cosine BrayCurtis Canberra

Eigenfaces(PCA)

Accuracy (%) 88.6 69.9 78.9 71.1 79.5 79.5Run Time

(sec)211.3 7.4 3.9 10.0 6.8 9.6

Fisherfaces(LDA)

Accuracy (%) 89.2 92.8 89.2 95.8 94.8 83.1Run Time

(sec)179.4 10.8 7.3 12.8 8.5 13.3

Euclidean CityBlock Cosine BrayCurtis Canberra

Eigenfaces(PCA)

Accuracy (%) 54.0 63.1 56.0 69.7 66.0Run Time

(sec)139.2 78.8 268.5 121.6 223.9

Fisherfaces(LDA)

Accuracy (%) 62.9 55.9 78.2 71.7 51.9Run Time

(sec)163.7 115.3 281.0 146.1 255.8

21 / 26

Page 22: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Our Work with Face Recognition

Analysis

Size Analysis

22 / 26

Page 23: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Our Work with Face Recognition

Analysis

Gender Analysis

23 / 26

Page 24: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Conclusion

Future Research

Future Research

I How can we optimize the face recognition algorithm further?

I How does race affect face recognition?

I How much would gender and race classification as a first stepincrease accuracy rates and reduce run time?

24 / 26

Page 25: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Conclusion

References

References

G. Guo, G. Mu, K. Ricanek (2010)

Cross-age face recognition on a very large database: the performanceversus age intervals and improvement using soft biometric traits

20th International Conference on Pattern Recognition 2010 , 3392-3395.

DK Hayati PHM Yassin, S.Hoque and F. Deravi (2013)

Age sensitivity of face recognition algorithms

Proc of the Fourth International Conference on Emerging SecurityTechnologies (EST), 2013

M. Turk, A. Pentland (1991)

”Eigenfaces for Recognition”

J. Cognitive Neuroscience , vol. 3, no. 1, 1991.

K. Ricanek Jr. and T. Tesafaye (2006)

”MORPH: A Longitudinal Database of Normal Age-Adult Progression”

IEEE 7th International Conference on Automatic Face and GestureRecognition (FGR06) , Southampton, UK, Apr. 2006, pp. 341345.

25 / 26

Page 26: Face Recognition on the MORPH-II Databasepeople.uncw.edu/chenc/STT592_Deep Learning/2017 NSF... · Face Recognition on the MORPH-II Database Introduction/ Background Morph-II Database

Face Recognition on the MORPH-II Database

Conclusion

References

The End

26 / 26