FACE RECOGNITION

PROJECT

Outline
1. Face image cropping and preprocessing
2. Appearance-based face recognition methods
Principal Component Analysis (PCA)
Linear Discriminant Analysis (LDA)
Locality Preserving Projections (LPP)
K-nearest neighbor as the classification method

3. Experimental Results

4. Demonstration

Face image cropping and preprocessing

Face Detection with OpenCV Library
Normalization
Cropping
Grayscale conversion
Fixing of locations of eyes (by OpenCV)
Fixing to the same resolution
Preprocessing
Histogram Equalization

Our Dataset

Appearance-based methods
PCA
dimensionality reduction method
produces a compact representation
LDA
supervised learning algorithm
find a subspace which separate different classes of objects
LPP
considers the manifold structure for face analysis

1. Training

2. Recognizing

Reference: [1]

Eigenfaces, Fisherfaces & Laplacianfaces

PCA
Eigenfaces

LDA

Fisherfaces

LPP

Laplacianfaces

PCA Reconstruction

Original Face

d=2

d=6

d = 10

d = 20

d = 40

d =60

Original Face

d=2

d=6

d = 10

d = 20

d = 40

d =60

Experimental Results
1. PCA with different energy percentage
2. Different metric types
3. Different partitions for validation

4. Different K for K-nearest neighbor

5. LPP with different PCA processing

PCA with different energy percentage

1
0.9
0.8
0.7
0.6

Accuracy

0.5

own dataset
0.4

ORL dataset

0.3
0.2
0.1
0
0.5

0.6

0.7

0.8

0.9

0.91

0.92

0.93

0.94

0.95

0.96

0.97

0.98

0.99

Energy Percentage

Different metric types

Accuracy

Accuracy
1

0.8

0.9

0.7

0.8
0.6
0.7
0.5

0.6
0.5

PCA

0.4

LDA

LPP

0.3

PCA
0.4

LDA

0.3

LPP

0.2

0.2
0.1

0.1

Cosine
dissimilarity

L1 distance

ORL Dataset

Euclidean
distance

Cosine Dissimilarity

L1 Distance

Our Dataset

Euclidean Distance

Different partitions for validation

Accuracy

Accuracy

0.9

0.9

0.8

0.8

0.7

0.7
0.6
0.6
PCA
0.5

LDA
LPP

0.4

0.5
PCA

0.4

LDA
LPP

0.3
0.3
0.2

0.2
0.1

0.1

0
3

Training Number

ORL Dataset

Training Number

Our Dataset

Different K for K-nearest neighbor

Accuracy

Accuracy

0.9

0.9

0.8

0.8
0.7
0.7
0.6
0.6
PCA
0.5

0.5

PCA

LDA
LPP

0.4

LDA

0.4

LPP
0.3

0.3

0.2

0.2

0.1

0.1
0
1

ORL Dataset

KNN_K

0
1

Our Dataset

KNN_K

LPP with different PCA processing

Accuracy
0.95

0.9

0.85

LPP
PCA
0.8

0.75

0.7
40

60

80

100

Number of
Components

Conclusion
PCA performs the best on ORL dataset
LDA performs the best on our dataset
Out dataset includes lots of variations
LPP does not perform as expected
This method still could achieve good results with proper setting
A more complex method calculating weight matrix may help

References
[1] Delac, K., Grgic, M., & Grgic, S. (2005). Independent comparative study of
PCA, ICA, and LDA on the FERET data set. International Journal of Imaging
Systems and Technology, 15(5), 252-260.
[2] Turk, M., & Pentland, A. (1991). Eigenfaces for recognition. Journal of
cognitive neuroscience, 3(1), 71-86.
[3] Belhumeur, P. N., Hespanha, J. P., & Kriegman, D. J. (1997). Eigenfaces vs.
fisherfaces: Recognition using class specific linear projection. Pattern Analysis
and Machine Intelligence, IEEE Transactions on, 19(7), 711-720.
[4] He, X., Yan, S., Hu, Y., Niyogi, P., & Zhang, H. J. (2005). Face recognition
using laplacianfaces. Pattern Analysis and Machine Intelligence, IEEE
Transactions on, 27(3), 328-340.
[5] Viola, P., & Jones, M. J. (2004). Robust real-time face detection.
International journal of computer vision, 57(2), 137-154.