BDPCA Plus LDA: A Novel Fast Feature Extraction Technique for Face Recognition 授課教授 :...

Preview:

Citation preview

BDPCA Plus LDA: A Novel Fast Feature Extraction Technique for Face Recognition

授課教授 : 連震杰 老師

組員 : 黃彥綸 何域禎

W. Zuo, D. Zhang, J. Yang, K. Wang, “BBPCA plus LDA: a novel fast feauture extractiontechnique for face recognition,” IEEE Transactions on Systems, Man, andCybernetics, vol. 36, no. 4, pp. 946-953, Aug. 2006.

Outline

• Introduction• Principal Component Analysis (PCA)• Bidirectional Principal Component Analysis(BDPCA)• Image Reconstruction• BDPCA plus LDA Technique• Experiments

Introduction

• Geometric-based approachesFeature detectionHigh recognition rate.Feature location is difference from people

to people.

Introduction

• Holistic-based approachesRobust recognition performance under

noise, blurring, and partial occlusion.EX:PCA(extract eigenfaces) 、 LDA(has SSS

problem)

Q&A

Q:What is small sample size (SSS) problem?A:In LDA the rank of Sw must be ,

then exist. For example, the ORL database image is 112x92 size, Sw :(112x92)x(112x92), and has 40 classifications, it must has 10344 pictures for training, but there is no a so large database.

C

ii CNN

1

1

1wS

Q&A

Q:How to solve the SSS problem?A:Before using LDA, we first do image

dimensionality reduction, such as PCA, BDPCA.

PCA

N

i

Tii xxxx

NSt

1

))((1

)( xxvy Tii PCAdi ,...,1

•Feature extraction:eigenfaces method•Data compression•Image dimensionality reduction•Fail to classfication=>LDA

Flowchart

Find LDA projector

Mapping data to BDPCA subspace

Mapping data to LDA subspace

KNN to get the face recognition rate

TestingTraining

Mapping data to LDA subspace

Mapping data to BDPCA subspace

Find BDPCA projector

Image dimensionality reduction

BDPCA• Bidirectional PCA(BDPCA)

Tmiiii xxxX ...21

N

ii

Ti

rowt XXXX

NmS

1

)()(1

rowk

rowrowr row

wwwW ...21

N

i

Tii

colt XXXX

NnS

1

))((1

colk

colcolc col

wwwW ...21

rTc XWWY

row projection matrix column projection matrix

niiii xxxX ...21

Y: Feature matrix

Image Reconstruction

xyvx ii TrcYWWX

PCA BDPCAOriginal image

Training

Testing

MSE curves

Training Testing

BDPCA plus LDA Technique

)det(

)det(maxarg

wT

bT

S

S

c

i

Tiiib N

NS

1

))((1

Tiji

C

i

N

jijiw yy

NS

i

)()(1

,1 1

,

wb SS Generalized eigendecomposition

yWz TLDA

rTc XWWY Mapping Y into its 1D representation y

mLDAW ,...,2,1

Between-class scatter matrix of y

Within-class scatter matrix of y

The LDA projector

The final feature vector

AdvantagePCA

mxn

(mxn)x1

(mxn)x(mxn)

bS

BDPCA

PCAd

mxn

rowtS

coltS

nxn mxm

colkrowk

Y

rowcol kk

CPU time

On ORL face database

Experiments

• To test the efficacy of BDPCA + LDA, we make use of two face databases, the ORL face database and the FERET database.

• Since our aim is to evaluate the efficacy of feature extraction methods, we use a simple classifier, the nearest neighbor classifier.

Comparisons of the recognition rates

ORL database FERET database

Conclusions

• The BDPCA +LDA has a much faster speed for facial feature extraction.

• The BDPCA + LDA needs less memory requirement because its projector is much smaller than that of the PCA + LDA.

• The BDPCA + LDA has a higher recognition accuracy over the PCA + LDA.

Recommended