112009/12/162009/12/16

Hierarchical Ensemble of GlobalHierarchical Ensemble of Global and Local Classifiers for Face and Local Classifiers for Face

RecognitionRecognition

Yu Su, Shiguang ShanYu Su, Shiguang Shan, Member, IEEE, Member, IEEE, Xilin Chen, Xilin Chen, , Member, IEEEMember, IEEE, and Wen Gao, and Wen Gao, Fellow, IEEE, Fellow, IEEE

222009/12/162009/12/16

I.INTRODUCTIONI.INTRODUCTION

This paper proposes a novel face recognition method This paper proposes a novel face recognition method which exploits both global and local discriminative which exploits both global and local discriminative features.features.Global features are extracted by 2-D discrete Fourier Global features are extracted by 2-D discrete Fourier transform. transform. Local feature are extracted by Gabor wavelet transform.Local feature are extracted by Gabor wavelet transform.The combination of global and local features plays a key role.Ensemble is also a key contributor to improve generalizability

332009/12/162009/12/16

I.INTRODUCTIONI.INTRODUCTION

Different Roles of Global and Local Features

442009/12/162009/12/162009/12/16 4

II. EXTRACTIONII. EXTRACTION

2-D Discrete Fourier Transformation ( DFT )2-D Discrete Fourier Transformation ( DFT )

A 2-D image of size M by N pixels,

0 ≦ u ≦ M - 1 and 0 ≦ v ≦ N - 1 are frequency variables.

R( u, v ) and I( u, v ) is real and imaginary components.

Extraction of Global Fourier Features

552009/12/162009/12/162009/12/16 5

II. EXTRACTIONII. EXTRACTIONGlobal feature extraction by 2-D DFT Global feature extraction by 2-D DFT

Some examples of inverse transform by using only the low frequency bands (about 30% of all the energy).

The real and imaginary The real and imaginary components named global components named global Fourier feature vector (GFFV).Fourier feature vector (GFFV).

662009/12/162009/12/162009/12/16 6

II. EXTRACTIONII. EXTRACTIONExtraction of Local Gabor FeaturesExtraction of Local Gabor Features

• Gabor Wavelet Transform ( GWT)

Gabor wavelet consists of a planar sinusoid multiplied by a 2-D Gaussian.

772009/12/162009/12/162009/12/16 7

II. EXTRACTIONII. EXTRACTIONGabor features are grouped into a number of feature Gabor features are grouped into a number of feature vectors named local Gabor feature vector (LGFV)vectors named local Gabor feature vector (LGFV)

882009/12/162009/12/16

II. EXTRACTIONII. EXTRACTIONPatch Selection via Greedy Search

in this paper, we propose a patch selection method to automatically determine the positions and sizes of the local patches

992009/12/162009/12/162009/12/16 9

Two layers of ensemble:

III. COMBININGConstruction of Hierarchical Ensemble Classifier

Li : Local Gabor feature vector (LGFV) CLi : Local Component Classifier (LCC) CL : Local Ensemble Classifier (LEC) CG : Global Classifier (GC) CH : Hierarchical Ensemble Classifier (HEC)

10102009/12/162009/12/162009/12/16 10

III. COMBINING

Firstly, the face images can be divided into two classes named intrapersonal pairs and interpersonal pairs.

Weight Learning for Component Classifiers

Secondly, for each image pair, a similarity vector can be obtained.

Last step, two classes of the N–dimensional samples are fed into FLD to get an optimal linear projection from N-D to 1-D.

11112009/12/162009/12/16

IV. RESULTDifferent Roles of Global and Local Features

Gabor feature are more sensitive to the detailed local variations.

12122009/12/162009/12/16

IV. RESULTThe performance improvement becomes trivial when the number of LCCs exceeds 30.

The performance of LEC is much better than that of the individual LCC (especially on Experiment 4).

13132009/12/162009/12/16

IV. RESULTHierarchical Ensemble Classifier

14142009/12/162009/12/16

ThanksThanks