1

Face Recognition Using

Face-ARG Matching

Bo-Gun Park, Kyoung-Mu Lee, Member, IEEE, and Sang-Uk Lee, Member, IEEE

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE

INTELLIGENCE, VOL. 27, NO. 12, DECEMBER 2005

Zheng-Wen Shen

2006/04/11

2

Outline

1. Introduction

2. The description of the Face-ARG

3. The Correspondence Graph

4. Similarity between two Face-ARGs and Face Recognition

5. Experimental Results

6. Conclusion

3

1. Introduction

A set of simple lines characterizing the structure of an object are sufficient to identify its shape and recognizable as gray-level images

A face is represented by the Face-ARG model All the geometric quantities and the structural

information are encoded in an Attributed Relational Graph

A set of nodes (line features) and binary relations between them

4

Two-phases

1. The correspondence graph of the reference Face-ARG and the test Face-ARG is constructed using the partial ARG matching algorithm through the stochastic analysis of the feature correspondence in the relation vector space.

2. The stochastic distance between the corresponding relation vector spaces of the extracted subgraphs is evaluated and compared for the identification of a face.

5

An example of the Face-ARG

representation and partial matching

6

2. The description of the Face-ARG

An ordered triple Face-ARG model defined as:

V = {V1,…,VN}

R ={rij}

F ={Ri|i=1,…,N}

Ri = {rij|j=1,..,N, j<>i}

7

3. The Correspondence Graph

Assume that two Face-ARGs, G1 and G2 are given by

8

A Correspondence Graph

A correspondence graph between G1 and G2:

9

A correspondence graph between G1

and G2

10

Block diagram and data flows for the

partial ARG matching process.

11

4. Similarity between two Face-ARGs

and Face Recognition

Similarity between Two Face-ARGs

12

Face Recognition

Select the best matched identity among the database which gives the highest similarity value above some prespecified threshold as the final recognition.

13

5. Experimental Results

For the evaluation of the proposed algorithm’s performance, it was tested on the AR face DB, which is composed of color images of 135 people (76 men and 59 women).

The AR face DB includes frontal view images with different facial expressions, illumination conditions, and occlusion by sunglasses and scarf.

All images used for experiments were normalized to face images of 120 by 170 pixels.

14

The ARG face database

15

Analysis on the Effect of imprecise line

extraction

Imprecise line extraction or noise can affect the overall performance of the line feature-based face recognition algorithms

Position error

Loss of line segments

broken line segments

16

Examples of feature variations due to

the imprecise line extraction. Position error

Missing lines broken lines

17

Position error

18

Missing lines

19

Broken lines

20

Recognition Results on the Real Face DB

21

6. Conclusion