Extraction of Region of Interests from Face Images Using Cellular Analysis

Extraction of Region of Interests from Face Images

Using Cellular Analysis

Speaker: Han-ping Cheng

Outline

• Introduction

• Proposed Work

• Results and Discussions

• Conclusion

• Future Works

Introduction

Face recognition system:

1. Face detection

- deals with the problem of face localization

2. Feature extraction

- finds the presence of facial features like eyes,

nose, nostrils etc.

3. Face recognition

- compares an input image against the database

and reports a match, if exists

Introduction

Face localization approachesUtilizing shape information:Ellipse fitting method, Mosaic images, Color information, Facial geometry andsymmetry, etc.

Facial feature extraction techniquesEigenface approach, 2D Gabor wavelets, anddiscrete cosine transform (DCT) based approach

Introduction

Cellular analysis of a face image• A novel algorithm for extracting the ROIs fro

m face images

Algorithm• Adaptive thresholding• Geometric properties of a face

Proposed Work

1. Face Localization

2. Constructing the Cellular Regions

3. Extraction of Regions of Interest

Proposed Work

minimum :min

maximum :max

)min(max2/1

:

:

global

global

factorsmoothing

thresholdadaptive

g

g

gglocal

adpt

prevlocaladpt )1(

Proposed Work

prevlocaladpt )1(

0)(

)1(

)1( ... )1(

... )1(

ni

locali

locallocal

tionclassificailasttheinusedthresholdthe

thresholdinitialthi

local :

:)(

0

Proposed Work

4 3, 2, 1,

minmax

i

adptii

The cell is said to contain a portion of the face. (occupied by the object)

Proposed Work

Let n be the number of cells occupied by the face.Then the following case are possible: i) n = 0: not a vertex of the ROI corresponding to the face ii) n = 1:

iii) n = 2:

iv) n = 3: v) n = 4:

vertex.270 a is otherwise, point; edgean is then occupied, are cellsadjacent twoif

90 angle internalth vertex wia is

vertex270 a is

vertexanot andregion face theinside is

Proposed Work

point edgean is V

vertex270 a is V

vertex90 a is V

2

3

1

The type of the vertex:

Proposed Work

1. Face Localization

2. Constructing the Cellular Regions

3. Extraction of Regions of Interest

Proposed Work

vertex theof type theii)

edgeincident theofdirection thei)

:by determined is vertex

any from edge outgoing The 4.

bottom towardsdirected be will

vfrom edge outgoing The .3

vertex)90 a

be chosen to always is (v

vex,start vert

thebe toassigned is then

20)( ,minmax If 2.

ordermajor

rowin scan ,point each For 1.

00g

start

start

start

g

Proposed Work

1. Face Localization

2. Constructing the Cellular Regions

3. Extraction of Regions of Interest

Proposed Work

• Region containment tree

Proposed Work

vj

Ri

Rj

1-ji1 :R regions, discovered previously the

t withcontainmen itscheck Rj,region neweach For

i

Proposed Work

Region ‘2’ represents the face region

1

2

3 4 5 6

Proposed Work

• Use priori knowledge about the face geometry to extract the ROI:

1. The center of the regions representing the

pair of eyes will approximately lie on the

same horizontal line

2. The center of the nostril region and mouth

region will lie approximately on the vertical

axis that passes through the center of the

face region 5parameter toleranceaConsider *

Proposed Work

• Region Merging(special case)

10xx and 10xx

and 20yy

: trueiscondition following theif

ROI single a tomerged are and

(2)max

(1)max

(2)min

(1)min

)2(max

(1)min

21

RR

Color Images

• RGB format

• Three different threshold

4 3, 2, 1,

minmax

i

adptii

Results and Discussions

• the precision of the extracted ROIs can be

controlled by varying resolution level

*c is the length of the cell

Results and Discussions

Results and Discussions

Results and Discussions

Conclusion

• Cellular representation of the ROIs

• The complexity is controlled by cell size

• Adaptive thresholding mechanism for gr

ay-scale and color image

• Region containment tree

Future Works

1. Compare with other ROI extraction

techniques

2. Designing a face identification system

on the basis of the extracted ROI

Thank you !