Irfan Ullah Department of Information and Communication Engineering Myongji university, Yongin, South Korea

parathyroidsanchovyΤεχνίτη Νοημοσύνη και Ρομποτική

17 Νοε 2013 (πριν από 3 χρόνια και 4 μήνες)

75 εμφανίσεις

Irfan

Ullah

Department
of Information
and
Communication Engineering

Myongji

university,
Yongin
, South Korea

Copyright © SunLightron (sl.avouch.org)


Introduction


Objective


Overview of proposed method


Eigenface recognition in clutter


Background representation


Classifier


Proposed method


Experimental results


Conclusions


Face

recognition

is

quite

a

difficult

task

because

faces

are

a

natural

class

of

complex,

multidimensional

objects
.



Fisher’s

linear

discriminant

(FLD)

and

Eigenface

recognition

(EFR)

methods

are

quit

well

when

input

test

patterns

is

a

face
.


EFR



If

the

threshold

is

set

high
,

it

ends

up

missing



If

the

threshold

is

lowered

to

capture

the

face,

it

gives

many

false

alarms



It

is

quite

sensitive

to

the

choice

of

the

threshold

value
.

Good

face

recognition

system



Detect

and

recognize

all

faces

in

a

scene



Not

missclassify

background

patterns

as

faces


Precautions



Few

false

alarms

will

render

the

system

ineffective



performance

should

not

be

too

sensitive

to

any

threshold

selection
.




Distance

from

eigenface

space

(DFFS)

and

distance

in

eigenface

space

(DIFS)

are

suggested

to

detect

and

eliminate

nonfaces

for

robust

face

recognition

in

clutter
.



We

show

that

these

are

not

sufficient

to

discriminate

against

arbitrary

background

patterns

in

the

absence

of

any

information

about

the

background
.


To handle clutter in still images requires


Good face detection

module to find face patterns


And feed only those patterns as input to traditional
EFR scheme


Within

Principal

component

analysis

(PCA)

to

robustly

recognize

faces

in

the

presence

of

clutter
.



Traditional

eigenface

recognition

(EFR)

method,

which

is

based

on

PCA
,

works

quite

well

when

the

input

test

patterns

are

faces
.



But

poor

when

recognize

faces

appearing

against

a

background



It

May

miss

faces

or

may

wrongly

associate

many

background

image

patterns

to

faces



To

remove

this,

learning

the

distribution

of

the

background

is

helpful


1.
Construct

an


eigenbackground

space


which

represents

the

distribution

of

the

background

images

corresponding

to

the

given

test

image
.



2.
The

background

is

learned


on

the

fly



3.
Provides

a

sound

basis

for

eliminating

false

alarms
.


4.
An

appropriate

pattern

classifier

is

derived


5.
Eigenbackground

space

together

with

the

eigenface

space

is

used

to

simultaneously

detect

and

recognize

faces
.


Tranning

set of face images


Mean


where L is the total number of training images



PCA

solves for a set of L orthogonal vectors where


Is maximum subject to


U
n

and
λ
n

are the eigenvectors and eigenvalues of


where


Weight vector corresponding to pattern T
i





The distance in face space (DIFS) is




Minimum DIFS
is declared as
recognised face


Distance from free space (DFFS) is defined as


If then test pattern is
nonface image


If threshold is the
smallest
,

real faces are not missed, but
many false alarms


If threshold is
smaller

, we will
miss some faces


The threshold for
DFFS and DIFS
need to be
higher
, but then we get
false alarms


Therefore,
EFR

is sensitive to
threshold


Properties of background

must be utilized to solve this issue


Nonface

patterns could be confused as face patterns


Learning


Universal background class


Background distribution local to a given test image


Utilization of face and background distributions


Reduce false alarms


Decrease sensitivity to the choice of threshold


A window pattern in the test image is classified (positively)
as a
background pattern

if its
distance

from the
eigenface

space
is greater than a certain (high) threshold



Background patterns are distributed into K clusters


Few clusters result
under
-
representation

of background class


Too many is not possible due to
limited training samples


Pattern centers are few as compared to background
patterns, and are used for learning
eigenbackground

space

1.
Eigenvectors of the covariance matrix of the set of
background pattern centers.

2.
The subspace spanned by the eigenvectors corresponding
to the largest eigenvalues of the covariance matrix is
called the
eigenbackground space
.

3.
Eigenvectors of C
b
is called “
eigenbackgroung images


4.
Image T is converted into eigenbackground components
by


Faceclass (w
1
)

Background class (w
2
)

Weight vector

Estimator for face

Reconstruction error in x

Estimator for background

Reconstruction error in x

X
b

is the estimate of x

Weight

Image pattern is classified as face if




is positive, else not. When , when the number

of eigenfaces and eigenbackground patterns are the same,
and when , i.e., when the arithmetic mean of the
eigenvalues in the orthogonal subspaces is the same


Reconstruction error function


A scheme that
recognizes faces
by searching a given
test image for patches of
image patterns of faces

appearing against a
cluttered background

Stages


Estimation of the
eigenface

space


Construction of the
eigenbackground

space


Recognition

Image recognition


Eigenface

space and the
eigenbackground

space

are learnt
using training images


The
test image is examined
again in the presence of faces at
all points in the image.


At
every pixel
in the test image, a
subimage

is cropped
about
that pixel to obtain the test patterns.


For each of these test window patterns, the
classifier

is used
to determine whether a
pattern is a face or not.





Let
subimage

pattern in the test image is


It is projected onto the
eigenface

and
eigenbackground




where is the threshold


The pattern is recognized as belonging to the
i
th

person if




where
q

is the number of face classes or people in the database


Compute eigenfaces


Identify Prominent Background Images



High threshold, far from eigenface space marked as
background



Calculate Background Pattern Centers


Using K
-
mean algorithm, to reduce background patterns


Obtain Eigenbackground Images


Eigenvectors from highest eigenvalues


Detect and Recognize Faces in the Scene


Detection of face by classifier and DIFS

Fig. 2. Architecture of the proposed system. (a) Computation of eigenfaces. (b)
Construction of eigenbackground space. (c) Face detection and recognition.

Fig
.

3
.

(a)

Test

case

where

a

person

appears

naturally

against

a

cluttered

scene
.

(b)

Results

for

the

traditional

EFR

technique
.

(c)

Results

using

the

proposed

method
.

(d)

Some

of

the

background

pattern

centers

returned

by

the

K
-
means

algorithm
.

(e)

First

eight

eigenbackground

images

for

the

background

local

to

the

test

image
.

(f)

Typical

eigenfaces
.

Fig
.

4
.

(a)

Test

images

with

different

complex

backgrounds
.

Results

for

(b)

traditional

EFR

and

(c)

the

proposed

method
.

Fig
.

5
.

Representative

results

for

the

proposed

method

on

some

more

test

images
.

Fig
.

6
.

(a)

Few

of

the

test

cases

where

the

proposed

method

had

false

alarms
.

(b)

Test

cases

where

the

person

is

not

in

the

training

set
.

Fig
.

7
.

Some

results

for

the

proposed

method

on

outdoor

images
.

(a)

Examples

of

side
-
view

of

faces
.

(b)

Different

illumination

conditions

for

two

individuals
.

(c)

Example

images

containing

several

people

within

the

same

image
.

Fig
.

8
.

Detection

rate

versus

FAR

the

proposed

method

and

the

traditional

EFR

method
.


when the scheme is
directly extended to recognize faces
in the
presence of background clutter, its
performance degrades
as it
cannot satisfactorily discriminate against nonface patterns.



The background space which is created “on the fly” from the
test image is shown to be very useful in distinguishing nonface
patterns.



The scheme gives very
good results
with almost
no false
alarms
, even on fairly complicated scenes.



For background learning,
one must decide the number of
background centers based on resolution of the image.

In order to reduce the global computational complexity of the
algorithm.



Instead of processing each and every pixel,
one could process every
alternate pixel along rows and columns



One could skip processing of some of the pixels in the immediate
neighborhood

of an already identified face.



If people appear against a relatively constant or slowly changing
clutter, background learning need be done either only once or very
infrequently.