You are on page 1of 4

4

th
Student Conference on Research and Development (SCOReD 2006), Shah Alam, Selangor, MALAYSIA, 27-28 June, 2006
Human Face Recognition using Zernike
moments and Nearest Neighbor classifier
Noraini A. J.
1
, P. Raveendran
1
, N. Selvanathan
2
1
Department of Electrical Engineering, Faculty of Engineering University Malaya 50603, Kuala Lumpur
2
Department of Artificial Intelligence, Faculty of Computer Science and Information Technology
University Malaya 50603, Kuala Lumpur, Malaysia
Abstract - This paper presents a human face recognition
system using an orthogonal invariant moment namely Zernike
moment (ZM) as a feature extractor and a simple Euclidean
square distance classifier or Nearest Neighbor. Changes in
illumination condition, pose, facial expression and others are
challenging task in recognizing face images. Solving these
problems requires a feature extractor that can generate
distinct features for each class of image and a classifier that
able to recognize and classify the face image precisely. The
experiments utilized database face images from Olivetti
research laboratory (ORL) consisting of 40 subjects of 10
images each where none of them are identical [1]. They vary
in position, rotation, scale, expression, with and without
glasses. The performance of the classification depends on the
moment order and classification error is observed below 10%.
Index terms - Euclidean square distance, orthogonal
invariant moment, Zernike moments, nearest neighbor
I. INTRODUCTION
Face recognition system is categorized as biometrics
similar to other recognition applications such as thumb
print, hand-writing, iris scanning and many others. The
used of moment invariant such as ZM moment in face
recognition has long been research way back early nineties
and now it is still being research although many of the
latest work are either the improvement of existing
techniques or hybrid techniques. Since it is orthogonal in
nature and rotational invariant that makes it not sensitive to
noise and has proven to be superior in terms of their feature
representation capability [2]. The face recognition systems
have wide range of application such as access control
systems, content-based video browsing, building or office
security, criminal identification and authentication in secure
systems like computers or bank teller machines [3]. To
build an automated system that can accomplish the above
mentioned objectives is not an easy task. Constraints such
as poses, illumination conditions, facial expressions, aging
and many others are still the main problem to achieve high
classification accuracy. However, a successful face
recognition system depends heavily on the particular choice
of feature extraction methods.
Irrespective of the method used, extracted features
must minimize the within-class face variability and
maximize between-class face variability in order to provide
sufficient discrimination among different faces [3]. The
application of ZM as feature extractor in face recognition
has interest several researchers such as A. Saradha et al,
who combines Fourier descriptors with ZM and classifies
the face images using Linear Discriminant Analysis (LDA)
[4]. J. Haddadnia et al utilizes ZM on localized and
segmented faces and classifies using Radial Basis function
neural network [5], and N.H.Foon et al combines Wavelet
Transforms (WT) and ZM as a feature vector using
Euclidean distance as the classifier [6].
Besides that, other orthogonal moment based feature
extraction methods such as, Lengendre moments, Pseudo
Zernike moments, Krawtchouk moments and others have
also gain attention lately. They have proven to be suitable
for handling images with binary patterns such as pattern
recognition, palm print verification and etceteras [7].
These moments acquire the characteristic of translation,
scale and rotation invariance and thus can be chosen for
image analysis and pattern recognition application [7].
Other feature extractors like PCA, Fourier descriptors and
others have also gain attention and use as a hybrid feature
to represent faces.
In this paper the classification results from the
experiments utilizing feature vectors extracted from ZM
with varying order is presented and analyzed. The data
input consists of 40% train images and 60% test images.
The rest of the paper is organized as follows. Section 2
describes the theory of ZM. Section 3 detailed out the
classifier used for face recognition. Experimental results
are presented in section 4. Section 5 analyses the
performance of the feature extractors and finally section 6
concludes the paper.
II. ZERNIKE MOMENTS
The kernel of Zernike moments are orthogonal
Zernike polynomials defined over the polar coordinates
inside a unit circle. The Zernike moment of order p is
defined as [].
u u u
t
t
d dr r r f r V
p
Z
pq pq
) , ( ) , (
) 1 (
2
0
1
0
) )
-
+
= (5)
where, V
pq
(r,u) denote Zernike polynomials of order p and
repetition q and is written as
u
u
jq
e r
pq
R r
pq
V ) ( ) , ( = (6)
1-4244-0527-0/06/$20.00 2006 IEEE 120
4
th
Student Conference on Research and Development (SCOReD 2006), Shah Alam, Selangor, MALAYSIA, 27-28 June, 2006
while - denotes complex conjugate. The radial polynomial,
R
pq
(r) is expressed as
( )
s p
r
q s p q s p
s
s p
A
A x
s
q p
s
pq
R
2
!
2
2
!
2
2
!
)! (
1
2 /
0

|
|
.
|

\
|
|
|
.
|

\
| +

=
_

=
=
|
.
|

\
|
(7)
where p is a non-negative integer, and q is an integer such
that q p is even, and p q s . The discrete
approximation of the continuous Zernike integral of
equation (1) is written as follows
1 0
), , (
) (
) , (
1
0
1
0
^
< <

=
_ _

ij
N
i
N
j
q j
ij pq
pq
r for
j i f
e r R
N p Z
ij
u

(8)
where, (p,N) is the normalizing constant based on the
mapping transformation [8].
Zernike moment is used as the feature extractor
whereby the order is varied to achieve the optimal
classification performance.
III. NEAREST NEIGHBOR CLASSIFIER
Nearest-neighbor classification method has a very
wide application and is regard as the simplest and powerful
method. It is used to compare the feature vector of the
prototype image and the feature vectors stored in the
database. This is obtained by calculating the Euclidean
square distance between the prototype image and the
database. For instance let C
1
, C
2
, C
3
C
k
be the K clusters
in the database. The class is found by measuring the
distance d(x
(q)
, k) between x
(q)
and the K th cluster C
k
.
The feature vector with minimum difference is found to be
the nearest matching vector [4]. The expression for the
minimum distance is given by
{ }
k
q
k
q
C x x x C x d e = : min ) , (
) ( ) (
(9)
In the experiments, the square distance is computed to
determine the minimum difference that best match the
feature vector of the stored database.
IV. EXPERIMENTAL RESULTS
The experiments were conducted on various numbers
of subjects from ORL database of 40 subjects where each
subject consists of 10 different orientations of the images.
Each image is of size 92x112 and to reduce the complexity
in computation, the images are resized to 64x64. Figure 1
shows the sample subject in ORL database with different
orientation used in the experiments. They are selected at
random where 4 from each subject are considered as train
images and the other 6 as test images.
Fig. 1: Some of the face images from ORL database
Table 1(a) and 1(b) show the various order considered
in the experiments based on 5 subjects and 20 subjects
respectively. The train images are image 1 to image 4.
TABLE 1(a)
% PERFORMANCE OF VARIOUS ORDERS (5 SUBJECTS)
order
No. of features
Class. accuracy (%)
8
25
83.33
10
36
83.33
12
49
90.00
14
64
96.67
16
81
96.67
2 to 8
23
93.33
2 to 10
34
93.33
2 to12
47
100.00
2 to 14
62
100.00
2 to 16
79
100.00
TABLE 1(b)
% PERFORMANCE OF VARIOUS ORDERS (20 SUBJECTS)
Order
No. of features
Class. accuracy (%)
2 to 8
23
88.33
2 to 10
34
89.17
2 to 12
47
91.67
2 to 14
62
91.67
2 to 16
79
91.67
1-4244-0527-0/06/$20.00 2006 IEEE 121
4
th
Student Conference on Research and Development (SCOReD 2006), Shah Alam, Selangor, MALAYSIA, 27-28 June, 2006
Table 2 and 3 listed some of the feature descriptors of the
sample image shown in Figure 1 that include moment order
0 and 1 and exclude moment order 0 and 1 respectively.
TABLE 2
ZM DESCRIPTORS OF SAMPLE IMAGE
(INCLUDE MOMENT ORDER 0 AND 1)
TABLE 3
ZM DESCRIPTORS OF SAMPLE IMAGE
(EXCLUDE MOMENT ORDER 0 AND 1)
The selection of the order at which ZM performs the best is
experimented by considering first, all the moments starting
from 0 and then omitting moments 0 and 1. It is observed
that classification accuracy is better when moments 0 and 1
are disregard. Moments of orders 2 to 10 and 2 to 12 have
shown somewhat optimal performance as in Table 1(a) and
1(b). Based on these results, the rest of the experiments as
the number of subjects is increased only consider ZM of
order 2 to 10 and 2 to 12 for comparison in terms of
percentage classification accuracy. Table 4 and Table 5
show the results obtained from the experiments considering
order 2 to 10 and 2 to 12 respectively. The number of
subjects considered in these experiments is 5, 20, 25, 30
and 40. Experiments conducted on 5 to 25 subjects do not
include those wearing glasses while 30 to 40 include those
with glasses but they are all subjected to constraints as
mentioned earlier. The experiments are not restricted to
one set of train images but the train images are picked at
random as tabulated in Table 4 and 5 respectively. The
number of subjects wearing glasses in the experiment of 30
subjects is 5 and increased to 15 when 40 subjects are
considered.
V. PERFORMANCE ANALYSIS
Table 1(a) shows the performance of ZM when
moments 0 and 1 are considered and disregard respectively.
The classification accuracy is very much bias to ZM with
moments 0 and 1 omitted. The reason being moments at
lower order are basically noise and this may creates
confusion in face recognition.
Comparing the results obtained in Table 4 and 5, the
best performance of 5 subjects is similar for both orders 2
to 10 and 2 to 12 that is 100%. At 20 subjects, order 2 to
10 has overall classification accuracy at 94.17 % while
order 2 to 12 is at 96.67% that is 2.5% higher. When 25
subjects are considered, classification accuracy of 94.67%
at order 2 to 10 and 96% at order 2 to 12 are achieved.
This shows that the performance of order 2 to 12 is 1.33 %
higher than order 2 to 10. 5 subjects with glasses are
included in the experiment with 30 subjects. The best
classification accuracy obtained for order 2 to 10 and 2 to
12 are 94.44% and 96.11% respectively. Order 2 to 12
outperforms by 1.67%. The number of subjects with
glasses is increased to 15 at 40 subjects for both orders.
Order 2 to 10 shows 90.83% classification accuracy while
order 2 to 12 shows 92.5% classification accuracy that is
1.67% higher. However, for both orders the classification
error is below 10%.
Looking at the results above, it is pretty obvious that
order 2 to 12 performs better compared to order 2 to 10. It
is also observed that the higher the order of moments the
higher the classification accuracy since higher order
moments confine better and has more detail information
about the face image [6].
VI. CONCLUSION
Orders 2 to 10 and 2 to 12 of ZM possess the best
features that optimally illustrate the face. The selection of
which order to consider that can give the best overall
classification accuracy is very subjective since it is very
Moment
Order Feature img1_c5 img2_c5 img3_c5 img4_c5
0,0 1 0.568217 0.589011 0.622875 0.513356
1,0 2 0.951478 0.913442 0.737145 0.969321
2,2 3 0.322923 0.460294 0.479715 0.366224
2,0 4 0.124056 0.210402 0.26807 0.073281
3,3 5 0.678247 0.706593 0.647946 0.752306
3,1 6 0.621506 0.583958 0.472364 0.627314
4,4 7 0.972441 0.94563 0.905653 0.946786
4,2 8 0.153475 0.229023 0.271778 0.132799
4,0 9 0.369332 0.265661 0.22445 0.354366
5,5 10 0.721863 0.750831 0.72899 0.695351
5,3 11 0.476047 0.539621 0.454635 0.580448
5,1 12 0.313247 0.374939 0.317518 0.262082
6,6 13 0.481687 0.545315 0.4615 0.480398
6,4 14 0.853394 0.821301 0.798999 0.799079
6,2 15 0.608804 0.248813 0.322386 0.658696
6,0 16 0.566112 0.602494 0.572533 0.537671
Moment
Order Feature img1_c5 img2_c5 img3_c5 img4_c5
2,2 1 0.322923 0.460294 0.479715 0.366224
2,0 2 0.124056 0.210402 0.26807 0.073281
3,3 3 0.678247 0.706593 0.647946 0.752306
3,1 4 0.621506 0.583958 0.472364 0.627314
4,4 5 0.972441 0.94563 0.905653 0.946786
4,2 6 0.153475 0.229023 0.271778 0.132799
4,0 7 0.369332 0.265661 0.22445 0.354366
5,5 8 0.721863 0.750831 0.72899 0.695351
5,3 9 0.476047 0.539621 0.454635 0.580448
5,1 10 0.313247 0.374939 0.317518 0.262082
6,6 11 0.481687 0.545315 0.4615 0.480398
6,4 12 0.853394 0.821301 0.798999 0.799079
6,2 13 0.608804 0.248813 0.322386 0.658696
6,0 14 0.566112 0.602494 0.572533 0.537671
1-4244-0527-0/06/$20.00 2006 IEEE 122
4
th
Student Conference on Research and Development (SCOReD 2006), Shah Alam, Selangor, MALAYSIA, 27-28 June, 2006
dependent on the choice of database that is either the
database has less constraints or otherwise, the combination
techniques employed and choice of the classifiers. In
general, the performance of ZM as a feature extractor
requires a supporting technique that able to improve the
overall performance of the face recognition system [4-6].
TABLE 4
% PERFORMANCE AT ORDER 2 TO 10
Class. accuracy(%)
Train images image1-image4 image5-image8 image9,10,1,2 image3_image6 image7-image10
No of subjects
5(50images)w/o glass 93.33 100.00 100.00 100.00 100.00
20(200images)w/o glass 89.17 94.17 92.50 87.50 91.67
25(250images)w/o glass 87.33 94.67 90.00 87.33 90.67
30(300images) w/glass 87.78 94.44 91.67 88.33 91.11
40(400images)w/glass 80.83 90.83 88.75 84.17 83.75
TABLE 5
% PERFORMANCE AT ORDER 2 TO 12
Class. accuracy (%)
Train images image1-image4 image5-image8 image9,10,1,2 image3_image6 image7-image10
No of subjects
5(50images)w/o glass 100.00 100.00 100.00 100.00 100.00
20(200images)w/o glass 91.67 95.83 95.83 90.83 96.67
25(250images)w/o glass 89.33 94.67 96.00 90.67 94.67
30(300images) w/glass 89.44 95.00 96.11 90.00 94.44
40(400images)w/glass 84.17 91.25 92.50 87.08 83.75
This can be done either looking in terms of
preprocessing of the images before extracting the features
using ZM or combining other feature extraction methods.
However the idea of how to select the order at which ZM
performs best is presented and this helps to make
consideration which order that can be utilized effectively
for further processing and thus developing an effective face
recognition system. Lots of factors need to be studied to
achieve high classification accuracy and able to develop a
robust face recognition system.
VII. REFERENCES
[1] Olivetti Research Laboratory,
http://www.uk.research.att.com:pub/data/att_faces.zip
[2] Belkassim S. O, Pattern Recognition with moment invariants-
Acomparative study and new results, Pattern Recognition, Vol. 24
No. 12(1991), pp1117-1138.
[3] Javad Haddadnia et-al, Neural Network Based face Recognition
with moment invariants, IEEE transaction, 2001.
[4] A. Saradha, S. Annandurai, A Hybrid Feature Extraction Approach
for Face Recognition
Systems, International Journal on Graphics, Vision and Image
Processing, V5, 2005.
[5] J. Haddadnia, K. Faez, Human Face
Recognition with Moments Invariant, Proceeding on NSIP 2001,
Conference NSIP 2001
6] N.H.Foon et al An efficient Method for Human Face Recognition
using Wavelet Transform and Zernike Moments, Proceedings of
the International
Conference on Computer Graphics, Imaging and Visualization
(CGIV04).
[7] Y.C.Chin, A.A.Kassim, Y.Ibrahim,Character recognition using
statistical moments, Image Vision computing, Vol. 17, pp 299-
307, 1999.
[8] R. Mukundan, K.K.Ramakrishnan,
Fast computation of Legendre and Zernike
moments, Pattern Recognition, Vol. 28, No. 9, pp1433-1442, 1995.
1-4244-0527-0/06/$20.00 2006 IEEE 123

You might also like