8

Th
Internationl Conference on Informatics and Systems (INFOS 2012)

Abstract
This paper proposes a fast and accurate ear
recognition method based on the principal component
analysis (PCA). Conventional PCA has problems such as
time and space complexity when dealing with large size
images. Dividing such images into small blocks and then
applying the PCA algorithm on each block separately can
solve such problems. In this paper, we divide ear images
into non-overlapping blocks. For each part, we extract the
PCA features and match each block using minimum
distance classifier. Then, combine the outputs of the
classifiers in abstract, rank, and score levels.


1. Introduction

Biometric systems use different physiological traits
such as face, iris, fingerprints, ear, finger knuckle, and
palm-print of individuals for recognition purposes. Ear
recognition method is one of the recent biometric
techniques. Ears play a significant role in forensic science
for many years, especially in the United States, where an
ear classification system based on manual measurements
has been developed by Iannarelli [1].
Principal component analysis (PCA) algorithm is
widely used by many researchers to extract features from
ear image [2-5]. Chang et al. [2], compared ear recognition
with face recognition using a standard PCA technique on
face and ear images. Ping Yan et al. [3] reported a good
recognition rate for an “Eigen-ear” PCA-based approach.
Alaa et al. [5] used feature combination to improve the
recognition rate of ear recognition system.
PCA is used in dimension reduction, feature extraction,
image compression, face detection, etc. The computation
of PCA requires eigenvalue decomposition (EVD) of the
covariance matrix of feature vectors [6,7]. However, The
problem of PCA algorithm is the high dimensionality.
Thus, it needs very large space and needs too much central
processing unit (CPU) time.
Many researches tried to decrease the dimensionality
and the CPU time for the PCA. Golub et al. [8] used
Jacobi’s method which diagonalizes a symmetric matrix
requires around O(d
3
+ d
2
n) computations (where n is the
number of feature vectors and d
2
is the size of the vector).
Roweis et al. [9] proposed expectation maximizing (EM)
algorithm for PCA, which is computationally effective
than EVD method for PCA. A method in [10], used to find
leading eigenvector which is less expensive method,
however, it can compute only one most leading
eigenvector [10]. Hall’s et al [11], Skarbek [12], and Liu
et al. [13] developed an algorithm to compute eigenspace
merging and they do not need to store the covariance
matrix of previous training samples.
In this paper, we decrease the dimensionality of the
original image and hence decrease the complexity of the
PCA algorithm through dividing the ear image into small
blocks. The PCA algorithm is applied to these non-
overlapping blocks, the we make a classification using
many classifiers and then combine the output of these
classifiers at abstract, rank, and score level. We compare
the proposed method to some of the state-of-the-art
methods. Experimental results show that the proposed
method is fast and achieve recognition rate better than
using the whole ear image in the state-of-the-art methods.


2. PCA Algorithm

The PCA is a linear subspace method and it finds a
linear transformation u which reduces d-dimensional
feature vectors to h-dimensional feature vectors (where h
< d). It can also reconstruct h-dimensional feature vectors
back to the d-dimensional feature vectors with some finite
error known as reconstruction error. The PCA is mostly
used in compression and reconstruction of high
dimensional feature vectors. Since the transformation is
from d-dimensional feature space to h dimensional feature
space and vice versa the size of u is dxh. The h column
vectors are the basis vectors. The first basis vector is in the
direction of maximum variance of the given feature
vectors. The remaining basis vectors are mutually
orthogonal and, in order, maximize the remaining
variances subject to the orthogonal condition. Each basis

Personal Identification Using Ear Images Based on Fast and Accurate Principal
Component Analysis

Alaa Tharwat

Electrical Engineering Department, Suez
Canal University, Egypt
engalaatharwat@hotmail.com


Abdelhameed Ibrahim, Hesham A. Ali
Computers and Systems Engineering
Department, Mansoura University, Egypt
{afai79, h_arafat_ali}@mans.edu.eg
8
Th
Internationl Conference on Informatics and Systems (INFOS 2012)

vector represents a principal axis. These principal axes are
those orthonormal axes onto which the remaining
variances under projection are maximum. These
orthonormal axes are given by the dominant/leading
eigenvectors (i.e. those with the largest associated
eigenvalues) of the measured covariance matrix. In PCA,
original feature space is characterized by these basis
vectors and the number of basis vectors used for
characterization is usually less than the dimensionality d
of the feature space [12].
For Ear images, let an image Γ(N x N) and represent
every image as a vector of dimensions N
2
. The training set
of ear images Γ
1
, Γ
2
,……, Γ
M
, thus, the average of the
training set is ¢ =
1
M
∑ Γ
I
M
ì=1
. Subtract each image from
the average (Normalization) ¢
ì
= Γ
I
−ψ. Then, create
data matrix A=[Φ
1
Φ
2
Φ
3
…….Φ
M
] (N
2
x M matrix) and
compute the covariance matrix as
C =
1
M
∑ ¢
n
¢
n
1
M
n=1
= AA
1
(1)
Compute the eigenvalus (γ
k
) and eigenvectors (V
k
),
and sort the eigenvectors according to the eigenvalus and
discard eigenvectors that have eigenvalus less than or
equal to zero.

If we want to test unknown ear image Γ (in the same
size of all used ear images) just subtract the mean from the
ear image and then transform it into its "Eigen-ear"
components (projected into ear space) :
¢ = Γ −¢ (2)
0
`
= ∑ u
ì
1
¢
K
ì=1
(3)

Consider a set of 2D ear images (MxN), with K
images in the training set. First the training set are
centered by calculating the mean of each images in
O(KMN) and a data matrix of centered ear images is
created with computational complexity O(KMN). The
complexity of constructing covariance matrix is
O(K(MN)
2
). Then calculate the eigenvalues and
corresponding eigenvectors of the covariance matrix, for
which the computational complexity is O(K(MN)
3
).
Consequently, eigenvectors are sorted in decreasing order
according to corresponding eigenvalues using merge sort
in O(K(MN) log
2
(MN)). Since the first L eigenvalues and
the associated eigenvectors are considered in the
construction of the reduced eigenspace, computation of the
reduced eigenspace is carried out in O(LMN) whereas the
projection of the images into this eigenspace requires
computational complexity of O(KLMN). These
computational complexity expressions are written in order
in the text, gets the following sum and the overall
computational efficiency as [14]; computational
efficiency of PCA= O(KMN)+ O(KMN)+ O(K(MN)
2
)+
O(K(MN)
3
)+ O(K(MN) log
2
(MN))+ O(LMN)+
O(KLMN) = O(K(MN)
3
)

3. Classification Level Fusion Method

Combining classifiers technique achieves good
recognition rates when the classifiers used are
independent. The fusion in this level may be in the
abstract, rank, or score levels.
Abstract (Decision) Level Fusion can be interpreted
as making a decision by combining the outputs of different
classifiers for a test image.
Rank Level Fusion sorts the output of each classifier
(a subset of possible matches) in decreasing order of
confidence so that each class has its own rank. The fusion
can be done through counting the ranks of each class and
the decision will be the class of the highest rank.
Score (Measurement) Level Fusion, apply fusion
rules on the vectors that represents the distance between
the test image and the training images; the output of each
classifier and the decision is the class that has the
minimum value. Assume that, the problem of classifying
an input pattern Z into one of m possible classes based on
the evidence provided by R different classifiers. Let x
i
⃗ be
the feature vector (derived from the input pattern Z)
presented to the i
th
classifier. Let the outputs of the
individual classifiers be P(x
j
|x
i
⃗), i.e., the posterior
probability of the pattern Z belonging to class ω
j
given the
feature vector x
i
⃗. Let c {1, 2, . . . , m} be the class to
which the input pattern Z is finally assigned. The
following rules is used to determine c [15]:
c= arg max
j
max
i
P(ω
j
| x
i
) (4)
c= arg max
j
min
i
P(ω
j
| x
i
) (5)
c= arg max
j
med
i
P(ω
j
| x
i
) (6)
c= arg max
j
avg
i
P(ω
j
| x
i
) (7)
c= arg max
j
Π
i
P(ω
j
| x
i
) (8)


4. Proposed ear recognition method

The proposed method is shown in Figure 1. We divide
each ear image into non-overlapping blocks or parts. For
each part, we extract the PCA features and match each
block using minimum distance classifier. Then, combine
the outputs of classifiers in abstract, rank, and score levels.
In abstract or decision level, we used majority voting
method to combine results from all blocks. And in rank
level we used borda count method. Finally, in score level
fusion we used minimum, maximum, product, mean, and
median to combine scores.
In the proposed approach, we divide the ear images
(MxN) into Q equal blocks. Thus, the block size will be
(MxN/Q). Calculate the mean requires O(KMN/Q) and
center the data requires O(KMN/Q) where K represents
the number of ear images in training set. The complexity
of calculating covariance matrix is O(K(MN/Q)
2
). The
complexity of calculating eigenvalues and eigenvectors
8
Th
Internationl Conference on Informatics and Systems (INFOS 2012)

and sorting it is O(K(MN/Q)
3
) and O(K(MN/Q) log
2

(MN/Q)) respectively. Finally, the computation of reduced
eigenspace and the projection of small blocks on it
requires O(LMN/Q) and O(KLMN/Q) respectively. Thus,
the complexity of the proposed PCA-based method is
O(K(MN/Q)
3
). We then propose to combine the outputs of
the classifiers in abstract, rank, and score level to increase
the recognition rate.

5. Experimental Results

In our experiments, we used ear images database
consists of 102 grayscale images (6 images for each of 17
subjects) in PGM format. Six views of the left profile from
each subject were taken under uniform, diffuse lighting.
Slight changes in the head position were encouraged from
image to image [16]. In the first experiment, we used the
PCA for the whole ear image using 2 and 4 training
images. In the second, third, and fourth experiments, we
divide the 2D ear images (64x64) into 4, 9 , and 16 blocks
respectively. Then, identify persons through these blocks
and combine the results of the classifiers on abstract level
(A. level), rank level (R. level), and measurements level
(M. level). Simple nearest neighbor classifier applied in
our experiments. We used Cosine, Euclidean, and City
Block distance. The computations of these distances are
shown as follows:

J
LucIìdcun
= ¸(X
s
−X
t
)(X
s
− X
t
)′ (9)

J
Cìt¡ BIock
= ∑ |X
s
− X
t
|
n
]=1
(10)
J
Cosìnc
= 1 −
X
s
X
t

_(X
s
X
s

)(X
t
X
t

)
(11)

where X
s
represents test image (vector), X
t
represents
training images (Vector), and n is the vector's size of test
and training images.
Table 1. shows the results for the four experiments. We
compare between the recognition rate of the classifiers
fusion for the small block and the whole ear image using
PCA. As shown in Table 1, we note that the recognition
rate of the combined small blocks of ear image is better
than the recognition rate of the whole ear image. For
example when divide the ear images into 4 blocks and use
4 training images we note that the recognition rate of small
blocks in abstract level equal to the whole ear image and
in rank level the results of small blocks better than the
whole ear image while the results in measurements level
approximately better than the whole ear image and
achieved superior results. When using 9 and 16 blocks the
results of measurements level achieved the best results and
the results of rank better than the whole ear image but the
results in abstract level lower than the whole ear image.

Figure 1. Block diagram of proposed algorithm

Table 1. Comparison between recognition rate of
classifiers fusion of small block and the whole ear
image using PCA.


2Traini
ng
4Train
ing
Single Classifier
Whole Ear
Image
Kumar et al. [4] 90-95
Alaa et al. [5]
PCA 83.5
PCA+LDA 85.9
PCA+DCT 85.9
Euclidean 89.7059 94.117
6
City Block 88.2353 94.117
6
Cosine 89.7059 94.117
6
4 Blocks
A. Level 79.41 94.117
6
R. Level 88.2353 97.058
8
M. Level
Min 89.7059 100
Max 77.9412 91.176
5
Mean 92.6471 97.058
8
Median 91.1765 94.117
6
Product 88.2353 94.117
6
9 Blocks
A. Level 73.5294 88.235
3
R. Level 91.1765 97.058
8
M. Level
Min 80.8824 91.176
5 Max 82.3529 88.235
3
Mean 89.7059 97.058
8
Median 91.1765 97.058
8
Product 89.7059 94.117
6
16 Blocks
A. Level 64.7059 82.352
9
R. Level 94.1176 97.058
8
M. Level
Min 79.4118 94.117
6
Max 76.4706 88.235
3
Mean 89.7059 94.117
6
Median 91.1765 94.117
6
Product 88.2353 94.117
6
Note that, the recognition rate inversely proportional to
the block size. For example, the results when using 4
blocks better than the results of 9 and 16 blocks. Finally,
we note that mean and median rules achieved results better
than all other methods used in measurement level.
8
Th
Internationl Conference on Informatics and Systems (INFOS 2012)

Table 2, shows the CPU time of calculating PCA
algorithm when using the whole ear image and blocks of
ear image (Proposed algorithm). We note that, the CPU
time when calculating PCA on blocks is very small
compared with the CPU time of the whole ear image. And
the CPU time decreased when the size of blocks
decreased.

Table 2. CPU time of the proposed PCA algorithm



Recognition rates of the state-of-the-art methods for ear
are shown in Table 1. In this table, [4], used PCA to
extract features from ear images and used database as in
our research, [5] also used PCA and the same database,
and combined the features of PCA and Linear
Discriminant Analysis (LDA), also combined PCA with
Discrete Cosine Transform (DCT). Results show that the
proposed fusion method increases the recognition rate in
smaller time in comparison to the state-of-the-art methods.

6. Conclusion

In this paper, we identified persons using ear images
based on the PCA algorithm. In our proposed method, we
solve the high dimension problem of the PCA algorithm
by partitioning the ear image into small blocks and then
combine the classifiers outputs of the minimum distance
classifiers in many levels. Identifying persons using the
proposed algorithm achieved good results than using the
whole ear image. CPU time is very smaller. Our future
work includes combining the outputs of classifying blocks
using other methods and use the feature level fusion to
combine these blocks.

References
[1] A. Iannarelli, “Ear Identification”, Forensic
Identification series, Paramont Publishing Company,
Fremont, California, 1989.
[2] K. Chang, K. W. Bowyer, S. Sarkar, B. Victor.
Comparison and Combination of Ear and Face Images in
Appearance-based Biometrics. IEEE Transaction Pattern
Analysis and Machine Intelligence, 25(9): 1160-1165,
2003.
[3] P. Yan, K. W. Bowyer. Empirical Evaluation of
Advanced Ear Biometrics. Proceedings IEEE Conference
Computer Vision and Pattern Recognition Workshop
Empirical Evaluation Methods in Computer Vision, pp.
41-48, 2005.
[4] Ajay Kumar, David Zhang. Ear Authentication using
Log-Gabor Wavelets. Proceedings of SPIE (2007),
6539(36):- pp.134-139, 2007.
[5] A. Tharwat, A. I. Hashad, G. I. Salama. Human Ear
Recognition Based On Parallel Combination Of Feature
Extraction Methods. The Mediterranean Journal of
Computers and Networks , 6(4):- , 2010.
[6] M. Turk, A. Pentland. Eigenfaces for Recognition.
Journal of Cognitive Neuroscience, 3(1):- 71-86 , 1991.
[7] M. Turk, A. Pentland. Face Recognition Using
Eigenfaces. Proceedings of IEEE Conference on
Computer Vision and Pattern Recognition Maui, Hawaii,
USA, pp. 586-591,1991.
[8] Golub, G.H., van Loan, C.F. Matrix Computations.
third edition John Hopkins University Press, Baltimore,
1996.
[9] Roweis, S. EM algorithms for PCA and SPCA. Neural
Information Processing Systems, 10:- 626–632, 1997.
[10] Schilling, R.J., Harris, S.L. Applied Numerical
Methods for Engineers Using Matlab and C. Brooks/Cole
Publishing Company, 2000.
[11] P. M. Hall, D. Marshall, and R. R. Martin. Merging
And Splitting Eigenspace Models. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 22(9):- 1042–
1049, 2000.
[12] W. Skarbek. Merging Subspace Models For Face
Recognition. In proceedings of the Computer Analysis of
Images and Patterns, 2756:- 606–613, 2003.
[13] L. Liu, Y. Wang, Q. Wang, T. Tan. Fast Principal
Component Analysis Using Eigen sPace Merging. IEEE
International Conference on Image Processing (ICIP), 6:-
457-460, 2007.
[14] Toygar, Ö., Acan, A. Boosting Face Recognition
Speed With A Novel Divide-And-Conquer Approach. In
Proceedings of International Symposium on Computer
and Information Sciences (ISCIS), pp. 430-439, 2004.
[15]A. Jain, K. Nandakumar, A. Ross. Score
normalization in multimodal biometric systems. Pattern
Recognition, 38(12):2270-2285, 2005.
[16] M. Á. Carreira-Perpiñán. Compression Neural
Networks for Feature Extraction: Application to Human
Recognition from Ear Images. (in Spanish), M.S. thesis,
Faculty of Informatics, Technical University of Madrid,
Spain, 1995.


CPU Time (Sec x10
-3
)
2 Training 4 Training
Whole Image 606.5625 615.2188
Parts of
Ear
Images
4 Blocks 36.1252 38.75
9 Blocks 9 9.7029
16 Blocks 2 3.5008