Content uploaded by Salim Chitroub
Author content
All content in this area was uploaded by Salim Chitroub on May 19, 2015
Content may be subject to copyright.
Multimodal Biometric Person Recognition System based on
Iris and Palmprint Using Correlation Filter Classifier
Abdallah Meraoumia1,Salim Chitroub1 and Ahmed Bouridane2;3
1Signal and Image Processing Laboratory, Electronics and Computer Science Faculty, USTHB.
P.O. box 32, El Alia, Bab Ezzouar, 16111, Algiers, Algeria
2School of Computing, Engineering and Information Sciences, Northumbria University,
Pandon Building, Newcastle upon Tyne, UK.
3Department of Computer Science, King Saudi University,
P.O. Box 2454, Riyadh, 11451, Saudi Arabia
Email: Ameraoumia@gmail.com, S_chitroub@hotmail.com, Ahmed.Bouridane@northumbria.ac.uk,
ABouridane.c@ksu.edu.sa
Abstract—Biometric system has been actively emerging in
various industries for the past few years, and it is continuing to
roll to provide higher security features for access control system.
Many types of uni-modal biometric systems have been
developed. However, these systems are only capable to provide
low to middle range of security feature. Thus, for higher
security feature, the combination of two or more uni-modal
biometrics is required. In this paper, we propose a multi-modal
biometric system for person identification using Palmprint and
Iris modalities. This work describes the development of a multi-
biometric personal identification system based on Minimum
Average Correlation Energy Filter (MACE) method (for
matching). The outputs of the sub-systems (Iris and Palmprint)
are combined using the concept of data fusion at matching score
level. We have also tried the various image fusion rulers to
choose the best one for Iris and Palmprint classification. The
experimental results showed that the designed system achieves
an excellent identification rate and provides more security than
uni-modal biometric system.
Index Terms—Biometrics, Identification, Iris, Palmprint,
MACE, PSR, Data fusion.
I. INTRODUCTION
Automatic personal identification from their physical and
behavioral traits, called biometrics technologies, is now
needed in many fields such as: surveillance systems, security
systems, physical buildings and many more applications.
These technologies makes use of the characteristics of person
such as fingerprint, iris, face, palmprint, gait, and voice, for
personal identification, which provides advantages over non-
biometric methods such as PIN, and ID cards [1].
In recent years, biometric identification has seen
considerable improvements in reliability and accuracy, with
some biometrics offering reasonably good overall
performance. However, Biometric systems based on uni-
modal biometrics are often not able to meet the desired
performance requirements for large user population
applications [2], due to problems such as noisy data, non-
university, spoof attacks, and unacceptable error rates.
Therefore, multi-modal biometric method is a novel solution
to overcome those problems. Recently, multi-modal
biometric techniques have attracted increasing attention and
interest among researchers, in the hope that the
supplementary information between different biometrics
might improve the identification performance [3].
The design of a multi-modal biometric system is strongly
dependent on the application scenario. A number of
multimodal biometric systems have been proposed in
literature that differ from one another in terms of their
architecture [4], the number and choice of biometric
modalities [5], the level at which the evidence is
accumulated, and the methods used for the integration or
fusion of information [6]. In this paper, palmprint and iris are
integrated in order to construct an efficient multi-modal
biometric identification. In this system, we propose to use
(Unconstrained) Minimum Average Correlation Energy Filter
(U)MACE method (for matching). The iris and palmprint
images are used as inputs of the matcher modules. The
outputs of the matcher modules {Max peak size or peak-to-
sidelobe ratio (PSR)} are combined using the concept of data
fusion at matching score level. We have also tried the various
image fusion rulers {Principal Components Analysis (PCA)
based and Discrete Wavelet Transform (DWT) based} to
choose the best one for iris and palmprint identification.
The remainder of the paper is organized as follows. The
proposed scheme for uni-modal biometric identification
system based on MACE filter is exposed in section 2. Section
3 gives a brief description of the region of interest extraction
for iris and palmprint. Section 4 present the matching
technique used. This section includes also an overview of
(U)MACE filter. The similarity measurement used is detailed
in section 5. The experimental results, prior to fusion and
© ICCIT 2012 782
Fig. 1. The block-diagram of the proposed uni-modal biometric identification system based on minimum average correlation energy.
after fusion, are given and commented in section 6. Finally,
section 7 is devoted to the conclusion and future work.
II. UNI-MODAL SYSTEM OVERVIEW
The proposed system is composed of two different
subsystems exchanging information in image level or
matching score level. Each sub-system exploits different
modalities (iris or palmprint). Each uni-modal biometric
system (for example Fig. 1 show a uni-modal biometric
identification system based on iris modality) consists of
preprocessing, matching (correlation process), normalization
and decision process. Modality (Iris and Palmprint)
identification with correlation filters is performed by
correlating a test image {transformed into the frequency
domain via a discrete fast Fourier transform} with the
designed filter (enrollment) also in the frequency domain.
The output correlation is subjected to an inverse fast Fourier
transform and reordered into the dimensions of the original
training image, prior to being phase shifted to the center of
the frequency square. The resulting correlation plane is then
quantified using performance measures (peak-to-sidelobe
ratio or max peak size ratio). Based on this unique measure, a
final decision is made.
III. PREPROCESSING PROCESS
A. Palmprint preprocessing
In order to localize the palm area, the first step is to
preprocess the palm images; we use the preprocessing
technique described in [7] to align the palmprints. In this
technique, Gaussian smoothing filter is used to smoothen the
image before extracting the Region Of Interest (ROI) and its
features. After that, Otsu’s thresholding is used for binarized
the hand. A contour-following algorithm is used to extract the
hand contour. The tangent of the two stable points on the
hand contour (they are between the forefinger and the middle
finger and between the ring finger and the little finger) are
computed and used to align the palmprint. The central part of
the image, which is 128x128, is then cropped to represent the
whole palmprint.
B. Iris preprocessing
The iris is the annular region of the eye bounded by the
pupil and the sclera (white of the eye) on either side. Like
most other biometric identification systems, the input eye
contained images need to be processed (eliminate
uninterested information) so that the characteristic iris
features can be extracted for comparison. The main pre-
processing steps consist [8] of localization of the inner and
outer iris boundaries (Boundaries tracking), localization of
eyelid boundaries (Noise detection). Finally, segmentation
and transformation from polar coordinates to a fixed size
rectangular image (Normalization).
IV. MATCHING PROCESS
For each class a single (U)MACE filter is synthesized.
Once the (U)MACE filter , has been determined, the
input test image f is cross correlated with it in the manner:
, ,
, (1)
Where the test image is first transformed to frequency
domain and then reshaped to be in the form of a vector. The
result of the previous process is convolved with the conjugate
of the (U)MACE filter. This operation is equivalent with
cross correlation with the (U)MACE filter. The output is
transformed again in the spatial domain. Essentially MACE
filter is the solution of a constrained optimization problem
that seeks to minimize the average correlation energy while at
the same time satisfy the correlation peak constraints. As a
result the output of the correlation planes will be close to zero
everywhere except at the locations of the trained objects that
are set to be correct where a peak will be produced.
MACE filter,H, is found using Lagrange multipliers in the
frequency domain and is given by [9]:
(2)
D is a diagonal matrix of sizex, (d is the number of pixels
in the image) containing the average correlation energies of
the training images across its diagonals. X is a matrix of size
Nxd where N is the number of training images and * is the
complex conjugate.
783
The columns of the matrix X represent the Discrete Fourier
coefficients for a particular training image . The column
vector u of size N contains the correlation peak constraint
values for a series of training images. These values are
normally set to 1.0 for images of the same class.
The UMACE filter like the MACE filter minimizes the
average correlation energy over a set of training images, but
does so without constraint (u), thereby maximizing the peak
height at the origin of the correlation plane. The UMACE
filter expression, H, is given by [10]:
(3)
V. SIMILARITY MEASUREMENT
Typically, the height of this peak can be used as a good
similarity measure for image matching (Fig. 2.(a)). Another
parameter, PSR, can be used for measuring the similarity
between two samples. PSR is a metric that measures the peak
sharpness of the correlation plane. For the estimation of the
PSR the peak is located first. Then the mean and standard
deviation of the 40x40 sidelobe region (excluding a 5x5
central mask) centered at the peak are computed. PSR is then
calculated as follows [11]:
(4)
Peak is the maximum located peak value in the correlation
plane, mean is the average of the sidelobe region surrounding
the peak and σ is the standard deviation of the sidelobe region
(Fig. 2.(b)).
VI. EXPERIMENTAL RESULTS AND DISCUSSION
A. Experimental database
To evaluate the performance of the proposed multi-
biometric identification scheme, a database containing
palmprint and iris images was required. In this work, we
construct a multibiometric database for our experiments
based on CASIA Iris database [12], and the Hong Kong
Polytechnic University (PolyU) Palmprint database [13]. The
multi-biometric database consists of six iris images and six
palmprint images per person with total of 100 persons. Two
samples, of each palm and iris, are randomly selected to
construct a training set and the rest of the samples are taken
as the test set.
B. Open set identification
1) Uni-modal test result: The goal of this experiment was to
evaluate the system performance when using information
from each modalities. For this, there are a total of 200
training images and 400 test images for each modality,
respectively. Therefore, there are totally 400 genuine
comparisons and 39600 impostor comparisons are generated
MACE and UMACE filters are applied to evaluate the
identification performance. The PSR and the peak matching
are determined and the identification performance is
illustrated in Fig. 3.
a) Iris test results: Fig. 3.(a) compares the performance of the
open set iris identification system under the two filters
(MACE and UMACE) and the two performance measures
(Peak size and PSR). The experimental results indicate that
the MACE filter and PSR matching perform better result than
the other cases in terms of Equal Error Rate (EER). For
example, if the MACE filter with peak matching is used, we
have EER = 2.500 % at the threshold To = 0.5079. In the case
of using the UMACE filter with peak matching, EER was
3.000 % at To = 0.5560. The UMACE filter with PSR
matching done an EER equal to 4.750 % at To = 0.5433. The
use of MACE filter with PSR matching improves the result
(2.287 % at To = 0.4624) for a database of 100 persons. The
system was tested with different thresholds and the results are
shown in Table. 1.
b) Palmprint test results: The Receiver Operating
Characteristics (ROC) shown in Fig. 3.(b) depicts the
performance of the open set palmprint identification system
at all filters and performance measures. Our identification
system can achieve a best EER of 0.250 % for To = 0.5590
and the maximum Genuine Acceptance Rate (GAR) =
99.7500 % in the case of MACE filter and PSR matching. For
example, the described identification system can recognize
palmprints quite accurately with an EER of 0.500 % and To =
0.6185 with MACE filter and Peak matching. When the To, is
0.6442, the EER is equal to 0.936 % with UMACE filter and
Peak matching. Finally, the UMACE filter and PSR matching
done an EER equal to 1.500 % at To = 0.6016. The
performance of the system identification under different
values of To, which control the FAR and the FRR, is shown
in Table 1.
In Fig. 3.(c), we compare the performance of iris and
palmprint modalities. The results show the benefits of using
the palmprint modality. Thus, the performance of the open set
uni-modal identification system is significantly improved by
using the palmprint data.
2) Multi-modal test result: The goal of this experiment was to
investigate the systems performance when we fuse
information from iris and palmprint modalities. Therefore,
information presented by different biometrics is fused to
Fig. 2. Similarity matching. (a) Max peak size and (b) PSR
784
make the system efficient using two levels, matching score
level and image level.
a) Fusion at the matching score level: The information
presented by two modalities (iris and palmprint) is fused at
the matching score level to make the system efficient. For
that, a series of experiments were carried out to selection the
best fusion rule that minimize the EER using the best uni-
modal result (MACE filter with PSR matching) for the two
modalities. Thus, to determine the best fusion rule, a
graphical relationship (ROC) can be established (see Fig.
4.(a)). We can observe that the SUM rule based fusion has
the best performance. Thus, the best result of EER is given as
0.010 % with To of 0.8911. The performance of the open set
identification system is significantly improved by using the
fusion. Finally, the performance of the open set identification
system under different fusion rule and a database size equal
to 100 is shown in Table 2.
b) Fusion at image level: Image fusion is the process by
which two or more images are combined into a single image.
Recently, several fusion techniques have been proposed by
various researchers [14]. The PCA and DWT-based approach
is widely used in image fusion.
• PCA-based image fusion: In order to see the performance
of the system, we usually plot, in Fig. 4.(b), the ROC curve
for all the case possible (MACE filter, UMACE filter, Peak
matching, PSR matching). From Fig. 4.(b), we can observe
the benefits of using the MACE filter with PSR matching.
For example, MACE filter with Peak matching give an EER
equal to 0.250 % at To = 0.5350. This system can achieve an
EER of 0.936 % for To = 0.6442 in the case of UMACE filter
with Peak matching. UMACE filter with PSR matching done
an EER equal to 1.500 % at To = 0.6016. Finally, in the case
of using the MACE filter with PSR matching, the system can
operate at a 0.046 % EER, and the corresponding threshold is
To = 0.6635. The experimental result shows that fusion of the
two modalities is much higher than the individual modality.
Finally, FAR, FRR and GAR of all four cases, is listed in
Table 3.
• DWT-based image fusion: To enhance the performance of
identification, the two modalities (iris and palmprint) are
fused in single image by using DWT techniques. Fig. 4.(c)
shows the result of DWT-based image fusion on the different
filters and matching measure. It is observed from this figure
that our fusion strategy provide a better performance than the
single modality especially in the case of MACE filter with
PSR matching. The system performance can be achieved a
minimum EER of 0.053 % and a threshold To = 0.6666,
followed by MACE filter with Peak matching, UMACE filter
with Peak matching and MACE filter with PSR matching at
EER equal to 0.250 %, 0.984 % and 1.500 %, respectively,
with a threshold To = 0.5380, 0.6418 and 0.6069,
respectively. Finally, the FAR, FRR and GAR are computed
and listed in Table 3.
To find the better methods (uni-modal, multi-modal
system), with the lowest EER, graphs showing the ROC
curve were generated (see Fig. 5.(a)). By the analysis of this
Fig. 3. Uni-modal open set system identification test results. (a) The ROC curves for Iris modality, (b) The ROC curves for Palmprint modality and (c)
Performance comparison.
785
plot, it can be observed that the performance of fusion at the
matching score level is most excellent than that of image
level and uni-modal system. A fusion at the matching score
level using SUM rule improves the result (EER = 0.010 %)
for a database size equal to 100. However, it can be
concluded that the fusion at the matching score level yields
much better identification results compared with image level.
Finally, the ROC curve of GAR against FAR for various
thresholds is shown in Fig. 5.(b) and the genuine and
impostor distributions are plotted in Fig. 5.(c).
C. Closed set identification
For the evaluation of the closed set identification system
performance, table 4, table 5 and table 6 shows the
performance of the uni-modal system, multi-modal system
based on matching score level and multi-modal system based
on image level. From these tables, it can be seen that the
performance of multi-modal system based on matching score
level (SUM rule) is much better than the others cases in terms
of Rank-One Recognition (ROR) rate. For example, if only
the iris is used, we have ROR = 91.250 % with a lowest Rank
of Perfect Recognition (RPR) of 74, the ROR is increase of
98.750 % with RPR equal to 45 if palmprint is used.
The system can be work with a ROR equal to 99.750 % in
the case of multi-modal system based on matching score level
and multi-modal system based on image level (PCA and
DWT) with a RPR equal to 2, 11 and 13, respectively. From
these results, we can concluded that the fusion at matching
score level give the best performance.
VII. CONCLUSION AND FUTURE WORK
In this paper, a multi-modal biometric identification system
based on fusion of two biometric traits, palmprint and iris,
has been proposed. Fusion of these two biometric traits is
carried out at the matching score level and image level. The
proposed system use minimum average correlation energy
filter for matching process. To compare the proposed multi-
modal system with the uni-modal systems, a series of
experiments has been performed in the two cases, open set
identification and closed set identification, and it has been
found that the proposed multi-modal system gives a
considerable performance gain over the uni-modal systems in
the two cases.Our future work will focus on the performance
evaluation in both phases (verification and identification) by
Fig. 4. Multi-modal open set system identification test results. (a) The ROC curves for the case of fusion at matching score level, (b) The ROC curves
for the case of fusion at image level (PCA based) and (c) The ROC curves for the case of fusion at image level (DWT based).
786
using a large size database and integration of other biometric
traits such as fingerprint or face to get the system
performances with a high accuracy.
REFERENCES
[1] Arun A. Ross, K. Nandakumar and A. K. Jain, “Handbook of
Multibiometrics”, Springer Science+Business Media, New York, 2006.
[2] A. Kumar and D. Zhang, “Improving Biometric Authentication
Performance from the User Quality”, IEEE transactions on
Instrumentation and Measurement, Vol. 59, No. 3, 2010, pp. 730-735.
[3] A.K. Jain, A. Ross, S. Prabhakar, “An introduction to biometric
recognition”, Circuits and Systems for Video Technology, Vol.14, 2004,
[4] Soumitra Kar, Swati Hiremath, Dilip G. Joshi, Vinod.K.Chadda and
Apurva Bajpai, “A Multi-Algorithmic Face Recognition System”,
International Conference on Advanced Computing and
Communications, ADCOM 2006, Surathkal, pp: 321-326, 2007.
[5] Abdallah Meraoumia, Salim Chitroub and Ahmed Bouridane,
“Palmprint and Finger-Knuckle-Print for efficient person recognition
based on Log-Gabor filter response”, Analog Integrated Circuits and
Signal Processing, 2011, Volume 69, Number 1, Pages 17-27.
[6] Dakshina Ranjan Kisku, Ajita Rattani, Phalguni Gupta, Jamuna Kanta
Sing, “Biometrics Sensor Image Fusion for Identity Verification”, IEEE
International Conference on Technologies for Homeland Security (HST
2009), MA-Greater Boston, USA, IEEE Press, pp. 433-439, 2009.
[7] David Zhang, Wai-Kin Kong, Jane You, and Michael Wong, “Online
Palmprint Identification”, IEEE transactions on pattern analysis and
machine intelligence, vol. 25, no. 9, september 2003, pp: 1041-1050.
[8] Kaushik Roy and Prabir Bhattacharya, “Optimal Features Subset
Selection and Classification for Iris Recognition”, Journal on Image and
Video Processing, Image Analysis and Recognition, ICIAR 2008, Pvoa
de Varzim, Portugal, 2008, LNCS, pp: 894-904.
[9] Aini Hussain, Rosniwati Ghafar, Salina Abdul Samad and Nooritawati
Md Tahir, “Anomaly Detection in Electroencephalogram Signals Using
Unconstrained Minimum Average Correlation Energy Filter”, Journal of
Computer Science 5 (7): 501-506, 2009.
[10] Rosniwati Ghafar, Aini Hussain, Salina Abdul Samad and Nooritawati
Md Tahir, “Umace Filter for Detection of Abnormal Changes In Eeg:
A Report of 6 Cases”, World Applied Sciences Journal 5 (3): 2008.
[11] Nooritawati Md Tahir, Aini Hussain, Salina Abdul Samad and
Hafizah Husain, “Posture Recognition Using Correlation Filter
Classifier”, Journal of Theoretical and Applied Information
Technology, 4(9), 2008, 767-773.
[12] CASIA Iris Image Database,: http://www.sinobiometrics.com.
[13] The Hong Kong Polytechnic University, PolyU Palmprint Database,
[Online]. Available: http:// www. Comp. polyu. edu.hk/ biometrics.
[14] Xu Xuebin, Zhang Deyuri, “NSCTWavelet: An Efficient method for
Multimodal Biometric Recognition Based on Pixel level Fusion”,
International Symposium on Industrial Electronics, Seoul, Korea,
July, 2009.
Fig. 5. Performance of the best system. (a) Comparison of uni-modal and multimodal system, (b) The ROC curves for the best system and (c) The
genuine distribution and the imposter distribution.
787
