Intelligent Information Management, 2009, 1, 166-171
doi:10.4236/iim.2009.13024 Published Online December 2009 (
Copyright © 2009 SciRes IIM
Multimodal Belief Fusion for Face
and Ear Biometrics
Dakshina Ranjan KISKU1, Phalguni GUPTA2, Hunny MEHROTRA3, Jamuna Kanta SING4
1Department of Computer Science and Engineering, Dr. B. C. Roy Engineering College, Durgapur, India
2Department of Computer Science and Engineering, Indian Institute of Technology Kanpur, Kanpur, India
3Department of Computer Science and Engineering, National Institute of Technology Rourkela, Rourkela, India
4Department of Computer Science and Engineering, Jadavpur University, Kolkata, India
Abstract: This paper proposes a multimodal biometric system through Gaussian Mixture Model (GMM) for
face and ear biometrics with belief fusion of the estimated scores characterized by Gabor responses and the
proposed fusion is accomplished by Dempster-Shafer (DS) decision theory. Face and ear images are con-
volved with Gabor wavelet filters to extracts spatially enhanced Gabor facial features and Gabor ear features.
Further, GMM is applied to the high-dimensional Gabor face and Gabor ear responses separately for quanti-
tive measurements. Expectation Maximization (EM) algorithm is used to estimate density parameters in
GMM. This produces two sets of feature vectors which are then fused using Dempster-Shafer theory. Ex-
periments are conducted on two multimodal databases, namely, IIT Kanpur database and virtual database.
Former contains face and ear images of 400 individuals while later consist of both images of 17 subjects
taken from BANCA face database and TUM ear database. It is found that use of Gabor wavelet filters along
with GMM and DS theory can provide robust and efficient multimodal fusion strategy.
Keywords: multimodal biometrics, gabor wavelet filter, gaussian mixture model, belief theory, face, ear
1. Introduction
Recent advancements of biometrics security artifacts for
identity verification and access control have increased
the possibility of using identification system based on
multiple biometrics identifiers [1–3]. A multimodal bio-
metric system [1,3] integrates multiple source of infor-
mation obtained from different biometric cues. It takes
advantage of the positive constraints and capabilities
from individual biometric matchers by validating its pros
and cons independently. There exist multimodal biomet-
rics system with various levels of fusion, namely, sensor
level, feature level, matching score level, decision level
and rank level. Advantages of multimodal systems over
the monomodal systems have been discussed in [1,3].
In this paper, a fusion approach of face [4] and ear [5]
biometrics using Dempster-Shafer decision theory [7] is
proposed. It is known that face biometric [4] is most
widely used and is one of the challenging biometric traits,
whereas ear biometric [5] is an emerging authentication
technique and shows significant improvements in recog-
nition accuracy. Fusion of face and ear biometrics has
not been studied in details except the work presented in
[6]. Due to incompatible characteristics and physiologi-
cal patterns of face and ear images, it is difficult to fuse
these biometrics based on some direct orientations. In-
stead, some form of transformations is required for fu-
sion. Unlike face, ear does not change in shape over the
time due to change in expressions or age.
The proposed technique uses Gabor wavelet filters [8]
for extracting facial features and ear features from the
spatially enhanced face and ear images respectively.
Each extracted feature point is characterized by spatial
frequency, spatial location and orientation. These char-
acterizations are viable or robust to the variations that
occur due to facial expressions, pose changes and
non-uniform illuminations. Prior to feature extraction,
some preprocessing operations are done on the raw cap-
tured face and ear images. In the next step, Gaussian
Mixture Model [9] is applied to the Gabor face and Ga-
bor ear responses for further characterization to create
measurement vectors of discrete random variables. In the
proposed method, these two vectors of discrete variables
are fused together using Dempster-Shafer statistical de-
cision theory [7] and finally, a decision of acceptance or
rejection is made. Dempster-Shafer decision theory
based fusion works on changed accumulative evidences
which are obtained from face and ear biometrics. The
proposed technique is validated and examined using In-
dian Institute of Technology Kanpur (IITK) multimodal
database of face and ear images and using a virtual or
chimeric database of face and ear images collected from
D. R. KISKU ET AL. 167
BANCA face database and TUM ear database. Experi-
mental results exhibit that the proposed fusion approach
yields better accuracy compared to existing methods.
This paper is organized as follows. Section 2 discusses
the preprocessing steps involved to detect face and ear
images and to perform some image enhancement algo-
rithms for better recognition. The method of extraction of
wavelet coefficients from the detected face and ear im-
ages has been discussed in Section 3. A method to esti-
mate the score density from the Gabor responses which
are obtained from the face and the ear images through
Gabor wavelets has been discussed in the next section.
This estimate has been obtained with the help of Gaus-
sian Mixture Model (GMM) and Expectation Maximiza-
tion (EM) algorithm [9]. Section 5 proposes a method of
combining the face matching score and the ear matching
score which makes use of Dempster-Shafer decision the-
ory. The proposed method has been tested on 400 sub-
jects of IITK database and on 17 subjects of a virtual
database. Experimental results have been analyzed in
Section 6. Conclusions are given in the last section.
2. Face and Ear Image Localization
This section discusses the methods used to detect the
facial and ear regions needed for the study and to en-
hance the detected images. To locate the facial region for
feature extraction and recognition, three landmark posi-
tions (as shown in Figure 1) on both the eyes and mouth
are selected and marked automatically by applying the
technique proposed in [10]. Later, a rectangular region is
formed around the landmark positions for further Gabor
characterization. This rectangular region is then cropped
from the original face image.
For localization of ear region, Triangular Fossa [11]
and Antitragus [11] are detected manually on ear image,
as shown in Figure 1. Ear localization technique pro-
posed in [6] has been used in this paper. Using these
landmark positions, ear region is cropped from ear image.
After geometric normalization, image enhancement op-
erations are performed on face and ear images. Histo-
gram equalization is done for photometric normalization
of face and ear images having uniform intensity distribu-
Figure 1. Landmark positions of face and ear images
3. Gabor Wavelet Coefficients
In the proposed approach the evidences are obtained
from the Gaussian Mixture Model (GMM) estimated
scores which are computed from spatially enhanced Ga-
bor face and Gabor ear responses. Two-dimensional Ga-
bor filter [8] refers a linear filter whose impulse response
function is the multiplication of harmonic function and
Gaussian function. The Gaussian function is modulated
by a sinusoid function. The convolution theorem states
that the Fourier transform of a Gabor filter's impulse
response is the convolution of the Fourier transform of
the harmonic function and the Fourier transform of the
Gaussian function. Gabor function [8] is a non- orthogo-
nal wavelet and it can be specified by the frequency of
the sinusoid and the standard deviations in both x and y
For the computation, 180 dpi gray scale images with
the size of 200 × 220 pixels are used. For Gabor face and
Gabor ear representations, face and ear images are con-
volved with the Gabor wavelets [8] for capturing sub-
stantial amount of variations among face and ear images
in the spatial locations in spatially enhanced form. Gabor
wavelets with five frequencies and eight orientations are
used for generation of 40 spatial frequencies. Convolu-
tion generates 40 spatial frequencies in the neighbour-
hood regions of the current spatial pixel point. For the
face and ear images of size 200 × 220 pixels, 1760000
spatial frequencies are generated. Infact, the huge di-
mension of Gabor responses could cause the perform-
ance degradation and slow down the matching process.
In order to validate the multimodal fusion system, Gaus-
sian Mixture Model (GMM) further characterizes these
higher dimensional feature sets of Gabor responses and
density parameter estimation is performed by Expected
Maximization (EM) algorithm. For illustration, some
face and ear images from IITK multimodal database and
their corresponding Gabor face and Gabor ear responses
are shown in Figure 2(a) and 2(b) respectively.
4. Score Density Estimation
Gaussian Mixture Model (GMM) [9] is used to produce
convex combination of probability distribution and in the
subsequent stage, Expectation Maximization (EM) algo-
rithm [9] is used to estimate the density scores. In this
section, GMM is described for parameter estimation and
score generation.
GMM is a statistical pattern recognition technique.
The feature vectors extracted from Gabor face and Gabor
ear responses can be further characterized and described
by Gaussian distribution. Each quantitive measurements
for face and ear are defined by two parameters: mean and
standard deviation or variability among features. Sup-
pose, the measurement vectors are the discrete random
variable xface for face modality and variable xear for ear
Copyright © 2009 SciRes IIM
(a) Face images and their gabor responses
(b) Ear images and their gabor responses
Figure 2. Face and ear images and their gabor responses
modality. GMM is of the form of a convex combination
of Gaussian distributions [9]):
face xpxp
)()( ),,()(
ear xpxp
)()( ),,()(
where M is the number of Gaussian mixtures and π(m) is
the weight of each of the mixture. In order to estimate
the density parameters of GMM, EM has been used.
Each of the EM iterations consists of two steps – Estima-
tion (E) and Maximization (M). The M-step maximizes a
likelihood function that is refined in each iteration by the
E-step [9].
5. Fusing Scores by Dempster-Shafer Theory
The fusion approach uses Dempster-Shafer (DS) decision
theory [7] to combine the score density estimation ob-
tained by applying GMM to Gabor face and ear re-
sponses for improving the overall verification results. DS
decision theory is considered as a generalization of
Bayesian theory in subjective probability and it is based
on the theory of belief functions and plausible reasoning.
DS decision theory can be used to combine evidences
obtained from different sources of system to compute the
probability of an event. Generally, DS decision theory is
based on two different ideas such as the idea of obtaining
degrees of belief for one question from subjective prob-
abilities for a related query and Dempster’s rule for fus-
ing such degrees of belief while they depend on inde-
pendent items of information or evidence [7].
DS theory combines three function ingredients: the
basic probability assignment function (bpa), the belief
function (bf) and the plausibility function (pf). Let ґFace
and ґEar be two transformed feature sets obtained from
the clustering process for the Gabor face and Gabor ear
responses, respectively. Further, m(ґFace) and m(ґEar) are
the bpa functions for the Belief measures Bel(ґFace) and
Bel(ґEar) for the individual traits respectively. Then the
belief probability assignments (bpa ) m(ґFace) and m(ґEar)
can be combined together to obtain a Belief committed to
a feature set C є Θ according to the following combina-
tion rule [13] or orthogonal sum rule
(), .
Face Ear
Face Ear
Face Ear
Face Ear
mC C
The denominator in Equation (3) is normalizing factor,
which denotes the amounts of conflicts between the be-
lief probability assignments m(ґFace) and m(ґEar). Due to
two different modalities used for feature extraction, there
is an enough possibility to conflict the belief probability
assignments and this conflicting state is being captured
by the two bpa functions. The final decision of user ac-
ceptance and rejection can be established by applying
threshold to m(C).
6. Experimental Results
The proposed multimodal biometrics system is tested on
the two multimodal databases, namely, IIT Kanpur mul-
timodal database of face and ear images and virtual mul-
timodal database consisting of face images taken from
BANCA face database and ear images taken from TUM
ear database.
6.1. Test Performed on IIT Kanpur Database
The results are obtained on multimodal database col-
lected at IIT Kanpur. Database of face and ear consists of
400 individuals’ with 2 face and 2 ear images per person.
The face images are taken in controlled environment
with maximum tilt of head by 20 degree from the origin.
However, for evaluation purpose frontal view faces are
used with uniform lighting, and minor change in facial
Copyright © 2009 SciRes IIM
D. R. KISKU ET AL. 169
expression. These face images are acquired in two dif-
ferent sessions. The ear images are captured with high-
resolution camera in controlled environment with uni-
form illumination and invariant pose. The face and ear
biometrics are statistically different from each other for
an individual. One face and one ear image for each client
are labeled as target and the remaining face and ear im-
ages are labeled as probe.
Table 1 illustrates that the proposed fusion approach of
face and ear biometrics using DS decision theory in-
creases recognition rates over the individual matching.
The results obtained from IITK multimodal database
indicates that the proposed fusion approach with feature
space representation using Gabor wavelet filter and
GMM outperforms the individual face and ear biometrics
recognition while DS decision theory is applied as fusion
rule. This fusion approach achieves 95.53% recognition
rate with 4.47% EER. It has been also seen that FAR is
significantly reduced to 3.4% while it is compared with
the individual matching performances for face and ear
biometrics. Receiver Operating Characteristic (ROC)
curve is plotted in Figure 3 for minute illustrations about
the computed errors and recognition rates. The proposed
fusion approach is also compared with the technique
discussed in [6] and it is found to be a robust fusion
technique for user recognition and authentication while
the combination of Gabor wavelet filter, GMM and DS
decision theory is used.
<--- False Acceptance Rate --->
<--- False Rejection Rate --->
Receiver Operating Characteristics Curves
Ear Recognition
Face Recognition
DS Theory Based Fusion
Figure 3. ROC curves for IIT Kanpur database
Table 1. Error rates of different methods
Methods FRR
Rate (RR) (%)
Face rec-
ognition 8.26 7.82 8.04 91.96
Ear recog-
nition 7.60 5.70 6.65 93.35
DS based
fusion 5.55 3.40 4.47 95.53
Figure 4. Sample ear images from TUM database
6.2. Test Performed on Virtual Database
To verify the usefulness of the proposed technique eva-
luated with more than one multimodal database, a virtual
multimodal database of face and ear images taken from
BANCA database [4] and TUM database [12], respec-
tively, is created. In reality, there does not exist any mul-
timodal database of face and ear images of same subjects,
except the IIT Kanpur database, for experiment. BANCA
face database [4] consists of 20×52 face images obtained
from 52 subjects, each having 20 face images. The face
images are presented with changes in pose, in illumina-
tion and in facial expression. The BANCA face database
considered as one of the complex databases for experi-
ment. On the other hand, ear images of Technical Uni-
versity of Madrid ear database [12] are taken with a
grayscale CCD camera. Each ear image has a resolution
of 384×288 pixels and 256 grayscales. Six ear instances
of the left profile from each subject are taken under uni-
form, diffuse lighting conditions and slight changes in
head position. There are 102 ear images in the database
of TUM which are taken from 17 different individuals.
Six face and six ear images are considered for the
creation of virtual database for the experiment. Face im-
ages of BANCA database are taken randomly. For uni-
form experimental setup, face and ear images are nor-
malized by histogram equalization. Apply uniform reso-
lution and scaling to all the face and ear images. A total
of 6×17 images is collected separately for each face and
ear modality. Some sample ear and face images of TUM
ear database and BANCA database are shown in Figure 4
and Figure 5 respectively.
Three images per person are used for enrollment in the
face and ear verification systems. For each individual, 3
pairs of face-ear are used for training the fusion classifier.
For evaluation purpose, remaining 3 pairs of face-ear are
used for verification and match score generation. The
performance of individual matchers as well as the DS
theory based fusion technique is presented through Equal
Error Rate (EER), False Reject Rate (FRR), False Accept
Rate (FAR) and recognition rate.
Table 2 illustrates the performance of the DS based fu-
sion classifier as well as the individual modality of face
Copyright © 2009 SciRes IIM
Figure 5. Sample face images from BANCA database
Table 2. Error rates determined from virtual multimodal
database are shown
Methods FRR
Rate (RR) (%)
Face rec-
ognition 6.18 5.44 5.81 94.19
Ear recog-
nition 5.88 4.04 4.96 95.04
DS based
fusion 4.7 2.98 3.84 96.16
and ear biometrics in terms of EER, FRR, FAR, recogni-
tion rates. The results obtained from the chimeric or vir-
tual database indicate that individual face and ear bio-
metrics perform well while Gabor wavelets and GMM
are used for feature characterization and score density
estimation respectively. Further, when DS decision the-
ory is used to fuse the individual scores obtained from
face and ear biometrics, it performs better than the indi-
vidual matchers. However, due to the less number of
subjects in virtual database in comparison to that in the
IIT Kanpur multimodal database, results obtained from
virtual database show better performance over the IIT
Kanpur database.
DS decision theory based fusion approach achieves the
recognition rate of 96.16% from the virtual database with
3.84% EER. It has also been seen that the FAR is sig-
nificantly reduced to 2.98% while it is compared with
that of the FAR obtained from IIT Kanpur database and
that of the FAR determined from individual matchers.
Receiver Operating Characteristics (ROC) curves are
shown in Figure 6 for the individual face and ear biomet-
rics along with the DS based fusion scheme.
Multimodal fusion of face and ear biometrics are
rarely available except the work presented in [6], which
has used appearance based techniques to fuse these two
biometric traits. Test performed on IIT Kanpur database
consists of 400 subjects with the proposed fusion ap-
proach exhibits robust performance while it is compared
with the performance based on virtual database having
very less number of subjects. When IIT Kanpur database
is used for evaluation, a pair of face-ear is used for train-
ing the fusion classifier and another pair of face-ear is
used for verification. Therefore, two images per person
are used for face and ear modalities in IIT Kanpur data-
base against 6 images per person are taken in virtual da-
tabase. The recognition rate obtained from virtual mul-
timodal database shows the robustness and efficacy for
the proposed fusion method while small numbers of sub-
jects are used with more instances for a single subject.
Test performed on virtual database also exhibits the in-
variant performance with various facial expressions, pose
changes, illumination changes in BANCA face database.
This could not be analyzed for IIT Kanpur database be-
cause face images in that database are almost uniform
and not much variation in facial expressions and pose
In [6], the authors have proposed a multimodal fusion
of face and ear biometrics using principal component
analysis and this principal component analysis is used for
extracting eigen-face and eigen-ear. Further, these ei-
gen-face and eigen-ear are fused for recognition. This
fusion classifier has been achieved 90.09% recognition
rate with 197 subjects of face and ear images. In contrast,
the proposed fusion classifier has achieved 95.53% and
96.16% recognition rates for IIT Kanpur and virtual da-
tabases respectively. The use of Gabor wavelets for fea-
ture characterization and GMM for score density estima-
tion with Dempster-Shafer decision theory based fusion
technique has provided higher recognition rates with the
substantial variations in the databases.
Figure 6. ROC curves for different matchers
Copyright © 2009 SciRes IIM
Copyright © 2009 SciRes IIM
7. Conclusions
The proposed fusion strategy combines information that
has been extracted through Gabor wavelet filters and
Gaussian Mixture Model estimator. Gabor wavelet filters
have been used for extraction of spatially enhanced face
and ear features which are viable and robust to different
variations. Using E-estimator and M-estimator in GMM,
reduced feature sets have been extracted from high di-
mensional Gabor face and Gabor ear responses through
parameter estimation. These reduced feature sets are
fused together by Dempster-Shafer decision theory. It has
been found that the technique exhibits increase in accu-
racy and significant improvement over the existing
[1] A. K. Jain and A. K. Ross, “Multibiometric systems,”
Communications of the ACM, Vol. 47, No.1, pp. 34–40,
[2] A. K. Jain, A. K. Ross, and S. Prabhakar, “An introduction
to biometrics recognition,” IEEE Transactions on Circuits
and Systems for Video Technology, Vol. 14, No. 1, pp.
4–20, 2004.
[3] A. Rattani, D. R. Kisku, M. Bicego, and M. Tistarelli,
“Robust feature-level multibiometric classification,”
Proceedings of the Biometric Consortium Conference – A
special issue in Biometrics, pp. 1–6, 2006.
[4] D. R. Kisku, A. Rattani, E. Grosso, and M. Tistarelli,
“Face identification by SIFT-based complete graph to-
pology,” Proceedings of the IEEE Workshop on Auto-
matic Identification Advanced Technologies, pp. 63–68,
[5] B. Arbab-Zavar, M. S. Nixon, and D. J. Hurley, “On
model-based analysis of ear biometrics,” First IEEE In-
ternational Conference on Biometrics: Theory, Applica-
tions, and Systems, pp. 1–5, 2007.
[6] K. Chang, K. W. Bowyer, S. Sarkar, and B. Victor “Com-
parison and combination of ear and face images in ap-
pearance-based biometrics,” Transaction on Pattern
Analysis and Machine Intelligence, Vol. 25, No. 9, pp.
1160–1165, 2003.
[7] N. Wilson, Algorithms for Dempster-Shafer theory, Ox-
ford Brookes University.
[8] T. S. Lee, “Image representation using 2D Gabor wave-
lets,” IEEE Transaction on Pattern Analysis and Machine
Intelligence, Vol. 18, pp. 959–971, 1996.
[9] L. Xu and M. I. Jordan, “On convergence properties of the
EM algorithm for Gaussian Mixtures,” Neural Computa-
tion, Vol. 8, No. 1, pp. 129–151, 1996.
[10] F. Smeraldi, N. Capdevielle, and J. Bigün, “Facial features
detection by saccadic exploration of the gabor decompo-
sition and support vector machines,” In the 11th Scandi-
navian Conference on Image Analysis, pp. 39–44, 1999.
[11] A. Iannarelli, Ear Identification, Forensic Identification
series, Fremont, Paramont Publishing Company, Califor-
nia, 1989.
[12] M. A. Carreira-Perpiñán, “Compression neural networks
for feature extraction: Application to human recognition
from ear images,” MSc thesis, Faculty of Informatics,
Technical University of Madrid, Spain, 1995.
[13] D. R. Kisku, M. Tistarelli, J. K. Sing, and P. Gupta, “Face
recognition by fusion of local and global matching scores
using DS theory: An evaluation with uni-classifier and
multi-classifier paradigm,” In the Proceedings of IEEE
Computer Soceity Conference on Computer Vision and
Pattern Recognition Workshop, pp. 60–65, 2009.