To appear in Biometrics: Theory, Applications and Systems (BTAS 09), Washington DC, September 2009
Abstract— Periocular biometric refers to the facial region in the
immediate vicinity of the eye. Acquisition of the periocular
biometric does not require high user cooperation and close
capture distance unlike other ocular biometrics (e.g., iris, retina,
and sclera). We study the feasibility of using periocular images
of an individual as a biometric trait. Global and local
information are extracted from the periocular region using
texture and point operators resulting in a feature set that can be
used for matching. The effect of fusing these feature sets is also
studied. The experimental results show a 77% rank-1
recognition accuracy using 958 images captured from 30
different subjects.
I. INTRODUCTION
CULAR biometrics has made rapid strides over the past
few years primarily due to the significant progress made
in iris recognition. The iris is the annular colored structure in
the eye surrounding the pupil and its function is to regulate the
size of the pupil thereby controlling the amount of light
incident on the retina. The surface of the iris exhibits a very
rich texture due to the numerous structures evident on its
anterior portion. The random morphogenesis of the textural
relief of the iris and its apparent stability over the lifetime of
an individual, have made it a very popular biometric. Both
technological and operational tests conducted under
predominantly constrained conditions have suggested the
uniqueness of the iris texture across individuals and its
potential as a biometric in large-scale systems enrolling
millions of individuals [1, 2]. Indeed, even the two irises of an
individual are observed to be different in their intricate
textural content.
Besides the iris, other ocular traits have been investigated
for human recognition, viz., the retinal and the conjunctival
vasculature.
1. Retinal vasculature: The blood vessel pattern on the retina is
believed to be unique across individuals [3]. Typically, a
coherent light source is used to illuminate the vasculature
pattern on the back of the eye and a CCD is used to image
this pattern. However, a cooperative subject is assumed for
procuring a good quality image that can be used during the
matching phase.
2. Conjunctival vasculature: The vasculature pattern observed
Manuscript received June 7, 2009 and revised Aug 15, 2009.
Unsang Park and Anil K. Jain are with the Department of Computer
Science and Electrical Engineering at Michigan State University (email:
parkunsa@cse.msu.edu
, jain@cse.msu.edu).
Arun Ross is with the Lane Department of Computer Science and
Electrical Engineering at West Virginia University (email:
arun.ross@mail.wvu.edu
).
on the sclera of the eye has also been suggested as a
potential biometric [4]. These blood vessels typically reside
in the conjunctiva and the episclera layers of the sclera
(although the term “conjunctival vasculature” is used to
denote both sets of vessels), and are revealed when the iris is
“off-axis” with respect to the imaging device. Thus, there is
significant potential in utilizing these vasculature patterns
along with the iris texture in a bimodal biometric system by
employing a multispectral sensor for image acquisition.
In spite of the tremendous progress made in ocular biometrics
(especially iris), there are significant challenges encountered
by these systems:
1. The iris is a moving object with a small surface area that is
located within the independently movable eye-ball. The
eye-ball itself is located within another moving object – the
head. Therefore, reliably localizing the iris in eye images
obtained at a distance from unconstrained human subjects
can be difficult [5]. Furthermore, since the iris is typically
imaged in the near infrared portion (700 – 900nm) of the
electromagnetic (EM) spectrum, appropriate invisible
lighting is required to illuminate it prior to image
acquisition.
2. Retinal vasculature cannot be easily imaged unless the
subject is cooperative. In addition, the imaging device has
to be in close proximity to the eye.
3. While conjunctival vasculature can be imaged at a distance,
the curvature of the sclera, the specular reflections in the
image and the fineness of the vascular patterns, can
confound the feature extraction and matching modules of
the biometric system [6].
Periocular Biometrics in the Visible Spectrum: A Feasibility Study
Unsang Park, Arun Ross, and Anil K. Jain
O
Fig. 1: Example periocular images from two different subjects: (a)(b)
without eyebrows and (c)(d) with eyebrows.
(a)
(b)
(c)
(d)
To appear in Biometrics: Theory, Applications and Systems (BTAS 09), Washington DC, September 2009
In this work, we attempt to mitigate some of these concerns
by considering a small region around the eye as an additional
biometric. We refer to this region as the periocular region. In
this work we explore the potential of the periocular region as a
biometric in color images. We do not use the near-IR
spectrum in this paper, although the eventual goal is to use a
multispectral acquisition device that can image the periocular
region in both the visible and near-IR spectral bands [7]. This
would ensure the possibility of combining the iris texture with
the periocular texture. The use of the periocular region has
several benefits:
1. In images where the iris cannot be reliably obtained (or
used), the surrounding skin region may be used to either
confirm or refute an identity.
2. The use of the periocular region represents a good
trade-off between using the entire face region or using only
the iris for recognition. When the entire face is imaged
from a distance, the iris information is typically of low
resolution; this means the matching performance due to the
iris modality will be poor. On the other hand, when the iris
is imaged at close quarters, the entire face may not be
available thereby forcing the recognition system to rely
only on the iris.
3. The periocular region can offer information about
eye-shape that may be useful as a soft biometric.
4. The depth-of-field of iris systems can be increased if the
surrounding ocular region were to be included as well.
The purpose of this work is to do a feasibility study on using
periocular information as a biometric. Thus, images obtained
in the visible spectrum are studied for this purpose.
II. P
ERIOCULAR RECOGNITION
The proposed periocular recognition process consists of a
sequence of operations: image alignment (for the global
matcher described below), feature extraction, and matching.
We adopt two different approaches to the problem: one based
on global information and the other based on local
information. The two approaches use different methods for
feature extraction and matching. We will first review the
characteristics of these two approaches, and describe each
intermediate process.
A. Global vs. Local Matcher
Most image matching schemes can be categorized as global
or local. The basic difference between global and local
methods is based on whether the features are extracted from
the entire image (or a region of interest) or from a set of local
regions. Representative global features are color, shape, and
texture [8]. Global features are represented as a fixed length
vector and the matching process simply compares these fixed
length vectors, which is very time efficient.
On the other hand, the local feature based approach first
detects a set of key points and encodes each of the key points
using the surrounding pixel values (resulting in a local key
descriptor) [9, 10]. Then, the number of matching key points
between two images is calculated as the match score. Since
the number of key points varies depending on the input image,
two sets of key points from two different images cannot be
Fig. 2:
Example images showing difficulties in periocular image
alignment.
(a) Example images showing eyelid movement
(b) Example images where multiple corner candidates are present
Fig. 3: Global descriptor construction process.
(a) Input image
(b) Iris detection
(d) Interest region sampling
(c) Interest point sampling
Fig. 4: Examples of local features and bounding boxes for descriptor
construction in SIFT. Each bounding box is rotated with respect to
the major orientation.
To appear in Biometrics: Theory, Applications and Systems (BTAS 09), Washington DC, September 2009
directly compared. Therefore the matching scheme has to
compare each key point from one image against all the key
points in the other image, thereby increasing the time for
matching. There have been efforts to achieve a constant time
matching using local features through the bag of words
representation [11].
In terms of the matching accuracy, local feature-based
approaches have shown better performance. When all
available pixel values are encoded into the feature vector (as is
the case when global features are used), it becomes more
susceptible to image variations especially with respect to
geometric transformations and spatial occlusions. The local
feature based approach, on the other hand, is more robust to
such variations because only a subset of distinctive regions is
used to represent an image. This has resulted in more active
research on local feature based image retrieval schemes [12,
13, 14].
Face, iris, and hand mostly adopt a global representation
scheme while fingerprint mostly adopts a local representation
scheme. The basic criterion for determining different
representations in image-based biometrics is whether the trait
under consideration has a common morphology across all
subjects. If we take the average of a hundred face, iris, or hand
images after proper scaling and alignment, the output will still
appear as a legitimate face, iris, or hand image. However, the
average of a hundred fingerprint images will not look like a
fingerprint image anymore. Therefore, the face, iris, or hand
images can be aligned in a certain common coordinate space
and encoded into a fixed length feature vector. However,
fingerprint and other general images need to be represented by
their local key points.
We use both global and local matching methods for
periocular recognition in order to take advantage of the fixed
length feature representation of the global scheme and the
distinctiveness of the local scheme.
B. Image Alignment
Periocular images contain common components (i.e., iris,
sclera, and eyelids) that can be represented in a common
coordinate system. Once a common area of interest is
localized, a global representation scheme can be used. The iris
or eyelids are good candidates for the alignment process. Even
though both the iris and eyelids exhibit motion, such
variations are not significant in the periocular images used in
this research, since the images were taken under similar
operational conditions as traditional iris recognition systems,
where variations due to the iris and eyelids are deliberately
constrained. While frontal iris detection can be performed
fairly well due to the approximately circular geometry of the
iris and the clear contrast between iris and sclera, the accurate
detection of eyelids is more difficult. The inner and outer
corners of the eye can also be considered as anchor points, but
there can be multiple candidates as shown in Fig. 2.
Therefore, we primarily use the iris for image alignment. A
public domain iris detector based on Hough transformation
was used for localizing the iris [15]. The iris can be used for
translation and scale normalization of the image, but not for
rotation normalization. However, we overcome the small
rotation variations using a rotation tolerant feature
representation.
The iris-based image alignment is only required by the
global matching scheme. The local matcher does not require
image alignment because the descriptors corresponding to the
key points can be independently compared of each other.
C. Feature Extraction
We extract global features using all the pixel values in the
detected region of interest that is defined with respect to the
iris. The local features, on the other hand, are extracted from a
set of characteristic regions.
From the center, C
iris
, and radius, R
iris
, of the iris, multiple
(=n
pi
) interest points p
1
, p
2
, …, p
npi
are selected within a
rectangular window defined around C
iris
with a width of 6×R
iris
and a height of 4×R
iris
as shown in Fig. 3. The number of
interest points is decided based on the sampling frequency
(1/D
p
) which is inversely proportional to the distance between
interest points, D
p
×R
iris
.
For each interest point p
i
, a rectangular region r
i
is defined
with a dimension of D
p
×R
iris
as an interest region. We
construct the key point descriptors from r
i
and generate a full
feature vector by concatenating all the descriptors. The feature
representation using partitioned image is regarded as a local
feature representation in some image retrieval literature [16,
17]. However, we consider this as a global representation
because all the pixel values are used in the representation
without considering the local distinctiveness of each region.
Mikilajczyk et al. [10] have categorized the descriptor types
as distribution-based, spatial frequency-based, and
differential-based. We use two well known distribution-based
descriptors: gradient orientation (GO) histogram and local
binary pattern (LBP) [18]. We quantize both GO and LBP into
8 distinct values to build an eight bin histogram. The eight bin
histogram is constructed from a partitioned sub-region and
concatenated to construct a full feature vector. A Gaussian
blurring with a standard deviation
σ
is applied on both GO and
LBP to smooth variations across local pixel values. This
sub-partition based histogram construction scheme has been
successfully used in SIFT [12] for the object recognition
problem.
The local matcher first detects a set of salient key points in
scale space. Features are extracted from the bounding boxes
for each key points based on the gradient magnitude and
orientation. The size of the bounding box is proportional to
the scale (i.e., the standard deviation of the Gaussian kernel in
scale space construction). Fig. 4 shows the detected key points
and surrounding boxes on a periocular image. While the
global features are only collected around the eye, the local
features are collected from all salient regions such as facial
marks. Therefore, it is expected that the local matcher
provides more distinctiveness.
To appear in Biometrics: Theory, Applications and Systems (BTAS 09), Washington DC, September 2009
Once a set of key points are detected, these points can be
used directly as a measure of image matching based on the
goodness of geometrical alignment. However, such an
approach does not take into consideration the rich information
embedded in the region around each interest points.
Moreover, when there is affine transformation or occlusion it
will be beneficial to match individual interest points rather
than relying on the entire set of interest points. We used a
publically available SIFT implementation [19] as the local
matcher.
D. Matching Scheme
For the global descriptor, the simple Euclidean distance is
used to calculate the matching distance. The distance ration
based matching scheme [12] is used for the local matcher
(SIFT).
E. Parameter Selection for Each Matcher
The global descriptor varies depending on the choice of
σ
and the frequency of sampling of interest points, 1/D
p
.
SIFT has many parameters that affects its performance.
Some of the representative parameters are the number of
octaves (n
o
), number of scales (n
s
), and the cut-off threshold
value, t
ex
, related to the contrast of the extrema points. The
absolute value of each extrema point in the Difference of
Gaussian (DOG) space needs to be larger than t
ex
to be
selected as a key point.
We construct a number of different descriptors for both the
global and local schemes by choosing a set of different values
for
σ
, D
p
, n
o
, n
s
, and t
ex
. The set of parameters that results in
the best empirical performance is selected to be used for the
global and local representations.
III. E
XPERIMENTAL RESULTS
A. Database
We collected 899 high-resolution face images from 30
different subjects in two different sessions (450 in session 1
and 449 in session 2, 14~15 images per subject in each
session) using a Canon EOS 5D Mark II camera. The camera
parameters were set to the following options: maximum
resolution (21.1 Mega pixels - 5616×3744), Auto-Focus,
Optical Vibration Reduction Image Stabilization, Portrait
Mode, ISO AUTO, and JPEG format. Each subject was asked
to sit ~4 feet away from the camera during data acquisition.
We manually cropped the periocular region from each face
image in two different ways: with and without eyebrows.
Some example periocular images are shown in Fig. 1. The
sizes of periocular images are in the range [419,892] for width
and [182,400] for height with no eyebrow. The periocular
images with eye brow shows height in the range of [265,713]
with the same width range as those without eyebrow.
We assembled two different databases, DB1 and DB2, for
the periocular recognition experiments. DB1 consists of 120
images with two (left and right eye) periocular images per
subject per session. DB2 consists of 958 images with 898
probe and 60 gallery images. Probe dataset contains 28~30
periocular images per subject and gallery contains 2
periocular images per subject. DB1 is used for parameter
selection and DB2 is used for evaluating the matching
Fig. 6: Rank-
1 accuracies of the local matcher (SIFT) with different
choices of parameter: (a)(c)(e) without eyebrow and (b)(d)(f) with
eyebrow.
(
n
o
)
(n
s
)
t
ex
= 0.001
(a)
(b)
(n
o
)
(n
s
)
t
ex
= 0.001
(n
o
)
(n
s
)
t
ex
= 0.003
(n
o
)
(n
s
)
t
ex
= 0.005
(n
o
)
(n
s
)
t
ex
= 0.005
(n
o
)
(n
s
)
t
ex
= 0.003
(c)
(d)
(e)
(f)
Fig. 5: Rank-1 accuracies of the global matcher (GO and LBP) with
different choices of parameter: (a)(c) without eyebrow and (b)(d) with
eyebrow.
(σ)
(σ
)
(σ)
(a) GO, without eyebrow
(b) GO, with eyebrow
(c) LBP, without eyebrow
(d) LBP, with eyebrow
(D
p
)
(D
p
)
(
D
p
)
(D
p
)
(σ)
To appear in Biometrics: Theory, Applications and Systems (BTAS 09), Washington DC, September 2009
performance.
Table 1: Periocular recognition accuracy (%) with respect
to the use of eyebrows and side information.
Without eyebrow
With eyebrow
L or R
eye *
Same
eye **
L or R
eye
Same
eye
GO 52.5 49.2 62.5 60.8
LBP
56.7
50.8
70.0
66.7
SIFT 71.7 74.2 70.8 70.0
GO+SIFT
76.7
80.8
80.0
75.8
LBP+SIFT 76.7 80.8 80.0 78.3
GO+LBP+SIFT
73.3
77.5
80.0
79.2
* Left (Right) eyes can match with Right (Left) eyes
** Left (Right) eyes cannot match with Right (Left) eyes
B. Recognition Accuracy
The recognition accuracy using the aforementioned
periocular feature set is assessed using the Cumulative Match
Characteristic (CMC) curve. For DB1, given N (=120) images
I
1
, I
2
, …, I
N
, every image I
i
is taken as the query and the rest of
the images are used as the gallery. For DB2, separate set of
probe and gallery images are used. Matching experiments on
DB1 are performed with and without eyebrows, and with and
without the Left/Right (eye side) information. When the
Left/Right information is used, Left (Right) side periocular
image can only match to the Left (Right) side. To take
advantage of the characteristics of both global and local
descriptors, we used a fusion scheme that combines the global
and local information. We used a score level fusion based on
weighted sum with min-max normalization. The weights are
empirically selected for both the global and local matchers.
Fig. 5 shows the rank-1 accuracy of the global matcher
using GO and LBP descriptors based on different
configuration of the parameters. The best performance was
observed to be 62.5% and 70.0% for the GO and LBP
descriptors, respectively. The performance of both GO and
LBP shows a dependency on D
p
rather than
σ
. The use of
eyebrow showed better recognition accuracy for both the GO
and LBP descriptors.
Fig. 6 shows the rank-1 accuracy of the SIFT matcher.
Larger values of t
ex
and n
s
than those shown in Fig. 6 resulted
in lower accuracy. With a large value of n
s
, the standard
deviation of the Gaussian kernel increases by a small amount
when constructing the scale space, resulting in smaller values
across the DOG space. This has a similar effect as increasing
t
ex
, which also decreases the matching accuracy. Larger n
o
helps in improving the matching accuracy, in general. The
accuracy decreases with the use of eyebrows. We believe this
is due to the noisy keypoints detected around the eyebrow,
which results in false matches thereby inflating the imposter
matching scores. The best rank-1 performance is obtained as
74.2% with no eyebrow and using information about the
location of the periocular region (i.e., left or right eye).
The matching accuracy of the best global matchers, local
matcher, and the resulting fusion schemes on DB1 and DB2
are shown in Fig. 7. The best performance is observed to be
80.8% and 77.3% by fusing the LBP based global matcher
with SIFT on DB1 and DB2, respectively.
The first five rows of Fig. 8 show examples of image pairs
that were not correctly matched at rank-1 by the LBP and
SIFT schemes but were correctly matched after fusion for the
first five rows. The last two rows show failure cases both
before and after fusion. Periocular images from different
subjects appear similar in the last two rows, resulting in the
false matches.
IV. C
ONCLUSIONS AND FUTURE WORK
We have proposed a method for using periocular images as
a biometric trait. Both global and local descriptors have been
explored for feature extraction and matching. Further, a
score-level fusion scheme was employed to enhance the
recognition accuracy. Based on the evaluation of a total of 958
Fig. 7: CMC curve of the global, local, and fusion matchers on (a) DB1
and (b) DB2.
(a)
(b)
