ICGST-BIME Journal, Volume 7, Issue 1, May, 2007
Biometric Template Classification: A Case Study in Iris Textures
E. Srinivasa Reddy and I. Ramesh Babu
Department of Computer Science, Acharya Nagarjuna Univerity, Guntur, A.P.,India
edara_67@yahoo.com, rinampudi@yahoo.com
Abstract
Most of the biometric authentication systems store
multiple templates per user to account for variations
in biometric data. Therefore, these systems suffer
from storage space and computation overheads. To
overcome this problem the paper proposes techniques
to automatically select prototype templates from iris
textures. The paper has two phases: one is to find the
feature vectors from iris textures that have less
correlation and the second to calculate DU measure.
DU measure is an effective measure of the similarity
between two iris textures, because it takes into
consideration three important perspectives: a)
information, b) angle and e) energy. Also, gray level
co occurrence matrix is used to find the homogeneity
and correlation between the textures.
Keywords Shaker iris, Jewel iris, Flower iris, Stream
iris, gray level co-occurrence matrix, Spectral
information divergence, Spectral angle mapper, DU
measure
1. Introduction
A typical iris biometric system operates in two
distinct stages: the enrollment stage and the
authentication stage [2]. During enrollment stage iris
textures are acquired and processed to extract a
feature set. The complex iris textures carry very
distinctive information. The feature set includes some
distinguished features like nodes and end point in the
iris texture. The stored feature set, labeled with the
user’s identity, is referred as a template. In order to
account for variations in the biometric based
authentication system relies on the stability from an
individual is susceptible to changes due to distance
from the sensor, poor resolution of the sensor and
alterations in iris texture due to cataract operations
and others.
Multiple iris features pertaining to different portions
of iris must be stored in data base. There is a trade off
between the number of templates, and the storage and
computational overheads introduced by multiple
templates. An efficient system must in fact classify
the templates automatically.
17
Past Work Done The recognition ability of classifiers
depends on the quality of feature used as well as the
amount of training data available to them. Image
features are mostly extracted on shape and texture of
segmented objects. Karu et al. [15] define a
methodology for automatically identifying textured
regions within an image so that feature extraction
algorithms are only used where texture can be
quantified. The classification is mainly based on
statistical measures used for texture classification
such as DU measures, autocorrelation, co-occurrence
matrices etc. Textured methods used can be
categorized as: statistical, geometrical, structural,
model-based and signal processing features
[13].Randen and Husoy conclude that most studies
deal with statistical, model-based and signal
processing techniques. Co-occurrence method is
proved best for texture classification by Conners and
Harlow [14].
No work has been carried out in classification of iris
textures to overcome the overheads due to space and
time for high speed recognition, until today. This
paper tries to classify iris textures on guidelines of
statistical texture classification methods as discussed
above.
The remainder of this paper is organized as follows:
Section 2 briefly summarizes basic types of iris
textures. Section 3 describes the necessity of template
selection and updating. Section 4 describes statistical
measures used to define the relation between pixels.
A detailed description of the proposed measurements
used to classify the textures is given in Section 5.
Section 6 and 7 reports the experiments and discuss
the results and Section 8 presents the detailed flow
chart of the proposed algorithm. Finally Section 9
presents the conclusion.
2. Types of Iris Textures
Depending on the texture of different human iris, we
can group the irises into four basic groups. They are
a) Stream iris, b) Jewel iris, c) Shaker iris and d)
Flower iris. Also, there are different combinations of
 ICGST-BIME Journal, Volume 7, Issue 1, May, 2007
there groups. As a preliminary study we have
concentrated to group the given data base into the
above four classes.
a) Stream Iris: It contains a uniform fiber structure
with subtle variations or streaks of color as shown in
figure 1.c. The structure of the iris is determined by
the arrangement of the white fibers radiating from the
center of the iris ( or pupil). In this image one can
notice that they are uniform and reasonably direct or
parallel.
Figure 1.a & b: Stream Iris
Figure 1.c: Stream iris Texture
b) Jewel Iris: It contains dot-like pigments in the iris.
The jewel iris can be recognized by the presence of
pigmentation or colored dots on top of the fibers as
shown in figure 2.c. The dots (or jewels) can vary in
color from light orange through black. They can also
vary in size from tiny (invisible to the naked eye) to
quite large.
Figure 2.a & b: Jewel iris
Fig 2.c: Jewel iris Textures
c) Shaker iris: It contains dot-like pigments and
rounded openings. The shaker iris is identified by the
presence of both flower like petals in the fiber
arrangement and pigment dots or jewels is shown in
figure 3.c. The presence of even one jewel in an
18
otherwise Flower iris is sufficient to cause the
occupant to exhibit shaker characteristics.
Figure 3.a & b: Shaker iris
Figure 3.c: Shaker iris textures
d) Flower iris: It contains distinctly curved or
rounded openings in the iris. In a flower iris the fibers
radiating from the center are distorted (in one or more
places) to produce the effect of petals (hence the
name flower) shown in figure 4. In this image one
can notice that they are neither regular nor uniform. A
flower iris may have only one significant petal with
the remainder of the iris looking like a stream.
Figure 4.a & b: Flower iris
Figure 4.c: Flower iris textures
3. Template Selection and Updating
The size of the same iris taken at different times may
be different in the image as a result of changes in the
camera to face distance. Due to simulation by light or
for other reasons, such as hippos, natural continuous
movement of the pupil, the pupil may be constricted
or dilated [3]. The problem of template selection
means, given a set of N iris images corresponding to
a single human iris, selected k templates that best
represent the similarity observed in N images. The
selected templates must be updated time to time,
since most of the users may under go cataract
operations due to aging. To account for such changes,
old templates must be replaced with newer ones.
 ICGST-BIME Journal, Volume 7, Issue 1, May, 2007

Some of the minutiae points like nodes and end d) Homogenity measures the closeness of the
points of iris textures may be added or deleted from distribution of elements in the GLCM to the
the iris template. Thus template selection refers to the GLCM diagonal.
process by which iris textures are chosen from a
given set of samples, where as template update refers We have derived only correlation to identify the
to the process by which existing templates are length of the feature vector which has unique
replaced. information from iris texture. A typical iris exhibits
rich texture information in the immediate vicinity of
4. Selection of feature vectors the pupil which tapers away in intensity as one moves
GLCM(gray-level co-occurrence matrix)[4], also away from the pupil. Thus, the iris templates are
known as the gray-level spatial dependence matrix, is taken by segmenting a portion of iris texture near the
a statistical measure used to characterize the texture pupil and from them the feature vectors are derived,
of an image by calculating how often pairs of pixel, by using correlation function. This is implemented by
with specific values and in a specified spatial performing canny edge detection technique to
relationship occur in an image and then by extracting produce binary image or respective iris texture and
statistical measures from this matrix. also by calculating the probability of occurrence of
By default, the spatial relationship is defined as the pixels.
pixel of interest and the pixel to its immediate right
(horizontally adjacent), but one can specify other
spatial relationships between the two pixels. Each
element (i,j) in the resultant GLCM is simply the sum
of the number of times that the pixel with value I
occurred in the specified spatial relationship to a pixel Figure 5.a Figure 5.b
with value j in the input image.
Initially, the dynamic range of the given image is
scaled to reduce the number of intensity values in
intensity image from 256 to 8. The number of gray-
levels, determine the size of the GLCM.
The gray-level co-occurrence matrix can reveal Figure 5.c Figure 5.d
certain properties about the spatial distribution of the Fig:5 Iris textures after canny edge detection
gray levels in the texture image. For example, if most
of the entries in the GLCM are concentrated along the a) Stream iris b)Jewel iris c0Shaker iris d)Flower
diagonal, the texture is coarse with respect to the iris The correlation Vs offset plot for different
specified offset. iris textures are given below.
However, a single GLCM might not be enough to
describe the textual features of the input image. For
example, a single horizontal offset might not be
sensitive to texture with a vertical orientation. For
this reason, gray co-matrix [5] can create multiple
GLCMs for a single input image. TO create multiple
GLCMs, specify an array of offsets to the gray co-
matrix function. These offsets define pixel
relationships of varying direction and distance. For,
example, one can define an array of offsets that
specify four directions (horizontal, vertical and two
diagonals) and four distances. In this case, the input
image is represented by 16 GLCMs. The statistics are Figure 6.a Correlation of Shaker iris texture
calculated from these GLCMs and then the average is
taken.

Statistical description of GLCM:

The statistical parameters that can be derived for
GLCM are contrast, correlation, energy and
homogeneity.
a) Contrast measures the local variations in the
gray-level co-occurrence matrix.
b) Correlation measures the joint probability
occurrence of the specified pixel pairs.
c) Energy provides the sum of squared Figure6.b Correlation of Jewel iris texture
elements in the GLCM.

19
 ICGST-BIME Journal, Volume 7, Issue 1, May, 2007

Du et al[26] developed SID-SAM mixed measure. It

is defined by a mixed measure,
M = SID(f,g) * tan(SAM(f,g)). ……. (3)

c) Energy: It is measured in terms of average power

difference. Instead of using 2-norm, we can use 1-
norm. APD (average power difference) is given by,

APD(f,g) = ||f-g|| *1/N, where ||x||1=Σ|x|. ……. (4)

Figure 6.c Correlation of Stream iris texture Finally DU measure is given by,

Du(f,g)=APD * Mixed measure . ……. (5)

6. Iris Classification
When a new iris texture is to be grouped into a
predefined class, the Du measure is calculated with
any one of the typical templates in that class. If the
Du score is very high, we can say that it does not
belong to that class. The new template belongs to a
class, with which its Du measure is very low in
comparison to others [1].

Figure 6.d Correlation of Flower iris texture 7. Experimental Results

We have taken for our study the iris image data base
From the above correlation plots, we can see the peak from CASIA iris image Data base[CAS03a] and
offsets which represent the periodic patterns that MMU iris data base[MMU04a]. CASIA iris image
repeat for every set of pixels. Correlation function data base contributes a total number of 756 iris
can be used to evaluate the coarseness. Larger texture images which were taken in two different time
primitives give rise to coarse texture and small frames. Each of the iris images is 8-bit gray scale
primitives give fine texture. However if the primitives with resolution 320 x 280. MMU data base
are periodic, then the correlation increases and contributes a total number of 450 images which were
decreases periodically with the distance [12]. For captured by LG Iris Access ® 2200.
example, in the case of flower iris, the peak offsets
are at 9 and 13 pixels. But, a common offset is Our studies conclude that all the iris images in the
observed at 9 for all textures. So, we can take a 9 database have four different types of textures: Flower,
pixel nearest neighbor as a feature vector to estimate Jewel, Stream and shaker. Thus there occur 6
DU measure in the case of iris textures. different combinations for matching. Hence SID,
SAM, APD values, for these 6 combinations, are
given in the table.
5. Du Measure
Du measure[1] is an effective measure of the
Combination SID SAM MM APD DU
similarity between two iris textures, because it takes
Stream 1.51657 0.34578 0.54618 8.3483 4.55997
into consideration three important perspectives: Vs
a) Information: It is measured in terms of SID Jewel
(spectral information divergence) between probability Stream 1.60659 0.26289 0.43236 14.859 6.4244
mass functions of the two textures f, g. Vs
Flower
SID(f,g ) = D(d||g) + D(g||f); ……. (1) Stream 1.8626 0.29232 0.56053 4.4424 2.4901
Here D(f||g) is the relative entropy of g with respect Vs
to f. Shaker
Where D(f||g) = Σ (f* log(f/g)) and D(g||f)= Σ (g * Flower 1.4104 0.40956 0.6123 6.491 3.9744
Vs
log(g/f)) Jewel
b) Angle: It is measured in terms of spectral angle Flower 2.08906 0.27165 0.58188 10.397 6.0498
mapper. It measures the angle between two vectors f Vs
and g given by, Shaker
SAM (f,g) = acos(<f,g>/(||f||*||g||) ……. (2) Jewel 2.1463 0.38986 0.88193 3.9059 3.447
Vs
Where <f,g> = Σ fi*gi, and ||f||2 is the 2-norm of Shaker
vector f, ||f||2=<f,f>1/2
Mixed Measure:
Table 1: Du measures for different combinations of
iris textures.

20
 ICGST-BIME Journal, Volume 7, Issue 1, May, 2007

When a new template belongs to the same class, then

the SID (relative entropy) and APD (average power
difference) become zero; thus Du measure
automatically becomes zero or very less. h1=√Σf(i,j)
g1=√Σg(i,j)
8. Flow Chart for DU measures k1=Σfg(i,j)
a) Flow Chart to find SID
Let f & g be two iris templates

Read two iris templates

SAM=acos(k1/h1*j1)
to be compared, f &g

Find the probability of pixels in the

STOP
given templates
f=f/Σ(f), g=g/Σ(g)

c) Flow chart to find Mixed Measure

SID SAM
Find the relative entropy of each
template,
e1= f*log2(f/g);
e2= g*log2(g/f);
MM=SID*Tan(SAM)

SID=e1+e2
Stop
Stop
d)Flow chart to find APD
If f & g are two iris templates
b) Flow chart to find SAM
Read two iris templates Read two iris templates
to be compared, f &g to be compared, f &g
N size of template N size of template

i=1,j=1
h=f-g;
h1=Σh;

i=N&
j!=N APD=h/N

Stop
f(i,j)=f(i,j)*f(i,j)
g(i,j)=g(i,j)*g(i,j)
fg(i,j)=f(i,j)*g(i,j) 9. Conclusion
In this paper we mainly concentrated on basic
classification of iris textures into four groups. Our
results show that there is a considerable variation in
DU measure for different combinations which are yet
to be standardized with a large data base. Future work
includes further classification of iris textures into sub

21
 ICGST-BIME Journal, Volume 7, Issue 1, May, 2007
groups. Also not only DU measures, but statistical
texture methods such as auto correlation, edge
frequency, primitive lengths, Law’s texture features
can be used for more efficient and accurate
classification.
10. References
[1] NIST, Advanced Encryption Standard
(AES),2001.http://csrc.nist.gov/publications/fips/
fips-197.pdf
[2] H.Heijmans, Morphological Image Operators,
AcademiPress, 1994.
[3] A.Juels and M.Sudan, “A Fuzzy Vault Scheme”,
proc. IEEEInt’l. Symp.Inf. Theory,A
Lapidothand E.Teletar,Eds.,pp.408,2002.
[4] R.C.Gonzalez amd R.E.Woods, Digital Image
Processing, Addison-Wesley, 3rd ed.,1992.
[5] Joaquim De Mira Jr, Joceli Mayer, “Image
Feature Extraction for application of Biometric
Identification of Iris – A Morphological
Approach”. proc IEEE Int’l Symp on Computer
Graphics and Image processing, SIBGRAPI’03
[6] Umut Uludag, and Anil K.Jain, “ Fuzzy Finger
Print Vault”, Proc. Workshop: Biometrics:
Challenges Arising from Theory to practice,
pp.13-16,2004,W.H.Press
[7] S.A.Teukolsky, W.T.Vetterling, and
B.P.Flannery, Numerical Recipes in C, 2.
Ed., Cambridge University press, 1992.
[8] Chuan Chin Teo, Hong Tat Ewe “An efficient
One Dimensional Fractal Analysis for Iris
Recognition”WSCG’2005, January 31
Feb4,2005,Plzen,Czech Republic.
[9] LiMA,Tieniu, “Efficient Iris Recognition by
Characterizing Key Local Variations”, IEEE
trans., Image Processing-2004.
[10] Robert Ives, Delores Etter, Yingzi Du, “Iris
Pattern Extraction using Bit Planes and Standard
Deviations”, IEEE conference on Signals,
systems and computers, 2004.
[11] Qi-Chuan Tian, Quan Pan, Yong-Mei Cheng“
Fast algorithm and application of Hough
Transform in iris segmentation”, proceedings of
third IEEE conference on machine learning and
Cybernetics, Shangai, 26-29, 2004 .
[12] Mona Sharma, Markos Markou, Sameer Singh
“Evaluaion of Texture Methods For Image
Analaysis”, pattern recognition letters.
[13] T.R.Reed and J.M.H.Buf, “A review of recent
texture segmentation and feature extraction
techniques”,Computer visison, Image Processing
and Graphics, vol.57,no. 3,pp. 359-372,1993.
22
[14] R.W.Conners and C.A.Harlow, “A theoretical
comparison of texture algorithms”, IEEE
transactions on Pattern Analysis and Machine
Intelligence, vol2,no. 3,pp.204-222, 1980.
[15] K.Karu, A.K.Jain and R.M.Bolle, “Is there any
texture in the image?”, Pattern Recognition, vol.
29,no. 9,pp-1437-1446,1996.
Biographies
Srinivas R.E. received the
B.Tech degree in Electronics &
Communication ENgienering
from Nagarjuna University,
India in 1988, M.S. degree from
Birla Institute of Technology
and Scince, India in 1997,
M.Tech degree in Computer
Science from Visveswaraiah Technological
University, India in 2000 and pursuing his Ph.D in
computer science in Nagarjuna Univeristy under the
guidance of Dr. I. Ramesh Babu. He is the member of
IEEE and presented 8 papers in international
conferences. His research interests includes image
processing, biometrics and pattern recognition
Ramesh Babu.I received the
B.Tech degree in Electronics &
Communication Engineering
from Mysore University, India in
1981, M.Tech degree in
Computer Engineering from
Andhra University, India in 1984
and Ph.D degree in computer Science & Engineering
from Nagarjuna University, India in 1994.
He is currently working as Head & Professor in the
department of computer science, Nagarjuna
University. Also he is the senate member of the same
University from 2006. His areas of interest are image
processing & its applications, and he is currently
supervising 10 ph.D students who are working in
different areas of image processing. He is the senior
member of IEEE, and published 35 papers in
international conferences and journals.