Iris Recognition February, 2015
Iris Recognition February 2015
IRIS RECOGNITION
Muhammad Usman
University of Management and Technology (UMT)
Lahore, Pakistan
E-mail: [email protected]
Abstract
Iris Recognition has gauged much attention for over a past few decades. No doubt, it is one of the most
accurate domains as far as security is concerned. People had worked a lot in this area. It deals with all, starting
from acquiring image from hardware devices to matching iris images using some mathematical, statistical,
probabilistic or intelligent models. This article evaluates and compares among few of the well-known state of
art iris recognition models.
1. Introduction
Authentication and authorization is the key
element in any secure of supervised environment.
The domain of Iris recognition has been gauging
attention for over a past few decades. With
advancements in iris matching algorithms, the
broader applications and new technology is
witnessing the demand of related tools and
applications. These systems are based on the basic
requirement of user to achieve authentication/
authorization in a trusted and accurate manner.
Previously, identity cards, passwords or some other
document used for this purpose. Such a procedure
was based on two queries. First is “Who you are? And the Second is “What you have?” This type of information creates dependency in Authentication
and authorization. They might expect a card, that
must be carried permanently thus security is weak
and prone to failure. To counter this problem,
biological features are used to perform authorization
and authentication. Various techniques and models
have been used to get feature extracted, as in this era
it is easy to perform such computations on a small
cheaply available device. Scientists have used
various biologicals features of a person for this
purpose such as gait, iris, thumbprints, speech, facial
features, vein structure, ear pattern, fingerprints,
sutures and etc. Biometric methods typically perform
numerous steps to yield an outline and generate a
specific code that contains peculiar information about
the biological features in a numeric and succinct
form. Ultimately a vector of such features enables a
system to uniquely recognize each user with the help
of a classifier.
As discussed earlier, Biometrics based on physical
or behavior characteristics are used to uniquely
Iris Recognition February, 2015
Iris Recognition February 2015
identify an individual. With the change of
individual’s requirement, iris recognition has been
considered the most accurate among most of
authorization and authentication techniques. Human
iris has an unusual construction and provides rich
texture information. This textural information
embedded in the iris uniquel for each human and is
inimitable over time, hence it is preferred choice as
tool for identification of an individual. Dissimilarities
that exist in the biological features seem to alter over
time due to factors like growth and aging. As
compared with other biometric features iris is the
most stable and reliable for identification.
Identity evasion is a serious concern and it’s a tremendous challenge to robust fool proof
identification system. Use of iris for identification
provides a great dividend since iris is unique, remains
even and secure throughout life period. Speed,
simplicity and precision are main benefits of iris
recognition system. It is an efficient technique and
conferring to conditions its error rate is also less.
2. RELATED WORK:
2.1 LMD (Local Mean Decomposition)
Iris can be recognized using LMD. It is totally a data-
driven method and it does not have any permanent
filter. Image preprocessing extracts only the iris features from the entire eye image. Wei-Yu Han [8]
worked on this domain. There are three parts in
which this extraction can take place. This technique
can be used to observe many data types. It can decay
a signal into different parts which we say product
functions. Tieniu Tan suggested [5] almost the same
algorithm with another technique. His technique is
based on textures of iris and even and obvious
lightning. In this paper, Algorithm used was “Multi –
modal fusion”, in which it takes the entire eye as a biometric pattern that contains the information of iris.
Moreover, structure of eye region also returns
important data that can be used as secondary
personas to match noisy iris image. Framework for
such environment is well anticipated here.
2.2 Fuzzy Integration
Fuzzy integration can also be helpful to recognize
iris. Fuzzy logic [2] opens a gate for a system to the
reason for the uncertainty. It is a helpful tool for
modeling composite machines. Though, it is time and
again not easy for developers to define the sets and
rules used by systems. In that account, a genetic
algorithm provides us with a solution that finds not
only the architecture but returns a fair suitable
recognition rate. In one of research papers, Patricia
Melin stated in [2]that, once it is confirmed that the
genetic algorithm can produce a sound optimization
result, in which the algorithm runs ten times to find
Standard Deviation and the number of layers.
Summary of this exercise with the results of the 10
experiments on the basis of comparisons with
different optimization methods is presented in her
work. Feature Extraction Methods for Iris
Recognition is a useful technique to get iris
recognized. It gets the image, recognizes the by
separating the iris portion from the entire eye, which
is referred as segmentation. In third step segmented
iris is normalized. Then available samples in database
are matched with the iris. Hamming distance can be
used to calculate the distance between two iris
templates.
2.3 Ridgelts Transformation
Another useful technique is to apply Ridgelet
transform on Iris. Ridgelet transforms [6] are the
mixture of Radon transforms and Wavelet
transforms. They are appropriate for extracting the
plentifully present textural data that is in an iris. The
method anticipated here uses the ridgelets to form an
iris mark and to symbolize the iris. Author
contributes towards creating an enhanced iris
Iris Recognition February, 2015
Iris Recognition February 2015
recognition system. Experimental has shown
promising results that states the precision of 99.82%.
2.4 DRS (Daugman’s Rubber Sheet)
Mr. P.P.Chitte suggested that Daugman’s Rubber Sheet [11] Model is a healthy technique for
respective domain. In his paper, he suggested iris
blend technique based on (ICA), (PCA), and
Daugman’s rubber sheet. Algorithm has confirmed to be progressively more precise and dependable.
Author has compared results obtained from all three
algorithms and has selected the best technique for
respective area. Real- time Iris Segmentation is a
technique used by “Juan Alejandra” which maps the iris texture in Daugman’s doubly coordinates. Author has also applied an adaptive Hough transform to
calculate the estimated centre of the iris. Based on the
result of the first, polar transform finds the first papillary boundary whereas anellipsopolar transform
finds the second. Here both images for visible wavelength can be processed in an identical conduct.
2.5 Geometric Key Based Algorithms
Iris Recognition using stabilized iris Encoding and
using geometric key based Iris Encoding are purely
decision based algorithms. A nonlinear approach for
concurrently explanation for local uniformity of iris
bit and the overall quality of the weight is used in
these algorithms. The technique efficiently penalizes
the fragile bits while all together it satisfies consistent
bits. Through Experiments the proposed approach
can root major improvement in iris matching
accuracy. Iris can also be identified by using neural
networks. In this method after getting and processing
the image, distance of iris between left to right and
from top to bottom is calculated. At last, they used
neural network for to train and test the algorithm. The
best accuracy is 97.5%. Iris Recognition based on
Local Mean Decomposition is also a good technique
for matching and recognition. It is done in four
simple stages: a) Quality assessment of Image, b)
Preprocessing the image, c) Extracting the features
out, d) Matching image with database. To assess the
performance of this approach, several similarity
measures are used to view the results based on
experiments using both the CASIA and ICE iris
image databases.
3. TECHNICAL FRAMEWORK
A basic Iris recognition building consists of three
pillars. After getting the image, first of all Iris region
is segmented, secondly some transformation is
applied to get the features extracted and lastly the
image is matched/ classified through some database.
The process is well explained in the figure below.
Iris Recognition
Iris SegmentationTraining &
Classification
Transformation &
Feature Extraction
Mathematical
Model(Alcohol
Assumption)
NIR(Image Near-
Infrared)
Fuzzy Integration
Corner Detection
Binary Segmentation
Method
Iris bits Stabilization
Encoding
Geometric Key based
Iris Encoding
LMD (Local Mean
Decomposition)
Multi-Model Fusion
Fuzzy Integration
Simple Mean
calculations
Ridgelts
Gabor Filter Bank
Daugman’s Rubber Sheet Model
Geometric key based
Iris Encoding
Neural Network
algorithm
Decision based
algorithm
Zernike Moments
Three level layered
architecture
Zernike Moments
Neural Network
algorithm
Decision based
algorithm
Figure 1: Iris Recognition
3.1 Alcohol consumption based
Iris Recognition February, 2015
Iris Recognition February 2015
Author named RichaSingh [1,2] described usage of
alcohol results in the dilatation/constricting the pupil
which causes change in texture ending with affecting
the recognition performance. If � � Is the image of
user “U” taken before alcohol consumption and �
is iris image after alcohol consumption. Let �� � � and �� � � be the distance
between the major and minor axis of pupil and �� � and �� � the distance
between major and minor axis of iris boundaries. The
relation below defines the area.
�� � � = � × �� � � × �� � �
�� � = � × �� �× �� �
Likewise, the area of the iris boundary can be
calculated by,
� � � = � × � � � × � � �
� � = � × � � × � �
Experiments performed to test alcoholic usage are
given below.
1. Matches pre alcohol iris images.
2. Matches post alcohol iris images.
3. Matches pre and post alcohol iris images.
The first two were done for achieving precision,
where as the third experiment is performed to test the
intake of alcohol. This technique can be used for
cooperative systems as well as non-cooperative ones.
First of all, the original images are preprocessed.
Then, ordinal actions and color histogram in the
image are identified to distinguish iris data, textual
representation and semantic label plays their part to
get eye patterns. Then, 4 matching scores are
obtained by different algorithms. Finally, a union
strategy is applied to make the final difference score.
They used 'IIITD Iris under alcohol influence'
database for comparison of different datasets. This
database consists of 220 images of iris that were
taken before alcohol consumption and 220 images
after alcohol consumption from different 55 subjects.
Subjects then consists of different age groups. This
database shows 95% accuracy rate.
3.2 Fuzzy Integration
Other technique named as fuzzy integration plays a
vital role to recognize and individual by Iris. Modular
neural uses iris image and it can be trained by a
model. Author uses genetic algorithms with multiple
techniques, like: gating network or fuzzy integration.
The Image preprocessing is done by applying
comparison wavelet 2D technique. Furthermore
Image is also enhanced by leaving the iris behind.
Figure 2: Iris Enhancement
Once preprocessing is done, the image is compressed
in to matrix that uses wavelet transform with some
Iris Recognition February, 2015
Iris Recognition February 2015
levels of decomposition.
Figure 3: Image Compression
Once this process is done, fuzzy integration is
applied on it.
They have used CASIA database for comparison
process. They took three modules which consists
different number of images. And then run the
algorithm ten times on each of the module to get the
best accuracy rate and high efficiency. They
concluded that using three layered neural networks
for iris recognition shows the best recognition rate of
99.37%.
3.3 Haar Wavelet Transform
Dolly Choudhary and Ajay Kumar Singh suggested
another feature extraction method for iris recognition.
They presented it after a long study on feature
extraction algorithms. Authors have discussed
feature extraction methods in which they have used
feature encoding along with corner detection. They
described Haar wavelet transform that is supported
with programmed Gabor filter Bank with some
statistical pattern based on recognition model.
Following are the steps that are involved in the whole
process.
a. Corner Detection Based Iris Encoding
Author captured a very good quality image and then
the texture of iris is stored. Estimated distance
between user and source of light is taken as 12 cm.
Now localization step involved two parts which are
calculation of vertical and horizontal distances such
as:
�� = {− − − ; ; }
� = {− ; ; − }
b. Feature extraction using Haar wavelet
Here, G vertical is obtained by convolving image
with C vertical and G horizontal is obtained by
convolving of image with C horizontal. � = ��� + � �
c. Comparison
Results are compared by using different
mathematical models. The image obtained is finally
Image Acquisition
Segmentation
Normalization
Feature Extraction
Matching
Iris Recognition February, 2015
Iris Recognition February 2015
divided in five levels based according to iris texture.
Figure 5: Image Division
Above is an iris image up to five level patterns cD1,
cD2, cD3 and cD4 are almost same so one of them
may be selected to decrease redundancy. Since cD4h
is identical to the previous one and it is the smallest
in size, it can be taken as a diplomat of all the
information the four levels carry. The last level does
not hold the same textures and it should be chosen as
a whole.
Authors used CASIA database for the comparison of
images. They used accuracy rate of 95.4% by using
two systems of FRR and FAR. High accuracy rate
was achieved by using three training set of iris
images. Authors took 60 images from different
subjects and then classified them in 10 different
groups for better recognition rate as well as for
achievement of high performance.
3.4 Daugman’s Rubber Sheet
Daugman’s Rubber Sheet is also one of a healthy technique for feature extraction. It is carried out
through variations characterized by the look and loss
of a significant image. A spatial filter bank is used
for feature extraction and matching shown below.
Figure 6: Rubber Sheet Algorithm
The aim behind the process is to obtain a code for iris
recognition. The process starts with an outline of iris
by explaining feature extraction. Image is processed
in two steps. First step is Graying and second is edge
detection. Then, a certain threshold is applied to the
resultant image. Furthermore, image is normalized by
applying Daugman’s rubber sheet model. The above
image shows the steps applied.
Authors used UBIRIS database for concluding their
results. They took 230 dataset of 5 images per
dataset. These 5 images were of same person and
hence they used data of 230 different people of
different ages. Accuracy rate was 98.79%.
3.5 Ridgelets
LeninaBirgale and ManeshKokare suggested that
ridgelets can also be used for feature extraction.
Mask can be designed by using following algorithm:
Iris Authentication
Image Processing
Graying
Segmentation
Edge Detection
Thersholding
Circular Hough Transform
Normalization
PCA?ICA?Daugman
Thersholding
Bit Pattern
Pattern Matching Encryption/
Decryption of Data
Iris Recognition February, 2015
Iris Recognition February 2015
Iris Database Test Iris
Iris Database Test Iris
Iris Database Test Iris
Iris Database Test Iris
Matching
Matching Result
Figure 7: Algorithm
� � = ∑ ∑ � cos ��� + � , � sin ���������=
59= + ���
Where: � � = Area of the iris and the pupil. � � = Maximum radius of the iris. � , ���= Coordinates of the centre of the pupil � = Initial radius of the iris.
The wavelets relay on the scales of the point position
and ridgelets relate the scales of the line positions.
So, wavelets are very superior at representing point
singularities. But, when there is need to represent
singularities, wavelets fails and ridgelets predict the
better solution. The feature vector size used is 1X4.
It also skips normalization and uses the ridgelet
transform for feature extraction. To take out the
signature of the iris they have calculated the first
level energy by using the following equation,
����� = = ∑ ∑ | � , |−=
−=
� , = Discrete function
Energy of above relation can be calculated by,
� = � ∑ ∑ , − �==
, = cth
wavelet decomposed sub-band. � = Size of wavelet decomposed sub-band. � = Mean value of cth
decomposed sub-band.
Euclidean distance metric is exercised to calculate to
equalize the value for the known pair of images in
database. “Zero distance” involves an ideal
counterpart and the image tells a difference as the
distance increases.
,� = √∑ − =
= ath
element of ith
database iris signature.
= ath
element of the jth
query iris signature vector.
Authors used three databases which are CASIA v1,
CASIA v3 and UPOL. They took experiments with
756 images by using CASIA v1, 660 images by using
CASIA v3 and 384 images by using UPOL database.
3.6 Hough and Ellipsopolar Transforms
Hough and Ellipsopolar Transforms used various
statistical models in process of iris recognition.
Proposed algorithm is binary segmentation method.
The first part finds a center O of an image by an ovel Hough transform, through tells us about concentric
rings. It then extracts a polar symbol by using O as
Iris Recognition February, 2015
Iris Recognition February 2015
origin. Here every border line edge has similar course,
which reduce the rate of edge detection.
An initial boundary Bi is noticed which takes input
by modeling a circle in Cartesian coordinate’s
gradient for each angle. It then smoothes the curve
obtained and ends with remapping to Cartesian
coordinates. Finally it fixes corner values by an
ellipse. This method is based on two hypotheses.
H0 ( B = P) , H1(B=L)
It maps concentric ovals along boundary“B”.AsP and L ideally should posses same center, so algorithm
should go with the normal values. Depending on the
result, e(energy) of one of the hypothesis is
discarded. If
e (P0) > e(L0) then H0 is rejected
ifP = P0, or L = L0.
It is now ready for rubber sheet model to be applied.
Authors used databases of Casia V3, Ubris.2, ICE,
FRGV and FERET for images comparison. They
took three different datasets for three different types
of databases. First dataset includes 1332 images,
second dataset includes 1000 images while third
dataset consists of 420 images. They have used
multiple algorithms for segmentation of iris on
datasets of images and after comparison of all
algorithms, they calculated that average accuracy rate
was recorded as 95% and minimum computation
time was 0.68s. They also determined the usability
rate of all algorithms and average usability rate was
92%.
3.7 Zernike Moments
Iris recognition using “Zernike Moments” is one of the fine decision based algorithms.
The iris segmentation begins with the image
enhancement. It uses retinex algorithm in which a
low pass filter is used that contains high frequency in
the enhanced image. The detected noisy pixels are
then packed with a binary mask to remove out all
such noisy pixels. The last step of iris segmentation is
estimation of noise such as eyelashes, shadow and
eyelids.
Authors concluded their results by using three
different databases such as UBIRIS.v2, FRGC,
CASIA v4-distance. They took two kind of images.
One which are visible in normal light and one which
are only visible using IR light. They took different
datasets for different database. 1015 visible images
were taken for UBIRIS.v2. 1234 visible images were
used for FRGC database. And 1016 IR images were
compared using CASIA v4 database. They concluded
UBIRIS.v2 database as best results shown. Accuracy
rate was high and performance was more than 90%.
CASIA was less usable because of high cost.
3.8 Decision Based-Occlusion Estimation
Image
Enhancement
Reflection
Detection
Coarse
Segmentation
Boundary
Refinement
Normalized Iris
image (Occlusion
Mask)
Occlusion
Estimation
Input Image
Iris & Pupil
LocalizationSegmented Iris
Normalization
Figure 9: Decision Based –Occlusion Estimation
In this technique, the decision is based on outputs of
two steps. First step is named as Iris bits stabilization
encoding. In which the obtained image is normalized
and then further preprocessed and mapped by using
nonlinear weighting strategy. Meanwhile image is
matched using a trained probability map which leads
through to the non linear weight estimation. Once the
features are extracted they are passed through a
weighted feature that ends in storing the image in
database. Second step is Zernike Moments Phase
Preprocessing &
MatchingFeature Extrraction Weighted Feature
Non – Linear
Weight Estimation Matching
Block Division &
Vectorization
Feature
Extraction(Zernike’s Moments)
Non- Linear Weight
Estimation
Zernike Moments
Phase Fratures
Matching
Scor
e1
Scor
e2
Final
ScoreDecision
Occlusion Mask
Normalized Iris
Image
Figure 8: Zernike Moments [1]
Iris Recognition February, 2015
Iris Recognition February 2015
Features. This section is further categorized in two
different parts. First the Normalized iris image is
vectorized and features are extracted through Zernike
Moments. Meanwhile, Occlusion mask obtained is
passed through under non – liner weight estimation.
Zernike Moments Phase features takes input as the
features extracted from Zernike moments and it
generates an output image that is used for cross –
phase matching. Finally images obtained from
algorithms are matched and scores are combined and
a trained model is used to take the decision.
They used three main databases for comparison such
as UBIRIS v.2, FRGC, CASIA.v4-distance. They
used three algorithms with different datasets and they
concluded that error rate was almost same while
performance was improved for algorithms.
Iris Recognition February, 2015
Iris Recognition February 2015
3.9 Gabor filter
Feature ExtractionLocal Mean
EstimationGeoKey Encoding
Geometric Key
Matching
Binary Features Matching Scor
e1Final
ScoreDecision
Phase
Quantization
Normalized Iris
Image
GeoKey
Transformation
Scor
e1
Figure 10: Gabor Filter Algorithm
Another simple approach for Iris Recognition is
using Geometric Key based Iris Encoding. In which
the normalized image is passed through a Log –
Gabor Filter.
Iris images can be recognized by decision based
algorithm. Image samples are collected and image is
then processed. Furthermore, the distance between
iris from left to right and top to bottom is calculated.
Biological neural network processing consists of
three steps. First the image is captured, and then the
respective image is processed by imagej tool for
feature extraction. Author suggested neural network
for classification. Feed forward back propagation
neural network are also produced here. A number of
neural networks are created for different group of
dataset, and finally the performance of each created
networks is measured.
Finally neural network is used for training and testing
purpose. Respective training algorithm and setting
multiple parameters for training and CASIA iris
database are used in this effort. The best accuracy
shown is 97.5%.
3.10 Neural Network Algorithm
Database
Image Processing
Feature Extraction
Length
Left to right
Breadth
Top to bottom
Dataset creation
Neural Network
model
Result
Figure 11: Neural Network Algorithm Flow
Figure 12: IR through Neural Network
It is one of the typically used algorithms. It has three
layers
1. Input layer
2. Output layer
3. Hidden layer
First two layers have some in between layers(hidden
layers), which helps in performing needed
calculations. First is the number of hidden layer and
second is the number neuron in each layer. Based on
these different training algorithm was designed.
Authors used three most frequently used databases
such as UBIRIS.v2, FRGC and CASIA-v4 distance.
Iris Recognition February, 2015
Iris Recognition February 2015
They took 1000 images for UBIRIS.v2 from 171
different subjects, 1085 images for FRGC from 149
different subjects and 935 images for CASIA from
131 different subjects.
4. Comparative Analysis
In the respective research, few methods are proposed
to recognize human iris. Though, every method has
pros and cons yet they are quite effective in terms of
accuracy and efficiency. Methods can be mapped
graphically as more efficient or more accurate. For an
overview, techniques discussed are listed below.
1. LMD (Local Mean Decomposition)
2. Multi-Model Fusion
3. Fuzzy Integration
4. Simple Mean calculations for feature
extraction
5. Ridgelets
6. Gabor Filter Bank
7. Daugman’s Rubber Sheet Model 8. Geometric key based Iris Encoding
9. Neural Network algorithm
10. Decision based algorithm
11. Iris recognition based on Zernike Moments
12. Three level layered architecture
Here some of the methods are decision based, few are
good in matching and few used receiver operating
characteristics to recognize human iris. Some of them
are based on retrieval values and a little of them uses
some classifier. To represent the accuracy and
efficiency of each algorithm, results from the each
respective algorithm are used as a source with the
same database.
On ROC curve, LMD, Fuzzy Integration and Neural
network algorithms showed healthy results. Since,
time complexity depends on the how complex the
mathematical operation used in the algorithm. Neural
Networks consists of three layers and for every layer
it performs calculations. So when it comes to time
complexity, it is not one of the suggested models.
Moreover, it uses more databases for training
purposes for all three layers.
Iris recognition based on Zernike Moments is a
purely precision recall technique in which the
algorithm splits in two different algorithms and at the
end scores are matched based on the decision.
Precision is how much selected items are relevant
whereas recall means how many relevant items are
selected. The algorithm achieves precision by
performing intersection between relevant and
received results. Mode of their result is then dividing
them with ones it retrieves. In recall, mode of the
result is divided by the relevant ones. The proposed
method has shown good precision and efficiency in
iris recognition.
LMD, Multi-Model Fusion and Simple Mean
calculation for feature extraction have the same time
complexity. Ridgelets, Rubber Sheet Model and
Geometric key based algorithm has shown promising
results in accuracy and space complexity. They use
less space and can produce an accurate iris match by
using CASIA – I database.
For comparisons, results produced in their algorithms
are used using recommended databases. If they are
mapped altogether, in terms of efficiency and
accuracy, they can be compared easily. Here neural
network is the best example of trade off between
accuracy and efficiency. Algorithms based on neural
networks showed good precision but when it comes
to efficiency (time complexity), they showed
ordinary performance.
Iris Recognition February, 2015
Iris Recognition February 2015
Figure 13: False Positives
Figure 14: True Positives
5. Conclusion
A fine comparison of above mentioned algorithms
for true positives and false positives is made.
Moreover, percentages for True Positives, False
Positives, efficiency and accuracy can also be
mapped altogether with the help of a bar graph.
Though every problem has its own area of concern
but to get the iris recognized, every algorithm has
used kind of same steps followed by different
mathematical models. Neural Networks used layered
architecture and Zernike Moments had used decision
based approach in which they calculate total of
collected scores and then decide for the sample to be
selected. Fuzzy integration is a quick way to get iris
recognized but it is not good when it comes to
address the issues like proper light or noise removal.
Our suggested approach is, if neural networks use the
mean value for calculating the layered architecture
False
Positives %
LMD (Local Mean
Decomposition)Multi-Model Fusion
Fuzzy Integration
Simple Mean calculations for
feature extractionRidgelts
Gabor Filter Bank
Daug a ’s Rubber Sheet Model Geometric key based Iris
EncodingNeural Network algorithm
Decision based algorithm
Iris recognition based on
Zernike MomentsThree level layered
architecture
True
Positives %
LMD (Local Mean
Decomposition)
Multi-Model Fusion
Fuzzy Integration
Simple Mean calculations for
feature extraction
Ridgelts
Gabor Filter Bank
Daug a ’s Rubber Sheet Model
Geometric key based Iris
Encoding
Neural Network algorithm
Decision based algorithm
Iris recognition based on
Zernike Moments
Three level layered
architecture
Iris Recognition February, 2015
Iris Recognition February 2015
approach, it can result with good efficiency and accuracy.
ROC Curves analysis
x = false positive rates (1-specificity)
y = true positive rates (sensitivity)
Fitt
ed
cur
ve:
y =
0.4
Ln(
x)
+
0.9
1
R^
2 =
0.8
985
Are
a
und
er
cur
ve
=
0.5
19
Est
imated ROC curve with
Column 1 = false positive rates (1-specificity)
Column 2 = true positive rates (sensitivity)
LMD
(Local
Mean
Decom
positio
n)
Multi-
Model
Fusion
Fuzzy
Integra
tion
Simple
Mean
calcula
tions
for
featur
e
extract
ion
Ridgelt
s
Gabor
Filter
Bank
Daugm
a ’s Rubber
Sheet
Model
Geome
tric key
based
Iris
Encodi
ng
Neural
Netwo
rk
algorit
hm
Decisio
n
based
algorit
hm
Iris
recogn
ition
based
on
Zernik
e
Mome
nts
Three
level
layere
d
archite
cture
False
Positives %11 10 15 22 32 15 20 13 12 20 14 15
True
Positives %89 90 85 78 68 85 80 87 88 80 86 85
Accuracy % 96.3 86.4 90.78 80.86 88.6 94.5 91.6 94 98.5 89.5 88.7 78.4
Effiency % 78 89.6 86.7 77.8 95.7 90.6 89 77 85.5 78 84.6 89.5
0
20
40
60
80
100
120
Pe
rce
nta
ge
s
Overall Balance
Iris Recognition February, 2015
Iris Recognition February 2015
Figure 15: ROC Curves
0
0.2
0.4
0.6
0.8
1
1.2
0 0.2 0.4 0.6 0.8 1 1.2
Straight Line
Wei-Yu Han
Sunpreet S. Arora, Mayank Vatsa,
Richa Singh
Patricia Melin, Victor Herrera
Dolly Choudhary
Lenina Birgale and Manesh Kokare
R.G.P.V. Bhopal, M.P
Mr. P.P.Chitte, Prof. J.G.Rana
Chun-Wei Tan, Ajay Kumar
Sunpreet S. Arora
Andreas Uhl and Peter Wild
Chun-Wei Tan, Ajay Kumar 2
Gajendra Singh Chandel, Ankesh
Bhargava2
Iris Recognition February, 2015
Iris Recognition February 2015
6. REFERENCES
1. Sunpreet S. Arora, MayankVatsa, Richa
Singh IIIT Delhi New Delhi, India Iris
Recognition (National Science And
Technology Council)
2. Genetic Optimization of Neural Networks
for Person Recognition Based on the Iris
3. Gajendra Singh Chandel1, Ankesh
Bhargava2 Sri Satya Sai institute of
science and technology, R.G.P.V. Bhopal,
M.P.
4. Iris Recognition based on Local Mean
Decomposition Wei-Yu Han1, Wei-Kuei
Chen1, ∗, Yen-Po Lee1, Kuang-shyr Wu1
and Jen-Chun Lee2 ChienHsin University,
Taoyuan, Taiwan 2 Department of
Electrical Engineering, Chinese Naval
Academy, Kaohsiung, Taiwan Received:
3 May. 2013, Revised: 28 Aug. 2013,
Accepted: 29 Aug. 2013 Published online:
1 Apr. 2014
5. Noisy iris image matching by using
multiple cues Tieniu Tan ⇑, Xiaobo
Zhang, Zhenan Sun, Hui Zhang National
Laboratory of Pattern Recognition,
Institute of Automation, Chinese
Academy of Sciences, P.O. Box 2728,
Beijing 100190, PR China.
6. Iris Recognition Using
RidgeletsLeninaBirgale* and
ManeshKokare**
7. Analysis of Template Aging in Iris
Biometrics Samuel P. Fenker and Kevin
W. Bowyer Department of Computer
Science and Engineering Univ. of Notre
Dame, Notre Dame IN 46556
8. IEEE Trans. Information Forensics and
Security, 2014 Efficient and Accurate at-
a-distance Iris Recognition Using
Geometric Key based Iris Encoding
Chun-Wei Tan, Ajay Kumar
9. IEEE Trans. Image Processing, 2014
Accurate Iris Recognition at a Distance
Using Stabilized Iris Encoding and
Zernike Moments Phase Features
10. Weighted Adaptive Hough and
Ellipsopolar Transforms for Real-time Iris
Segmentation Andreas Uhl and Peter
Wild∗ Multimedia Signal Processing and
Security Lab Department of Computer
Sciences, University of Salzburg, Austria
11. Multi-stage Visible Wavelength and Near
Infrared Iris Segmentation Framework⋆
Andreas Uhl and Peter Wild
12. Evaluation of the effects of Gabor filter parameters on texture
classificationFrancesco Bianconia, Antonio
Fern´andezbaDipartimentoIngegneriaIndu
striale, Universit`adegliStudi di Perugia
Via G. Duranti 63, 06125 Perugia (Italy)
bDepartamento de Disen˜oenIngenier´ıa, Universidad de Vigo E.T.S.I.I. - Campus
Universitario, 36310 Vigo (Spain)