The document proposes a secure personal identification system based on human retina recognition. The system uses retinal vascular patterns, which are unique to each individual, for identification. It consists of three stages: 1) preprocessing to extract the vascular pattern from retinal images, 2) feature extraction to identify feature points like endings and bifurcations, and represent them as vectors, and 3) matching input images to templates by calculating distances between feature vectors. Experimental results on two retinal image databases achieved over 97% accuracy, demonstrating the potential of the proposed system for high-security identification.

1. A Secure Personal Identification System Based on
Human Retina
Joddat Fatima †, Adeel M. Syed ‡ and M. Usman Akram∗
Department of Computer & Software Engineering
Bahria University, Islamabad, Pakistan.
Email: joddat.fatima@gmail.com†, adeelmuzaffar@gmail.com‡, usmakram@gmail.com∗
Abstract—Biometrics are the personal physiological and be-
havioral characteristics which are mostly used for personal
recognition. Today, biometric based security systems such as
fingerprint, iris and face recognition are used everywhere es-
pecially in high security areas. Human retina, not same as iris,
is another source of biometric system which provides the most
reliable and stable means of authentication. In this paper, we
present a new system based on vascular pattern of human retina
which can be used for security purposes. The proposed algorithm
consists of three stages; i.e. preprocessing, feature extraction and
finally the matching process. In preprocessing, it extracts the
vascular pattern from input retinal image using wavelets and
multilayered thresholding technique. Second stage extracts all
possible feature points and represents each feature point with
a feature vector. The proposed system matches the template
feature vectors and input image feature vector using Mahalanobis
distance. The experimental results demonstrate that the proposed
system performs retina recognition with high accuracy.
I. INTRODUCTION
Biometrics are used in automated personal recognition and
identification systems especially at high security regions [1].
The commonly used biometrics include fingerprint, iris, facial
and speech recognition [1]. Although there are highly accurate
systems based on these biometrics but all of these can easily be
forged. Human retina contains vascular pattern which is unique
in every individual and can be used in biometric system [2].
Unlike traditionally used biometric systems, vascular pattern
of human retina is the most reliable and stable source for
biometrics. It is not easy to forge it as it lies at back end of
human eye and is not directly accessible [2]. There is always
a confusion regarding iris and retina as they both belong to
human eye but their functions and patterns are completely
different. Iris is the colored region between the pupil and sclera
whereas retina is located at back region of eye. The systems
based on iris recognition use iris pattern for uniqueness and
identification whereas the foundation of retinal recognition
is the pattern of blood vessels present in human retina [2].
Figure1 shows the digital images of iris and retina.
Digital retinal images are used in automatic identification
systems. They are acquired using digital fundus camera con-
sisting of low power microscope and is used by eye specialists
for treatment of retinal diseases. The acquisition of human
retina is not as simple as for other biometrics like fingerprints,
face, iris etc and the fundus cameras are also very expensive
[2]. It is the acquisition process which stops the wide and
common use of retinal recognition systems but still they are
Fig. 1. Difference between iris and retina. a) Digital image of iris; b) digital
image of human retina
suited for high security military and Government areas due to
their stability and reliability. Figure2 shows the human retinal
and vascular pattern extracted from digital retinal image.
Fig. 2. Human Retinal a) Digital image of human retina; b) Vascular pattern
The uniqueness and importance of blood vessels and its
pattern in human retina were discovered by two eye specialists,
Dr Carleton Simon and Dr Isodore Goldstein while studying
on eye diseases back in 1935 [3]. They discovered that every
eye owns a unique vascular pattern which can be used for
personal authentication. Advancing in the same field, Dr Paul
Tower studied on the vascular patterns of twins in 1950 and
discovered that they are unique even among identical twins
[4]. In his study, he showed that retinal vascular patterns the
least similar among resemblance factors typical of twins.
A few studies have been done on retinal recognition. Main
focus was on the accurate extraction of blood vessels. A
number of methods have been proposed for blood vessel detec-
tion and segmentation [5]-[11] for computer aided diagnostic
system to detect retinal diseases, but their use for retinal identi-
2013 IEEE Symposium on Industrial Electronics & Applications (ISIEA2013), September 22-25, 2013, Kuching, Malaysia
978-1-4799-1125-7/13/$31.00 ©2013 IEEE 90

2. Fig. 3. Flow diagram of proposed retinal recognition system
fication is very rare in literature. Landmark based methods for
retinal registration were presented in [12]-[13] using vessel
bifurcations and crossover points as feature points. Another
retinal registration method based on location of optic disc was
presented by [14]. A cross correlation coefficient based retinal
identification was done in [15]. They first registered the input
image and then matching is done by correlating the vascular
pattern. Bevilacqua et al. [16] proposed a vascular bifurcation
and cross over point based system for retinal authentication.
The comparison was done by means of accumulation matrix.
In this paper, we present a new method for secure personal
identification based on human retina. The main contribution
which we have made during this research is the generation of
database for retinal recognition which is very rare and only
one database is available online for retinal recognition. The
paper includes other contributions as well such as accurate
segmentation of vascular pattern, reliable extraction of feature
points and formulation of feature vectors for vascular feature
points.
The remaining paper is arranged as follows: Section II
contains the proposed system and explains all three stages
in detail. The experimental results are given in section III
followed by conclusion in last section.
II. PROPOSED SYSTEM
Biometric based security systems are the reliable source of
security for highly sensitive areas. Human retina consists of
blood vessels which form a unique pattern and it is almost
impossible to forge it in a false individual. The proposed
system consists of three stages; i.e. preprocessing, feature
extraction and vascular matching. In preprocessing, it performs
blood vessel enhancement and segmentation to extract the
vascular pattern from digital retinal image. Then, it extracts
and validates the feature points from vascular pattern just like
fingerprint minutiae points and makes a feature vector for each
point. In last stage, the system performs vascular matching us-
ing mahalanobis distances. Figure 3 shows the complete flow
diagram for proposed system. Like any other biometric system,
our system broadly consists of two modules i.e. enrollment
and identification modules. In enrollment modules, it captures
the retinal images from a number of persons and saves their
feature vectors in databases by doing off line processing. The
identification module performs online processing by capturing
the retinal image of a person which is to be identified and
system matches his/her feature vector with the ones which are
already saved in the database to extract the true identity of
person.
A. Preprocessing
The basis of retinal recognition is the vascular pattern
which shows uniqueness among different human beings. It is
important to extract the complete vascular structure accurately
form input retinal image. The first phase of our proposed
system is preprocessing in which system extracts the vascular
pattern using Gabor wavelet and multilayered thresholding
technique. Preprocessing is also performed to remove noise
and extra region from input fundus image [17].
1) Blood Vessel Enhancement: In proposed system, we
have used Gabor wavelet with a fixed scale value of 4 which
is selected empirically and a change of 100
in orientation.
2013 IEEE Symposium on Industrial Electronics & Applications (ISIEA2013), September 22-25, 2013, Kuching, Malaysia
91

3. Fig. 4. Flow diagram for proposed multilayered thresholding technique for blood vessel segmentation
This leads to 18 different wavelet responses out of which we
selected maximum wavelet response for each pixel [18].
2) Blood vessel segmentation: The wavelet transformation
enhances the blood vessels by giving high responses for vas-
cular areas but still thick vessels have high wavelet response
as compare to thin vessels. So it is hard to find one optimal
threshold value for accurate vascular extraction without any
supervised algorithm. This is of great importance especially
in case of PDR as new abnormal blood vessels are normally
very thin. We presented a recursive supervised multilayered
thresholding-based method for accurate vascular segmentation
[18]. Figure 4 shows the flow diagram of proposed multilay-
ered thresholding technique for vascular pattern segmentation.
The extracted vascular pattern consists of blood vessels
of variable thickness. In order to make the width of blood
vessels equal to single pixel, we apply morphological thinning
operation [19]. Figure 5 shows the step by step output for
preprocessing stage.
Fig. 5. Preprocessing: a)Green Channel retinal image; b)Enhanced blood
vessels; c)Segmented blood vessels; d)Thinned blood vessels
B. Feature Extraction
The preprocessing phase extracts the thinned vascular struc-
ture from input retinal image. Next phase is feature extraction.
1) Feature Points Extraction: The main features of vascular
pattern are vessel ending and bifurcation points (fig. 6) just like
fingerprint minutiae points [1].
Fig. 6. Structures of vascular features; i.e. vessel ending and bifurcations
The system uses crossing number method (eq. 5) to extract
the vascular endings and bifurcations [20]. E(p) = 3 and
E(p) = 1 correspond to vessel bifurcations and endings
respectively.
2013 IEEE Symposium on Industrial Electronics & Applications (ISIEA2013), September 22-25, 2013, Kuching, Malaysia
92

4. 2) Feature Points Validation: The feature points obtained
in feature extraction phase need to be validated before the
matching process. In order to eliminate false endings and bifur-
cations due to spurs and small breaks, we apply a windowing
technique. In this, we take a window of size 9 × 9 with all
initial values equal to 0 and place the candidate feature point;
i.e. ending or bifurcation, at the center of window. Initialize the
center point with a value of −1. Vessel endings have only one
connected branch, assign 1 to all connected pixels and count
0 to 1 transitions while moving in clockwise direction along
the boundary of window. The count should be equal to 1 for a
valid vessel ending. We repeat the same procedure for vessel
bifurcations and count transitions. Just like vessel ending, if
the count is equal to 3 then it will be a valid bifurcation. Figure
7 shows the true vessel endings and bifurcations and also the
validation processes for both.
Fig. 7. Feature Validation: a) Vessel ending; b) Validation of vessel endings;
c) Vessel Bifurcation; d) Validation of vessel bifurcation
3) Feature Set Formation: Once the feature points are
extracted and validated, the system forms a feature vector
for each feature point. The system forms a feature vector
for each candidate feature point by calculating its distance
and relative angles with its four nearest feature points. The
proposed system represents each feature point with a feature
set of form < φ11, φ12, φ13, φ14, d11, d12, d13, d14 > where
φxy and dxy are the relative angle and distance between two
feature points x and y respectively. The relative angles are used
to have rotation invariant matching process. Figure 8 shows the
formation of feature vector for a feature point.
C. Feature Matching for Retinal Identification
Feature vectors for different retinal images are saved in
database. The last stage identifies the input retinal image
by matching the feature vectors of query image with the
ones stored in database. The proposed system calculates a
Fig. 8. Four nearest feature points and their relative orientations for center
feature point
Feature Distance (FD) by calculating Mahalanobis distance
[21] between feature vectors of all feature points in query
image and templates stored in database. The matching phase
finally computes a score value which represents that how
similar two retinal image are when compared with each other.
This value is computed by counting the number of feature
vectors from database matched with the query image feature
vectors on the basis of FD.
III. EXPERIMENTAL RESULTS
The person recognition system based on biometrics require
thorough testing and validation. There are only a few publicly
available databases for retinal recognition as compared to
fingerprint or facial recognition. VARIA [23] is the only
database that is truly formed for retinal recognition systems. It
includes 233 retinal images of 139 different persons with a res-
olution of 768 × 584. To further evaluate the proposed retina
identification system, we designed our own database with
the help of armed forces institute of ophthalmology (AFIO),
Pakistan and we named it as Retinal Identification DataBase
(RIDB). RIDB is collected for 20 different individuals with
5 images per person and overall it contains 100 images of
resolution 1504 × 1000. In order to check the validity of
person identification, the system is tested on both retinal image
databases and results are shown in Table-II.
TABLE I
RECOGNITION RATE OF PROPOSED METHOD
Database Total Correctly Wrongly Recognition
images recognized recognized rate (%)
VARIA 233 232 1 99.57
RIDB 100 97 3 97
In order to further check the validity of proposed system, we
2013 IEEE Symposium on Industrial Electronics & Applications (ISIEA2013), September 22-25, 2013, Kuching, Malaysia
93

5. used false acceptance rate (FAR), false rejection rate (FRR),
equal error rate (ERR) and recognition rate as performance
merits and they are defined as follows
• False Acceptance Rate (FAR) is the rate of individuals
which are incorrectly accepted. It measures the number of
cases in which input vector is matched to a non matching
vector present in the database.
• False Rejection Rate (FRR) is the rate of individuals
which are incorrectly rejected. It measures the number
of cases in which input vector is not matched to actual
vector present in the database.
• Receiver Operating Characteristic (ROC) is a curve
which shows the relation between FAR and FRR and us-
ing ROC we can trade off between these two parameters.
ROC curves are used to compute the area under the curve
(AUC).
• Equal Error Rate (EER) is the rate at which both FAR
and FRR have same values. The value of the EER can
be easily obtained from the ROC curve. In general, the
device with the lowest EER is most accurate.
• Recognition Rate (RR) is computed by running the system
on test data and it is the number of persons which are
correctly classified.
In biometrics, there is always a trade off between FAR and
FRR and their error rate depends on the value of threshold. If
we lower the threshold, it will reduce the FRR but will increase
the FAR until we reach a point (zero FRR) where FRR has
a value of 0. In the same way, zero FAR is a point where
FAR has a zero value and it can be achieved by increasing the
threshold value. We computed FAR and FRR values against
different values of threshold and computed the values for zero
FRR, EER and zero FAR. Table II summarizes the FAR and
FRR values against different threshold values. Figure 9 shows
TABLE II
FAR AND FRR VALUES VS THRESHOLD
Threshold FAR FRR
0.24 0.41 0
0.30 0.33 0.01
0.50 0.09 0.03
0.54 0.05 0.05
0.60 0.02 0.09
0.65 0 0.10
the FAR and FRR curves for proposed system. The values of
FAR, FRR and EER are highlighted with different threshold
values. It shows that at a threshold value of 0.542, we have
EER for FAR and FRR which is 0.0557 for proposed system.
IV. CONCLUSION
Automated person identification is very important to im-
prove the security level especially in highly sensitive areas.
Human retina provides a reliable source for biometric based
system which is almost impossible to forge. In this paper,
we presented an automated system for person identification
based on human retina. We proposed a three stage algorithm
Fig. 9. FAR vs FRR curves for proposed system with different threshold
values
in this paper consisting of preprocessing, feature extraction
and vascular matching. The first step enhanced and segmented
blood vessel using wavelets and multilayered thresholding. In
order to use vascular pattern for matching, the proposed system
used vessel endings and bifurcations as feature points and
formed translation and rotation invariant feature vectors based
on relative angles and distances. Matching is performed using
Mahalanobis distance and results demonstrated the validity of
proposed system.
REFERENCES
[1] S. Prabhakar, S. Pankanti, and A. K. Jain, “Biometric recognition:
Security and privacy concerns”, IEEE Security and Privacy, vol. 1, No.
2, pp. 33-42, 2003.
[2] Ravi Das, “Retinal recognition Biometric technology in practice”,
Keesing Journal of Documents & Identity, vol. 22, pp. 11-14, 2007.
[3] Simon, C., Goldstein, I., “A New Scientific Method of Identification”,
New York State Journal of Medicine, vol. 35, No. 18, pp. 901-906, 1935.
[4] Tower, P., “The Fundus Oculi in Monozygotic Twins: Report of Six Pairs
of Identical Twins”, Archives of Ophthalmology 54, pp. 225-239, 1955.
[5] J. Staal, M. D. Abramoff, M. Niemeijer, M. A. Viergever and B. van
Ginneken, “Ridge-based vessel segmentation in color images of the
retina”, IEEE Trans. Med. Imag., vol. 23, No. 4, pp. 501-509, 2004.
[6] J. V. B. Soares, J. J. G. Leandro, R. M. Cesar, H. F. Jelinek and M.
J. Cree., “Retinal vessel segmentation using the 2-D gabor wavelet and
supervised classification”, IEEE Trans. on Med. Imag, vol. 2, No. 9,
1214-1222, 2006
[7] G. G. Yen andW.-F. Leong: “A sorting system for hierarchical grading of
diabetic fundus images, A preliminary study”. IEEE Trans. Inf. Technol.
Biomed., vol. 12(1), 118-130 ,2008.
[8] S. Chaudhuri, S. Chatterjee, N. Katz, M. Nelson, and M. Goldbaum:
“Detection of blood vessels in retinal images using two-dimensional
matched filters”. IEEE Trans. Med. Imag. 8(3), 263-269,1989.
[9] M. U. Akram, S. Khalid, S. A Khan, “Identification and Classification
of Microaneurysms for Early Detection of Diabetic Retinopathy”, Pattern
Recognition (Elsevier), Vol 46, No.1, 107-116, 2013
[10] D. Marn, A. Aquino, M. Emilio G. A., and J. M. Bravo, “A New
Supervised Method for Blood Vessel Segmentation in Retinal Images
by Using Gray-Level and Moment Invariants-Based Features”, IEEE
Transactions on Medical Imaging, vol. 30, no. 1, 2011.
[11] A. Hoover, V. Kouznetsova, and M. Goldbaum, “Locating blood vessels
in retinal images by piecewise threshold probing of a matched filter
response”. IEEE Trans. Med. Imag. 19(3), 203-210,2000.
[12] M. Ortega, M.G. Penedo, J. Rouco, N. Barreira, and M.J. Carreira,
“Retinal verification using a feature points-based biometric pattern”,
EURASIP J. Adv. Signal Process, pp. 1-13, 2009.
2013 IEEE Symposium on Industrial Electronics & Applications (ISIEA2013), September 22-25, 2013, Kuching, Malaysia
94

6. [13] Z.W. Xu, X.X. Guo, X.Y. Hu, X. Chen, and Z.X. Wang, “The identifi-
cation and recognition based on point for blood vessel of ocular fundus”,
In Proc. ICB 2006 LNCS 3832, pp. 770-776, 2006.
[14] H. Farzin, H. Abrishami-Moghaddam, and M.-S. Moin, “A novel retinal
identification system”, EURASIP J. Adv. Signal Process, pp. 1-10, 2008.
[15] K. Fukuta, T. Nakagawa, Y. Hayashi, Y. Hatanaka, T. Hara, H. Fujita,
“Personal Identification Based on Blood Vessels of Retinal Fundus
Images”, Proc. of SPIE 6914, 2008.
[16] V. Bevilacqua, L. Cariello, D. Columbo, D. Daleno, M. D. Fabiano,
M. Giannini, G. Mastronardi, M. Castellano, “Retinal Fundus Biometric
Analysis for Personal Identifications”, ICIC LNAI 5227, pp. 1229-1237,
2008.
[17] A. Tariq, M. U. Akram, “An Automated System for Colored Retinal
Image Background and Noise Segmentation”, IEEE Symposium on
Industrial Electronics and Applications (ISIEA 2010), pp. 405-409, 2010.
[18] M. U. Akram and S. A. Khan, “Multilayered Thresholding Based Blood
Vessel Segmentation for Screening of Diabetic Retinopathy”, Engineering
with Computers (EWCO), Vol. 29, No. 2, pp. 165-173, 2013.
[19] R. C. Gonzalez and R. E. Woods, “Digital image processing”, Prentice
hall, second edition, 2002.
[20] A. K. Jain, R. Bolle, and S. Pankanti, “Biometrics: Personal Identifica-
tion in a Networked Society”, Springer- Verlag New York Inc (C), 1999,
[21] R. O. Duda, P. E. Hart, and D. G. Stork, “Pattern Classification”, New
York, Wiley, 2001.
[22] A. Hoover, M. Goldbaum, “Locating the optic nerve in a retinal image
using the fuzzy convergence of the blood vessels”, IEEE Trans. on
Medical Imaging, vol 22, No. 8, pp. 951-958, 2003.
[23] VARIA,“Varpa retinal images for authentication”,
http://www.varpa.es/varia.html.
2013 IEEE Symposium on Industrial Electronics & Applications (ISIEA2013), September 22-25, 2013, Kuching, Malaysia
95