Automatic Feature Extraction from Ocular Images

Institute of Telecommunications
University of Technology & Life Sciences
85-796 Bydgoszcz, Poland
Abstract—Ocular images processing is an important task in: i) biometrics system based on retina and/or sclera images, and
ii) in clinical ophthalmology diagnosis of diseases like various vascular disorders. We presents a general framework for
image processing of ocular images with a particular view on feature extraction. The method uses the set of geometrical and
texture features and based on the information of the complex vessel structure of the retina and sclera. The feature extraction
contains the image preprocessing, locating and segmentation of the region of interest (ROI). The image processing of ROI
and the feature extraction are proceeded, and then the feature vector is determined for the human recognition and
ophthalmology diagnosis.
Keywords-Retina image, Conjunctiva image, Feature extraction, Gabor transform, Texture features
1. Introduction
The general eye anatomy is presented in Fig.1. In this
paper an approach is presented in which the vessel system of
the retina and conjunctiva is automatically detected and
classified for human identification/verification and
ophthalmology diagnosis.
Fig.1. The eye anatomy.
The retina is a thin layer of cells at the back of the
eyeball of vertebrates. It is the part of the eye which converts
light into nervous signals. It is lined with special
photoreceptors which translate light into signals to the brain.
The main features of a fundus retinal image were defined as
the optic disc, fovea, and blood vessels. Every eye has its
own totally unique pattern of blood vessels. The unique
structure of the blood vessels in the retina has been used for
biometric identification and ophthalmology diagnosis.
The conjunctiva is a thin, clear, highly vascular and
moist tissue that covers the outer surface of the eye (sclera).
Conjunctival vessels can be observed on the visible part of
the sclera.
A biometric system is a pattern recognition system that
recognizes a person on the basis of a feature vector derived
from a specific physiological or behavioral characteristic that
the person possesses. The problem of resolving the identity
of a person can be categorized into two fundamentally
distinct types of problems with different inherent
complexities: (i) verification and (ii) identification.
Verification (also called authentication) refers to the problem
of confirming or denying a person's claimed identity (Am I
who I claim to be?). Identification (Who am I?) refers to the
problem of establishing a subjects identity.
We propose a new modality for eye-based personal
identification that uses the blood vessel in ocular images
(retina, conjunctiva). The acquisition process requires
collaboration from the user and it is sometimes perceived as
intrusive (Fig.2).
Fig.2. Typical Eye Vessel Acquisition and Biometrics
System.
In diagnostic ophthalmology the main problem is the
detection of feature points of vessel structure (bifurcations
and crossovers) and its misclassification.
The paper is arranged as follows: in Section 2 a
description of the preprocessing (color transformation, edge
detection, etc.) method is presented. Section 3 describes the
Open Journal of Applied Sciences
Supplement：2012 world Congress on Engineering and Technology
34 Cop
y
ri
g
ht © 2012 SciRes.

extraction of geometrical features method used. Section 4
describes the extraction of texture features method used.
Section 5 shows the results and Section 6 provides some
conclusions.
Images which are considered in this paper as Retina-1
and Conjunctiva-1, are displayed in Fig.4.
Fig.3. The flowchart of the proposed stage for biometrics
and clinical diagnosis.
a) b)
Fig.4. Retina-1 (a) and Conjunctiva-1 (b) images.
I. PREPROCESSING
Before performing feature extraction, the original eye
images are subjected to some image processing operations,
as:
2. Color transformation.
Most digital images are stored in RGB color space. RGB
color space is represented with red (R), green (G), and blue
(B) primaries and is an additive system. RGB color space is
not perceptually uniform, which implies that two colors
with larger distance can be perceptually more similar than
another two colors with smaller distance, or simply put, the
color distance in RGB space does not represent perceptual
color distance.
Fundus images contain full color information. The first
step is to separate RGB channels and from the three color
channels the green (G) component of RGB color space for
blood vessel recognition is chosen (Fig. 5). Additional to
represent retinal characteristic we are using luminance
component (Y) from color space (Fig. 6).
To represent eye characteristic we using luminance
component (Y) from color space.
a) b) c) d)
Fig.5. Original retina image in RGB color space (a) and Red
(b), Green (c) Blue (d) channels.
a) b) c)
Fig.6. Retina image in color space: a) Y component and
components (b), (c) respectively
(1)
1. Image stretched.
The contrast level is stretched according to
(2)
is the color level for the output pixel (x, y) after the contrast
stretching process. is the color level input for data the pixel
(x, y). max - is the maximum value for color level in the
input image. min - is the minimum value for color level in
- constant that defines the shape of the
stretching curve.
2.Detection of the optic disc.
The optic disk is characterized by grey values that are
brighter than the background values. The variance of
intensity of adjacent pixels was used for detection and
recognition of the optic disc. First variance image is form
and the location of the maximum of this image is the centre
of the optic disc (Fig.7).
3.Edge detection.
Cop
y
ri
g
ht © 2012 SciRes.35

To obtain the vessel binary image several alternatives
method can be used from morphological to multi-resolution
analysis methods. We use the typical edge detection Canny
algorithm with local threshold. The results of the vessel
edge detection are shown in Fig.8.
a) b)
Fig.7. The retina image after image preprocessing (a) and
Optic disk and macula (b).
a) b)
Fig.8. Vessel edge detection of Retina-1 (a) and
Conjunctiva-1 (b) images.
3. Extraction of Geometrical Features
For each vessels line we specify vessel bifurcations
characteristic points and cross points of vessel intersections
characteristic points, information derived from connected
number of point pis used (Fig. 9).
When p=1, the connected number of p is defined by the
next equation
(3)
(4)
where: S=(1,3,5,7) and means (1- p ).
Topological properties of the pixel p are shown in Table
1.
TABLE I
TOPOLOGICAL PROPERTIES OF p
THE VALUE OF OR 3 4
PROPERTY OF PIXEL p Branch Cross
The feature vector corresponding to vessel topology and
consecutively the number of bifurcations and cross points
are stored in the feature vector. Moreover, the coordinates
of all the extracted characteristic points are stored. The
feature vector for each vessels consists of the following
parts: - 2 numbers corresponding to the number of
bifurcation points and cross points in each vessels, -
subvector in which the coordinates of the bifurcation points
are stored, - subvector in which the coordinates of the
cross points are stored.
a) b)
Fig.9. Geometrical features of Retina-1 (a) and Conjunctiva-
1 (b) images.
The correspondence between the vessel in an image and
the vessel templates is based on the similarity between their
characteristic points. The characteristic points are
computed for each vessel template. The characteristic
points of the vessel image are then compared with the
characteristic points of each vessel template.
Using the correspondence between the vessel
characteristic points and vessel template characteristic
points, we can calculate the total number of matching points
and obtain the matching results. The process is illustrated in
Fig. 10.
Fig.10. The correspondence between the vessel
characteristic points in retina image
4. Texture Feature from Gabor
Wavelet Transform
Gabor wavelet is a powerful tool to extract texture
features. Gabor functions are Gaussians modulated by
complex sinusoids. In two dimensions they take the form
(Fig. 11):
36 Cop
y
ri
g
ht © 2012 SciRes.

(5)
where (x, y) is the pixel position in the spatial domain, is
the scaling parameters of the filter , is the radial center
frequency, , and specifies the orientation of the Gabor
filters.
The second term of the Gabor filter, , compensates for
the DC value because the cosine component has nonzero
mean while the sine component has zero mean.
Fig.11. Real (a) and imaginery (b) parts of Gabor wavelets
and Gabor kernels with different orientations (c).
Gabor filtered output of the image is obtained by the
convolution of the image with Gabor function for each of
the orientation/spatial frequency (scale) orientation.
Given an image
(6)
The normalized retina or conjunctiva images are divided
into blocks. The size of each block in our application is .
Each block is filtered with Eq. (6).
A set of parameters of the Gabor filters is used as and .
In this case we have 12 filters. But the Gabor feature vector
with all the 12 filters becomes very redundant and
correlative and the dimension of Gabor feature vector is
large. The dimension of the Gabor feature vector with 12
filters will result in where is the size of each ROI block
and t is number of blocks.
To reduce dimension of feature vector, we select few
Gabor filters (Fig. 12) without degrading the recognition
performance. In this case we have only 6 filters. Thus we
have pattern vector with elements for each blocks.
In our application we used only 3 global texture features
for each t blocks:
(7)
(8)
(9)
where k, l is block image dimension.
The feature vector is constructed using and as
feature components.
We defined the vectors of features as follows:
(10)
The first part of the contains the number of bifurcation
and crossing points with corresponding coordinates for each
tblocks. The features in second part of are listed as
follows
Fig.12. Selected Gabor filter bank.
5. Conclusion
A new method for recognition retina vessel and
conjunctiva vessel images was presented. This method
based on geometrical, and Gabor features. This paper
analyses the details of the proposed method. Retina vessel
and conjunctiva vessel images can be used for personal
identification. Experimental results have demonstrated that
this approach is promising to improve retina recognition for
person identification. Furthermore conjunctiva vessel
images proposed method is suitable to improve eye
diagnosis.
REFERENCES
[1] K.G. Goh, M.L. Lee, W. Hsu, H. Wang, ADRIS: An
Automatic Diabetic Retinal Image Screening System,
Medical Data Mining and Knowledge Discovery,
Springer-Verlag, 2000.
[2] W. Hsu, P.M.D.S. Pallawala, M.L. Lee, A.E. Kah-Guan,
The Role of Domain Knowledge in the Detection of
Retinal Hard Exudates, IEEE Computer Vision and
Pattern Recognition, Hawaii, Dec 2001.
[3] H. Li, O. Chutatape, Automated feature extraction in
color retinal images by a model based approach, IEEE
Trans. Biomed. Eng., 2004, vol. 51, pp. 246 -- 254.
[4] N.M. Salem, A.K. Nandi, Novel and adaptive
contribution of the red channel in pre-processing of
colour fundus images, Journal of the Franklin Institute,
2007, p. 243--256.
[5] C. Kirbas, K. Quek, Vessel extraction techniques and
algorithm: a survey, Proceedings of the 3rd IEEE
Symposium on BioInfomratics and Bioengineering
(BIBE' 03), 2003.
[6] S. Chang, D. Shim, Sub-pixel Retinal Vessel Tracking
and Measurement Using Modified Canny Edge
Detection Method, Journal of Imaging Science and
Technology, March-April 2008 issue.
[7] T. Chanwimaluang, G. Fan, An efficient algorithm for
extraction of anatomical structures in retinal images,
Proc. IEEE International Conference on Image
Processing, pp. 1093--1096, (Barcelona, Spain.
[8] H. Farzin, H. Abrishami-Moghaddam, M.S. Moin, A
novel retinal identification system, EURASIP Journal
on Advances in Signal Processing, vol. 2008, Article ID
280635, 2008.
Cop
y
ri
g
ht © 2012 SciRes.37

[9] R.S. Choras, Image Feature Extraction Techniques and
Their Applications for CBIR and Biometrics Systems,
International Journal of Biology and Biomedical
Engineering, Issue 1, vol.1, pp.6--16, 2007.
[10] R.S. Choras, Iris Recognition - M. Kurzynski,
M.Wozniak (Eds.): Computer Recognition Systems 3,
AISC 57, pp. 637--644, Springer-Verlag Berlin
Heidelberg 2
[11] R.S. Choras, Iris-based person identification using
Gabor wavelets and moments, Proceedings 2009
International Conference on Biometrics and Kansei
Engineering ICBAKE 2009, pp. 55--59, CPS IEEE
Computer Society, Los Alamitos, California,
Washington, Tokyo
[12] D. Gabor, "Theory of communication," J. Inst. Elect.
Eng., 93, pp. 429--459, 1946
38 Cop
y
ri
g
ht © 2012 SciRes.