Histogram of Oriented Gradient based Automatic

Detection of Eye Diseases

Dandi Setiawan
dandisetiawan001@gmail.com
Informatics Engineer, STMIK Tasikmalaya

Abstract—Eye diseases are burning diseases nowadays. Eye

illness identification is one of the basic issues in computer vision.
Three such diseases have been considered in this paper, cataract,
conjunctivitis, and stye. A cataract occurs once protein builds up
within the lens of an eye and makes it cloudy which prompts the
decrease in vision. Conjunctivitis or pink eye is a condition where
the conjunctiva of the eye inflamed by an infection or by an
allergic reaction. And the third sort of illness i.e. stye is an
infection that causes a young red, painful lump near the edges of
the eyelid. Generally, an eye doctor uses a slit lamp camera to Fig. 1. Anatomy of eye
find these sicknesses. But because of lack of specialist eye doctor
and slit lamp camera in rural areas are the main problem of the the sort of eye disease, the patient is affected by.Normally
belated in detecting those diseases. In this paper first, the these diseases are detected by eye specialist doctors and due to
captured eye images collected from different patients and
processed for improvement. Then HOG used for detection of the
the shortage of specialist doctors in the remote areas find it
feature vector. Finally recognition of disease done with the difficult to diagnose. The proposed algorithm is efficient as it
assistance of minimum distance classifier. This planned method is has a low response time, high accuracy and is easy to
economical, computationally quick and price very low. The implement.
proposed system result with average accuracy is 96.5 percent in
classification. The main contribution of this paper is the detection of
cataract, conjunctivitis, andstye infected eye from the normal
Keywords—cataract; conjunctivitis; stye; HOG; minimum
eye. For extraction of the feature, we use Histogram of
distance classifier; MATLAB.
Oriented Gradient (HOG) and classification of eye disease
from the normal eye and it is accomplished using minimum
I. INTRODUCTION
distance classifier. The proposed algorithm is used to make it
easier for people to classify the different eye infections at
Human beings can be affected with an assortment of eye
home using just an image of an eye.
sicknesses and as such, there is a need of finding those
sicknesses especially in the case of dedicates organ like eyes. The paper compresses in taking after way. In an area I the
The basic parts of an eye which are used to find out eye introduction related with our analysis work is given. In
sicknesses are iris, pupil, sclera,and eyelids. Fig. 1 shows the segment II related work is given. In segment III depicts the
anatomy ofthe eye. Cataract, conjunctivitis, and stye are some complete framework. In segment IV the exploratory outcomes
such eye sicknesses, if these sicknesses are not seen over a found from the operation utilizing MATLAB image
large period of time in later stages, will result in sightlessness. processing toolbox are clarified. At long last in area V finish
The symptoms associated with cataract embrace clouded and of our expected research.
blurred vision, an issue with vision in the dark and sensitivity
with light and glare. Conjunctivitis has symptoms as well as
II. RELATED WORK
pink or red colorize in the exposed white space of the eye
(sclera) tearing and swelling of redness. The symptoms AmitBhadra et al. [1] presented aprogrammed algorithm for
associated with stye embrace a lump on the lid that’s like a cataract and conjunctivitis eye disease detection. They extract
boil or a pimple, lid pain, lid swelling, and tearing occur. The features of an optical eye image such as big ring space, small
signs and symptoms discovered will be distinguished based on ring space of the lens, the eye ellipse and the intensity of an
affected space for classification. For pre-processing of an
image, they convert an RGB image to gray
and median filter is employed for removal of noise and Amol B. Jagadale and D. V. Jadhav [10] introduced early
smoothing of an image. So as to detection of an inner and recognition and classification of cataract utilizing slit lamp
outer circle, they used Hough circle transform in open CV images by Hough circular transform. The slit light
library for cataract detection also calculate the average photography is the fundamental methods for atomic cataract
grayscale value for conjunctivitis detection and establish the imaging and in view of reflected light. The camera and slit are
read mean value. set sideways to the visual axis of the eye to catch a picture of
the cross -area of the eye. After that ophthalmologist assessed
Meimei Yang et al. presented a characterization of a retinal the level of cataract by contrasting the photo seen through the
picture for programmed cataract detection[2]. In pre- camera with a progression of standard photos, which is named
processing part they used an improved Top-Bottom hat as clinical reviewing, for example, Lens Opacities
transformation and a trilateral filter is employed to decrease Classification System III (LOCSIII).
the noise of an image, they extracted the light and texture
message of the image. For classification, they choose the back V Harini and V Bhanumathi [11] found a programmed,
prorogation (BP) neural network. advantageous and practical cataract characterization
framework. The fundamental goal is to build up a fundus
For automated detection of Ad-Cs sickness [3], picture examination for a wide range of cataract. At first
MelinGunay et al. planned a system setup. They measure the preprocessing is accomplished for the fundus picture, trailed
vascularization and intensity of redness in pink eyes once by include extraction in spatial and transform domain,
segmenting the region of infection in corneal images. They arrangement, and reviewing. The fundus picture is delegated
presented isolation of sclera region using an automated grab- as non-cataract and cataract picture utilizing SVM classifier.
cut methodology that known the seed region from the image The cataract picture is reviewed as mild and serious cataract
itself. utilizing RBF Network.
Eye sickness detection supported covariance,
MdAlamgirHossain [4] setup a completely unique approach. III. METHODOLOGY
Using the covariance approach and scoring technique they
extract intensity pattern to anterior sickness and it is then This section describes the procedure and approach that has
possible to analyze different eye disease and also realize the been adopted to detect cataract, conjunctivitis and stye in an
phase of the sickness. eye using image processing based methods. Fig.2 shows the
flowchart explains the steps that have been implemented.
JoydeepTamuli et al. [5] presented an image processing based
methodology to spot and grade conjunctivitis infected eye
according to its sorts and followed by Principal Component
Analysis (PCA) for extraction of discriminatory feature and Start
classified using supervised learning methodology as multi-
class SVM and KNN. The intensity of infected eye is
calculated using a significant red plane. Plot confusion matrix Read Image of eye
was used to calculate the accuracy of these classifiers.
JyotiPatil and AnantChaudhari [6] arranged one algorithmic
program for intensity observation of conjunctivitis utilizing Resize Image to 240*320 & convert
DIP. They utilized worldwide thresholding strategy for a RGB image to Gray-Scale
division of picture utilizing least and most extreme estimations
of pixels. They watch increment in force of redness of eye
with increment in days of contamination.
Image Enhancement using Median
JagdishNayak [7] proposed a calculation for programmed Filter
grouping of normal, cataract and post cataract disorder. They
removed the elements of an optical eye picture, for example,
Big Ring Area (BRA), Small Ring Area (SRA), Edge Pixel HOG implementation and Feature
Count (SPC) utilizing canny technique, object perimeter. They Selection
accomplished 94% sensitivity and 93.75% specificity.
For automatic cataract detection,Yunendah Nr Fuadah et
al. [8] developed a mobile android application. The interfaces Classification using Minimum
of MCataract application consists of main activity and startan Distance
activity that has a function, taking an image or loading the
image from the gallery, cropping to get the pupil space as
ROI, showing the pupil area obtained, and designation the Classification as
condition of an eye image. The aim of this setup is to analyze Eye with Cataract /
the performance of statistical texture features (uniformity, Eye with Conjunctivitis /
contrast, dissimilarity, correlation, homogeneity). Eye with Stye /
ManpreetKaur et al. [9] proposed a low-cost smartphone based Normal Eye
system integration with microscopic lenses that permits
patients in remote areas to own regular eye check-up and
determine retinal sickness. They connected a mobile to the Stop
computer and created a network that runs a Java framework on
a computer.
Fig. 2. Flowchart of methodology
 cataract, conjunctivitis, stye and normal eye. The
The algorithm works as follows:- algorithm is verified by testing the images from test
1. The images of eyes with three different types of dataset.
diseases (cataract, conjunctivitis, stye) and normal
eye have been collected of different patients from
Maharashtra Institute of Medical Education &
Research (MIMER Talegao).
2. After collecting the images of different patients, the
images were read and as the input image dimension
can be of any size proper adjustment is needed to
resize the image if the dimension is out of range.
Resize the image to 240* 320. After that convert
RGB image to grayscale image that eliminate the hue
and saturation information while retaining the
luminance.
3. After resizing the image, median filter is utilized to
remove the noise in smooth patches or smooth areas
of a signal, but negatively affect edges, and making it
easier to identify key features.
4. For feature extraction we use Histogram of Oriented
Gradient (HOG) feature descriptor [12] which
includes following steps.
• Gradient (incline/smooth change of something
between two points)computation: The first step of
calculation is to make sure the normalized colour and
gamma values. And then calculate the gradient
values. The most common method is to apply 1-D Fig. 3. Sample images for cataract, conjunctivitis, stye and normal eye used.
centred, point separate derivative mask in left-and-
right and up-and-down direction. This method In fig. 3 some sample images used in the proposed work have
demands filtering the colour or intensity data of the been shown. They are cataract, conjunctivitis, stye and normal
image with the following kernels. types (from left to right).
[-1, 1, 0] and [-1, 0, 1]T.
• Spatial and orientation binning: IV. EXPERIMENTAL RESULT
The second step of calculation is creating the cell
histograms. Each pixel within the cell casts a We have tested total 120 images containing 30 cataract
weighted vote for an orientation based on the values images, 30 conjunctivitis images, 30 stye images and 30
found in the gradient computation. These cells are in normal eye images. Minimum distance classifier indicates the
rectangular or radial in shape and histogram channels closest distance between the test data and train data that was
spread over 0-180 degree or 0-360 degree depending used in classification to determine cataract, conjunctivitis and
stye disease.Fig. 4 shows that how cataract disease is detected
on gradient. From the vote weight, pixel contribution
using given sample image and also it gives the histogram of
is the gradient magnitude or square root of gradient sample image for further processing. Similarly Fig. 6, Fig. 8
magnitude. and Fig.10 show the procedure for conjunctivitis, stye disease
• Normalization and descriptor blocks : and normal eye detection.
For changes in lighting up and contrast, the
incline/smooth change of something between two
points strengths must be locally normalized, which
requires grouping the cells together into larger,
spatially connected blocks.These blocks usually
overlap, meaning that each cell gives more than once
to the final descriptor. These blocks in the form of
four 8x8 pixels cells per block (16x16 pixels per
block) with 9 histogram channels. Let Ȟbe the non-
normalized vector containing all histograms in a
given block.‫ ۅۅ‬Ȟ ‫ۅۅ‬kbe its k-norm for k=1 and ɟ be
some small constant.Then the normalization factor is
given in following formula.
L1-norm: ƒ=
‖‖
భା
‖‖ е

5. Now minimum distance classifier was used to

analyse the data and recognize the disease.This
classifier utilized to arrange unknown picture
information to classes which limit the separation Fig. 4. A case of eye image with cataract
between the picture information and the class in
multi-feature space. A training set was constructed to The result of cataract detection in Graphical Use Interface
train the algorithm and four classes were made for (GUI) shown in Fig. 5.There are five steps to recognize the
eye disease.Firstly add the image from collected data for pre-
processing and image enhancement is done using median
filter. HOG is used to calculate the X & Y gradients. Finally
for recognition we use minimum distance classifier.

Fig. 8. A case of eye image with stye

Fig. 5.GUI of Cataract disease detection system

Fig. 9. GUI of Stye disease detection system

Fig. 6. A case of eye image with conjunctivitis

Fig. 7.GUI of Conjunctivitis disease detection system Fig. 10. A case of normal eye

Above Fig.7 shows conjunctivitis diseases detection in GUI

using MATLAB software. Table I shows the accuracy of classifier used in a proposed
system for different kinds of eye images which takes 30
training and testing images for each type of diseases.
TABLE IACCURACY OF PROPOSED METHOD WITH TYPES OF [10] Amol B. Jagadale, D. V. Jadhav, “Early detection and categorization of
DISEASES AND NORMAL EYE IMAGES cataract using slit-lamp images by hough circular transform,”
International Conference on Communication and Signal Processing, pp.
Minimum Distance 0232-0234, April 6-8, 2016, India.
No of No of Classifier [11] V. Harini, V Bhanumathi, “Automatic cataract classification system,”
Kind of Training Test International Conference on Communication and Signal Processing, pp.
Images
Samples Samples
No of samples
0815-0819, April 6-8, 2016, India.
correctly % Result
[12] NavneetDalal, Bill Triggs, “Histograms of Oriented Gradients for
classified human detection,” IEEE Computer Society Conference on Computer
Cataract 30 30 30 100% Vision and Pattern Recognition (CVPR'05), vol. 1, pp. 886-893, 2005.
Conjunctivitis 30 30 29 99.67% [13] Huiqi Li, JooHwee Lim, Jiang Liu, Damon Wing Kee Wong,
Stye 30 30 30 100% NganMeng Tan, Shijian Lu, Zhuo Zhang, Tien Yin Wong,
Normal eye 30 30 28 93.34% “Computerized systems for cataract grading,”2nd International
Conference on Biomedical Engineering and Informatics,pp. 1-4, 2009.

V. CONCLUSION

The proposed algorithm demonstrates the approach of

cataract, conjunctivitis and stye disease identification by
extracting best features from pupil, sclera, and eyelid of eye
utilizing HOG. HOG is a pure gradient based and captures the
object shape information well. This algorithm easily
detectsthese diseases from home using a simple digital image
which can be helpful in rural areas or hard to reach such areas
where the availability of eye specialist doctors may be less.
Our algorithm clearly classifies normal eye and eye with
cataract, conjunctivitis, stye disease using minimum distance
classifier. Result indicates 100% accuracy for cataract over 30
image, 99.67% accuracy for conjunctivitis over 30 images,
100% accuracy for stye, 93.34% accuracy for the normal eye.
Accuracy can further be increased by collecting extra dataset
for large scale trials.In the future, we will work on collecting a
more standard database using more feature and identifying
another eye sickness such as subconjunctivalhemorrhage,
chalazion, dacryocystitis.

REFERENCES

[1] AmitAsishBhadra, Manu Jain, SushilaShidnal, “Automated Detection of

Eye Diseases,” IEEE WiSPNET conference, pp. 1341-1345, 2016.
[2] Meimei Yang, Ji-Jiang Yang, Qinyan Zhang, Yu Niu, Jianqiang Li,
“Classification of retinal image for automatic cataract detection,” IEEE
15th International Conference on e-Health Networking, Applications and
Services, pp.674-679, 2013.
[3] MelihGunay, IremKucukoglu, EvginGoceri, TanerDanisman,
FadiAlturjman, “Automated detection of adenoviral conjunctivitis
disease from facial images using machine learning”, IEEE 14 th
International Conference on Machine Learning and Applications, pp.
1204-1209, 2015.
[4] MdAlamgirHossain, DebabrataSamanta, GoutamSanyal, “Eye diseases
detection based on covariance,” IRACST- International Journal of
Computer Science and Information Technology & Security (IJCSITS),
Vol. 2, No. 2, April 2012.
[5] JoydeepTamuli, Aishwarya Jain, Aaishwarya V. Dhan, AnupamaBhan,
Malay Kishore Dutta, “An image processing based method to identify
and grade conjunctivitis infected eye according to its types and
intensity,” 8th International Conference on Contemporary Computing
(IC3), pp. 88-92, 2015.
[6] JyotiPatil, Anant L. Chaudhari, “Intensity observation of conjunctivitis
using DIP,” International Journal of Applied Information Systems
(IJAIS), Vol. 4, No. 4, October 2012.
[7] JagadishNayak, “Automated classification of normal, cataract, and post
cataract optical eye images sing SVM classifier,” World Congress on
Engineering and Computer Science (WCECS), Vol. I, pp. 23-25,
October 2013.
[8] YunendahNurFuadah, Agung W. Setiawan, Tati L. R. Mengko,
Budiman, “Mobile cataract detection using optimal combination of
statistical texture analysis,” 4th International Conference on
Instrumentation, Communications, Information Technology, and
Biomedical Engineering (ICICI-BME), pp. 232-236, November 2-3,
2015.
[9] ManpreetKaur, JaspreetKaur, RavinderKaur, “Low cost cataract
detection system using smart phone,” International Conference on Green
Computing and Internet of Things (ICGCIoT), pp. 1607-1609, 2015.