T SHANTHI et al.

: FUNDUS IMAGE CLASSIFICATION USING HYBRIDIZED GLCM FEATURES AND WAVELET FEATURES
DOI: 10.21917/ijivp.2021.0337

FUNDUS IMAGE CLASSIFICATION USING HYBRIDIZED GLCM FEATURES AND

WAVELET FEATURES
T. Shanthi, R.S. Sabeenian, K. Manju, M.E. Paramasivam, P.M. Dinesh and R. Anand
Department of Electronics and Communication Engineering, Sona College of Technology, India

Abstract second opinion at an ease [7]. According to the report from World
We find the usefulness of computers in every field including medical Health Organization (WHO), India had 69.2 million people
field. Scanning the affected part has become a standard study. suffering with diabetes in 2015 and the diabetic population is
Diagnosing a disease at the right time, i.e. early detection, from the expected to rise to 98 million in 2030. The annual death rate of
study of images enables the physician to take right decision and provide diabetic patients in India is about 1 million. This alarming
proper treatment to the patient. With the alarming growth of situation calls for immediate attention and redress. About 90% of
population, it is difficult for every individual patient to get a second
people with diabetes for long period of time are at higher risk of
opinion from medical expert. In these situations, computer-aided
automatic diagnosis system will be much helpful. Diabetic retinopathy
developing diabetic retinopathy. The number of patients per
is a disorder that arises from increase in blood glucose level. Based on ophthalmologist is very high in the order of 1 lakh patients per
the severity, it has been distinguished into four stages. Diagnosing ophthalmologist [8]. In this context, automation in diagnosis
diabetic retinopathy at an early stage from retinal images and providing facilitates the process of diagnosis in an efficient manner. An
proper treatment will save the patient from severe vision loss. The automatic system will be of immense help to the doctors in saving
proposed method adopts hybridized GLCM features and wavelet their precious time so that more attention can be paid to the people
features to classify the fundus images according to the severity of the at high risk [9]. Diagnosis of diabetic retinopathy (DR) at an early
disease. The method is tested with fundus images collected from Indian stage and taking right treatment can save the patients from severe
Diabetic Retinopathy Dataset. vision loss. This paper discusses the automatic diagnosis of
diabetic retinopathy (DR) through the analysis of DR fundus
Keywords: images using combination of GLCM and WDM features [10].
Fundus Image, GLCM, WDM Features, Diabetic Retinopathy,
Classification

1. INTRODUCTION
A picture describes a scene efficiently and conveys the
information in better way. Human beings are good at interpreting
details from an image [1]. More than 90 percentage of data
communicated to brain is visual data. Human brain responds and
process visual data 60,000 times better than any other form of Fig.1. Vision of Normal person and DR Patient
data. Image processing systems needs representation of image in
digital form. Latest inventions of high end cameras and imaging
techniques make this possible at an ease [2]. A digital image is a
1.1 LITERATURE REVIEW
two dimensional array of numbers, where the numbers represent Image classification is an analytical approach to categorize
the intensity values of the image at various spatial locations. objects in an image or the whole image into classes or themes.
These pixels possess spatial coherence which can be inherited by Image processing is seldom carried out without image
performing arithmetic operations like addition, subtraction etc. classification. Exemplary researches have been and are being
[3]. The statistical manipulations of the pixel values help to carried out in image processing in general, but specifically in
develop an image processing techniques for variety of image classification [11].
applications. Image classification is a process in which a machine
In [12], the author present methods of classification are
is used to classify the images utilizing the spatial relativity of the
described. Feature extraction and classifiers are two important
pixels [4].
components of an image classification system. Features should be
Computer-assisted medical services are growing at a fast pace. selected properly and with due relevance. The purpose of
Diagnosing a disease correctly is equivalent to providing 50% of employing a classifier is to get enhanced accuracy in
the required treatment. Most of the loss in human life is because classification. Pioneer researchers have proposed new approaches
of incorrect diagnosis of disease or diagnosing a disease at its final to extract features of interest and thereby to reduce the
stage [5]. Recently a growing number of research projects have computation time.
been initiated in medical science where computers play a crucial
In [13], the author have attempted to detect diabetic
role. Automatic diagnosis of certain diseases will be of great
retinopathy automatically using a back propagation neural
service to mankind. Computer assisted diagnosis is the need of the
network. The network was evaluated with 179 images obtained
hour as more number of patients require the service of experts for
from 32 normal subjects and 147 diabetic patients. The images of
their advice and for taking proper treatment [6]. Automated
size 700×700 were captured using Canon fundus camera. Red free
diagnosis system helps the physicians to diagnose the disease
images were obtained by extracting green plane from the input
correctly and saves a lot of time. It also helps the persons to get a

2372
ISSN: 0976-9102 (ONLINE) ICTACT JOURNAL ON IMAGE AND VIDEO PROCESSING, FEBRUARY 2021, VOLUME: 11, ISSUE: 03

color image. The network has three layers, input layer with 400 the image. Hence the green component is extracted from the
nodes, hidden layer and output layer with 4 nodes. Recognition original RGB image as shown in Fig.3. These preprocessing steps
rates for the detection of hemorrhages, vessels, and exudates were mentioned are used to improve the overall quality of the image.
73.8%, 91.7% and 93.1% respectively.
In [14], the author classify DR images with the accuracy of
2.2 GLCM FEATURES
93% by segmenting vessels, exudates and microaneurysm from The Gray Level Co-occurrence Matrix method - likewise often
the fundus image. The algorithm has been tested with the publicly named as the Spatial Gray Level Dependence Matrix (SGLDM)
available DRIVE, DIARTDB1 and MESSIDOR databases by Method. This method is based on the intensity values of the
using at trained SVM classifier. image. The features when fed to the machine learning algorithm
performs the classification effectively [1] [2].
1.2 FUNDUS IMAGE DATASET DESCRIPTION
The second-order statistics are attained by considering a set of
The IDRID is a treasure of considerable number of retinal pixels related to each other in positive three-dimensions. The
images representing all grades of Diabetic Retinopathy (DR), Gray Level Cooccurrence matrices provide rare mathematical
Diabetic Macular Edema (DME), and healthy retina. A digital statistics on the texture. Texture feature calculation use the
fundus camera, namely, Kowa VX-10 alpha (Fig.2) was used to content of the GLCM to give a measure of the variation in
register the images. The images were taken in such a manner that intensity at the pixel of interest. GLCM matrix is a square matrix
the focus was centred near macula with 500-field of vision. The that has same number of rows and same number of columns.
camera captured the fundus images with a resolution of GLCM matrix is N×N matrix, where N denotes the number of
42884288 pixels. The images were produced in JPG format. possible gray levels in an image. For example, a 3-bit image will
Each image occupied a storage space of 800 KB. have a GLCM matrix of size 8×8. The rows and columns are the
gray values (0-7). GLCM matrix of image depends on the
direction and offset values. The direction can be anyone among
the eight possible directions as shown in the Figure1. The offset
represents distance between pixels.

Healthy Retina Mild NPDR Moderate NPDR

Severe NPDR PDR Fig.3. Co-occurrence Matrix

Fig.2. Fundus images showing healthy retina and retina with This paper utilizes 8 different features that are computed from
different grades of diabetic retinopathy GLCM Matrix [13]. These 8 features are elaborated as follows.
2.2.1 Entropy:
The rest of the article is organized as follows. Section 2
presents the method proposed in this paper. Section 3 describes in Entropy measures the disorder or complexity of an image. If
detail about the experimentation and discussions. The conclusion the entropy is large when the image is not texturally uniform and
is provided in section 4. if the texture is complex, then entropy becomes high. Energy and
entropy are inversely proportional to each other, it means if
entropy is high, energy will be low and vice-verse. Entropy can
2. METHODOLOGY USED be calculated for image f(x,y) using the following equation,
N 1 N 1
An image classification algorithm needs a feature extraction Entropy  E    f  x, y  *log  f  x, y   (1)
followed by a classifier. The classifier in this paper is fed with two x 0 y 0
sets of features and then followed by hybrid features. GLCM and
WDM are the features used in this paper. 2.2.2 Contrast:
Contrast measures the spatial frequency of an image and
2.1 PREPROCESSING different moment of GLCM. It is difference between the highest
and the lowest values of a contiguous set of pixels. Contrast can
In the proposed method, the fundus image from IDRID
be calculated for image f(x,y) using the following equation,
database is taken as the input. The input image is preprocessed so
N 1 N 1
that it enhances the result in further steps. In preprocessing stage, Contrast C     x  y  f xy
2
(2)
the input image of size 4288×2848 is resized to 150×150. The x 0 y 0
clarity of the blood vessels are prominent in the green channel of

2373
T SHANTHI et al.: FUNDUS IMAGE CLASSIFICATION USING HYBRIDIZED GLCM FEATURES AND WAVELET FEATURES

2.2.3 Mean: information, it is used as input for the next level of decomposition.
Mean is the normal value of the whole amount of data. This process continues for further level of decomposition.
Mathematically,  x1 x2 
The input matrix is divided into 22 cells as 
x4 
N 1 N 1
and the
 x3
 f  x, y  wavelet decomposed matrix is obtained as
Mean  M   x 0 y 0
(3)
M N W WVP 
WP   AP  (9)
2.2.4 Standard Deviation: WHP WDP 
The standard deviation is the square root of variance divided where,
by the entire number of samples. Mathematically expressed as,
WAP is approximation component = 1/4(x1+x2+x3+x4),
 x
2
Wvp is vertical component = 1/4(x1+x2-x3-x4),
SD  (4)
N WHP is horizontal component = 1/4(x1-x2+x3-x4),
2.2.5 Variance: WDP is diagonal component = 1/4(x1-x2-x3+x4).
This statistic measures the heterogeneity and it is strongly From each level of wavelet decomposition matrix, the
correlated with first order statistical variable such as standard covariance matrix and correlation matrices are computed.
deviation. Variance can be calculated for image f(x,y) using the
following equation, 3. EXPERIMENTS AND DISCUSSIONS
N 1 N 1
Variance V     i    f xy
2
(5) The features are fed to the classifiers individually and then in
x 0 y 0 mixed combination. The accuracies obtained using GLCM feature
where μ is the mean of an input image. spaces is provided in Table.1. The accuracy obtained from
variance and homogeneity were higher than other features. The
2.2.6 Homogeneity:
accuracies obtained using WDM feature spaces is provided in
Homogeneity is the consistency in the arrangement of an input Table.2. The accuracy obtained from W4 and W5 are higher than
image f(x,y). If the arrangement of an input image follow a regular other features. These high performing features are combined and
pattern then it is said to be homogenous. Mathematically can be fed to the classifier. The accuracies obtained using hybrid feature
expressed by the following equation, spaces is provided in Table.3. The accuracy obtained from
N 1 N 1 f xy homogeneity and fourth level WDM is higher than other features
Homogenous  H    (6)
1 x  y
2
x 0 y 0
Table.1. Accuracy results using GLCM feature spaces
2.2.7 RMS Contrast: Class E C M SD V H RC S
Root Mean Square (RMS) measure the standard deviation of DR Grade 0 76.6 75.6 85.3 77.4 87.2 88.8 79.6 77.1
the pixel intensities. It does not depend upon any angular
frequency or spatial distribution of contrast of an input image. DR Grade 1 82.5 84.1 92.2 78.3 84.4 86.2 80.1 78.4
Mathematically it can be expressed as DR Grade 2 75.1 83.5 89.8 94.8 96.9 95.8 87.5 91.9
N 1 N 1 DR Grade 3 84.6 87.6 84.8 90.5 97.9 93.9 92.7 88.4
  I I 
1
RMS Constrast  RC  
2
ij (7)
MN x 0 y 0
Table.2. Accuracy results using WDM feature spaces
where, I is the normalized pixel intensity values between 0 and 1
Class W1 W2 W3 W4 W5 W6
2.2.8 Smoothness:
DR Grade 0 82.63 87.65 95.32 97.43 92.29 87.83
The input image f(x,y) is highly correlated between the
DR Grade 1 82.56 88.13 92.23 98.34 94.43 86.22
adjacent pixels then we can say the image are autocorrelated (the
autocorrelation of input data with itself after shifting one pixel). DR Grade 2 85.17 83.52 94.80 94.87 96.90 95.82
The autocorrection of input image can be calculated by the DR Grade 3 84.60 87.61 96.82 96.59 97.99 93.91
following equations.
autocorrelation f  x, y     f  x, y  * f  x  x  y  y dxdy (8) Table.3. Accuracy results using hybrid feature spaces
Class V-W3 V-W4 H-W3 H-W4
2.3 WDM FEATURES DR Grade 0 82.63 94.43 95.32 97.43
The wavelet Decomposition matrix is obtained by using DR Grade 1 88.56 93.90 92.23 98.34
wavelets. The advantage of wavelet features is that the image can DR Grade 2 85.17 97.99 94.80 94.87
be analysed at different resolutions. The Haar wavelet is used in
this paper. The first level of decomposition gives one approximate DR Grade 3 84.60 87.61 96.82 96.59
component, diagonal component, vertical and horizontal Average 85.24 93.48 94.79 96.80
components. As the approximate component contains more

2374
ISSN: 0976-9102 (ONLINE) ICTACT JOURNAL ON IMAGE AND VIDEO PROCESSING, FEBRUARY 2021, VOLUME: 11, ISSUE: 03

4. CONCLUSIONS Engineering and Informatics, Vol. 1, No. 2, pp. 459-466,

2009.
A lot of research work is available for image classification [6] R.S. Sabeenian and V. Palanisamy, “Comparison of
problem but most of the cases are binary classification. The Efficiency for Texture Image Classification Using MRMRF
proposed work aims at providing multiclass classification with and GLCM Techniques”, International Journal of
satisfactory results. In future the proposed work can be further Computers Information Technology and Engineering, Vol.
extended to include more classes with minimum loss and 2, No. 2, pp. 87-93, 2008.
maximum classification accuracy. The proposed method is tested [7] Robert M. Haralick and Karthikeyan Shanmugam, “Textural
with individual features and then with the hybrid features. The Features for Image Classification”, IEEE Transactions on
noticeable characteristic of the method proposed is that it is very Systems, Man, and Cybernetics, Vol. 6, pp. 610-621, 1973.
simple to implement with promising results. The results can be [8] Manik Varma and Andrew Zisserman, “A Statistical
improved further by adopting better classifiers. Approach to Texture Classification from Single Images”,
International Journal of Computer Vision, Vol. 62, No. 1-2,
pp. 61-81, 2005.
REFERENCES
[9] A.K. Johan and Joos Vandewalle, “Least Squares Support
Vector Machine Classifiers”, Neural Processing Letters,
[1] R. Anand, T. Shanthi, M.S. Nithish and S. Lakshman, “Face
Vol. 9, No. 3, pp. 293-300, 1999.
Recognition and Classification using Google NET
[10] B. Scholkopf and A.J. Smola, “Learning with Kernels:
Architecture”, Advances in Intelligent Systems and
Support Vector Machines, Regularization, Optimization,
Computing, Vol. 1048, pp. 261-269, 2020.
and Beyond”, MIT Press, 2001.
[2] S. Veni, R. Anand and D. Vivek, “Driver Assistance through
[11] K.P. Bennett and A. Demiriz, “Semi-Supervised Support
Geo-fencing, Sign Board Detection and Reporting Using
Vector Machines”, Advances in Neural Information
Android Smartphone”, Advances in Intelligent Systems and
Processing Systems, pp. 368-374, 1999.
Computing, Vol 1057, pp. 361-372, 2020.
[12] T. Shanthi, R.S. Sabeenian and R. Anand, “Automatic
[3] G. Kylberg. “The Kylberg Texture Dataset v. 1.0, Centre for
Diagnosis of Skin Diseases using Convolution Neural
Image Analysis”, Technical Report, Swedish University of
Network”, Microprocessors and Microsystems, Vol. 76, pp.
Agricultural Sciences and Uppsala University, pp. 1-132,
1-20, 2020.
2020.
[13] T. Shanthi and R.S. Sabeenian, “Modified Alexnet
[4] R. Anand, S. Veni and J. Aravinth, “An Application of
Architecture for Classification of Diabetic Retinopathy
Image Processing Techniques for Detection of Diseases on
Images”, Computers and Electrical Engineering, Vol. 76,
Brinjal Leaves using K-Means Clustering Method”,
pp. 56-64, 2019.
Proceedings of International Conference on Recent Trends
[14] Mryka Hall-Beyer, “GLCM Texture: A Tutorial”, Available
in Information Technology, pp. 1-7, 2016.
at
[5] R.S. Sabeenian and V. Palanisamy, “Texture-Based Medical
http://www.ucalgary.ca/UofC/nasdev/mhallbey/research.ht
Image Classification of Computed Tomography Images
m, Accessed at 2017.
using MRCSF”, International Journal of Medical

2375