Accepted Manuscript

Title: Automatic classification of cast iron grades using

support vector machine

Authors: K. Gajalakshmi, S. Palanivel, N.J. Nalini, S.

Saravanan

PII: S0030-4026(17)31603-0
DOI: https://doi.org/10.1016/j.ijleo.2017.11.183
Reference: IJLEO 60097

To appear in:

Received date: 27-2-2017

Accepted date: 22-11-2017

Please cite this article as: Gajalakshmi K, Palanivel S, Nalini NJ, Saravanan
S, Automatic classification of cast iron grades using support vector machine,
Optik - International Journal for Light and Electron Optics (2010),
https://doi.org/10.1016/j.ijleo.2017.11.183

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
Automatic classification of cast iron grades using support vector machine

K. Gajalakshmi1*, S. Palanivel1, N.J. Nalini1, S.Saravanan2

1
Department of Computer Science and Engineering, Annamalai University, Annamalainagar-608002, Tamilnadu,
India.
2
Department of Mechanical Engineering, Annamalai University, Annamalainagar-608002, Tamilnadu, India.

T
*Corresponding author:

IP
K. Gajalakshmi

R
Research Scholar

SC
Department of Computer Science and Engineering

Annamalai University
U
N
E-mail: gajusaran@gmail.com
A

Abstract:
M

In this study, classification of three grades of cast iron viz., gray, malleable and white,
based on their texture is attempted, using Haralick features extracted from gray level co-
D

occurrence matrix (GLCM) and histogram features extracted from local binary pattern (LBP).
TE

The features were extracted from three hundred images stored in a database and are utilized to
train and test the support vector machine (SVM), to classify microstructures. The experimental
EP

results show that LBP based feature extraction achieves high accuracy when compared to
GLCM based features in classifying cast iron grades.
CC

Key words: Cast iron, Gray level co-occurrence matrix, Local binary pattern, Support vector
machine, Confusion matrix.
A

1 Introduction
Cast iron is broadly used as the base structures of manufacturing machines, rollers,
valves, pump bodies, mechanical gears and in automobile, structural, decorative sectors, owing
to its higher durability, ability to cast variety of shapes and is relatively inexpensive [1]. The
main families of cast irons viz., malleable, gray and white cast iron are defined by the

1
morphology and the distribution of graphite particles, which dictates their physical and
mechanical properties. The concentration of graphite in cast iron is determined by
microstructural analysis, performed by a material scientist visually, based on ASTM standards,
is exhaustive, time consuming and provides scope for error. Hence, automation of
microstructural analysis through textural image analysis is of immense importance.
Image processing is an essential and demanding sensor technology of the future, applied
in manufacturing industries for quality management and in non-destructive testing viz.,
inspection and characterization based on images of various components [2]. Several

T
researchers [3, 4] adopted statistical image analysis through gray level co-occurrence matrix

IP
(GLCM) feature extraction technique in metallography and other applications. Ohser et al. [5]

R
established a novel method to classify different metals based on their microstructure by
applying morphological image transformations. Meanwhile, Reboucas et al. [6] developed a

SC
new approach to calculate the nodule density of graphite in ductile cast iron, employing the
level set technique. Roberts et al. [7] employed support vector machine to classify flake

U
graphite in the microstructure of various cast iron grades. On the other hand, Velichko and
N
Mucklich [8] classified the cast iron microstructure based on graphite particles using three
A
dimensional parameters. Though classification of cast iron based on graphite particles was
reported by earlier researchers, classification of different grade of cast iron viz., malleable, gray
M

and white having varied graphite texture using support vector machine is scarce and is
attempted herein. In this study, features of various grades of cast iron images are categorised
D

by its edge features, analyzed using Gray Level Co-occurrence Matrix (GLCM) and Local
TE

Binary Pattern (LBP) techniques, whose performance measures are evaluated.

2. Methodology of proposed work
EP

The rapid development of computer vision and pattern recognition technologies is

applied in the classification of cast iron microstructure by performing three significant steps
viz., image database collection, feature extraction and classification. The step by step procedure
CC

attempted in classifying the grades of cast iron viz., gray, malleable and white is detailed in
Fig.1. Gray cast iron contains thread shaped grains, while malleable cast iron has sphere shaped
A

grains, whereas, the white cast iron has gray combinations spread on the surface. Pre-
processing steps were performed on the images collected, followed by feature extraction using
gray level co-occurrence matrix (GLCM) and local binary pattern (LBP). The extracted
features are classified by Support Vector Machine (SVM), and the performance is determined
using the measures accuracy and F-score
2.1 Feature Extraction

2
 The classification of cast iron grades in microstructure is carried out by employing
SVM technique, which demands dual input viz., features and predicted grade. The
preprocessing steps viz., resizing and improving the contrast, followed by transformation into
gray scale, for the entire images stored in the database. In this proposed work, features are
extracted by two techniques viz., GLCM and LBP, and the results are analysed.

2.1.1 Gray Level Co-occurrence Matrix

Gray level co-occurrence matrix (GLCM), a second order statistical method, represents

T
the textural features of an image having different pixel brightness distribution (gray level) in a

IP
matrix form [9-13]. This method calculate the co-occurrence of a pixel ‘i’ with a certain

R
intensity in relation with another pixel ‘j’ at a certain distance ‘d’ having orientation ‘θ’. The
outcome of the texture calculation is a single numerical number, which represents the entire

SC
window. This numerical value is positioned in the centre pixel of the window, and is shifted
by one pixel continuously and the process is repeated for calculating a new GLCM and a texture

U
measurement. In this way, the entire image is built up of texture values. Each cell in a window
N
is selected over an occupied image cell, which means that the centre pixel of the window cannot
A
be an edge pixel of the image. If a window has a dimension of N x N, a strip (N-1)/2 pixels
wide around the image will remain unoccupied. The unoccupied edge pixels are filled with the
M

nearest texture calculation. Similarly, Thirteen Haralick texture features based on Eqs. (1- 13)
were computed from each co-occurrence matrix of each image, to create a set of feature vectors
D

[14]. In addition, features affecting the spatial distribution of the gray levels in an image were
TE

determined. The number of rows (i) and columns (j) is equal to the number of gray levels (G)
used in image and each matrix element V (i, j), the value of cell (i, j), is normalized as P (i, j).
EP

2
Energy :  Pi, j 
i, j
(1)

 Pi, j log Pi, j 

CC

Entropy :  i, j
(2)

1
Homogeneity :  Pi, j  (3)
1  i  j 
i, j 2
A

2
Inertia :  i  j Pi, j
i, j
(4)

i    j   Pi, j 
Correlation :  i , j (5)
2
Shade :  i  j  2 3 Pi, j  (6)
i, j

3
 4
Prominence :  i  j  2 Pi, j
i, j
(7)

2
Variance :  i    Pi, j 
i, j
(8)

Difference Entropy: -∑ ( )log ( ( )) (9)

Sum Entropy:- ∑ ( )log ( ( )) (10)

T
Sum average : ∑ () (11)

IP
Difference variance:∑ −( ( ))) (12)

R
SC
Sum of variance: ∑ − ( ( )) (13)
where    x   y  ii j Pi, j    j j iPi, j 

and   i   i j j
U
  Pi, j     j     Pi, j 
2 2

i
N
x y
A
2.1.2 Rotation Invariant and Histogram Fourier LBP
M

Local binary pattern (LBP), a powerful means of describing texture, labels the image
pixels by thresholding the 3 x 3 neighborhood of each pixel with the center value and summing
the thresholded values weighted by the powers of two. The operator can be extended to
D

neighborhoods of different sizes by defining a circular neighborhood denoted by (p, r), where
TE

‘p’ represents the number of sampling points and ‘r’ is the radius of the neighborhood [15].
These sampling points around pixel (x, y) lie at coordinates (xp, yp) = {x + r cos (2πp/p), y−r
EP

sin (2πp/p)}. When the sampling point does not fall at integer coordinates, the pixel value is
bilinearly interpolated. Now the LBP label for the center pixel (x,y) of image f (x,y) is obtained
CC

through
P 1
LBPP, R ( x, y)   s( f ( x, y)  f ( x p , y p ))2 p (14)
P0
A

Where s(z) is thersholding function

1, z  0
s( z )   (15)
0, z  0
In the computation of the LBP histogram, uniform patterns are used so that the
histogram has a separate bin for every uniform pattern and all the non-uniform patterns were

4
assigned to a single bin. The original rotation invariant LBP operator, denoted as LBPriu2, is
achieved by rotating each bit pattern circularly to a minimum value. For instance, the bit
sequences 1000111, 11101000 and 00111010 arise from different rotations of the same local
pattern and they all correspond to the normalized sequence 0001111. All the patterns from one
°
row are replaced with a single label [16]. If the image ‘I’ is rotated by α = a , based on the

reasoning above, this rotation of the input image causes a cyclic shift in the histogram along
each of the rows,

T
hI  (Up(n, r  a))  hI (Up(n, r )) (16)

IP
Based on the property, which states that rotations induce shift in the polar representation (p, r)

R
of the neighborhood, from which features and histogram are collected by a cyclic shift. Using
the Discrete Fourier Transform (DFT), let H (n, ∙) be the DFT of nth row of the histogram hI

SC
(UP (n, r)), i.e.
p 1
H ( n, u )  h (U p ( n , r )) e  i 2 ur / P (17)
r 0
I

U
N
Now DFT provides a cyclic shift of the input vector causes a phase shift in the DFT coefficients
[17]. If hI (UP (n, r)) = h(UP (n, r − a)), then
A
H ' (n, u )  H (n, u )e i 2ur / P (18)
M

and therefore, with and 1 ≤ n1, n2 ≤ P − 1,

H ' (n1 , u)H ' (n2 , u)  H (n1 , u)e i 2ua / P H (n2 , u)e i 2ua / P  H (n1 , u)H (n2 , u),
D

(19)
Where, H (n2, u) denote the complex conjugate of H (n2, u). This shows that with any 1 ≤ n1,
TE

n2 ≤ P −1 and 0 ≤ u ≤ P −1, the features are invariant to cyclic shifts of the rows of hI (UP (n,
r)) and consequently, they are invariant also to rotations of the input image I(x, y).
EP

LBPu2  HF(n1 , n2 , u)  H (n1 , u)H (n2 , u) (20)

CC

H(n, u)  H(n, u)H(n, u) (21)

The Fourier magnitude spectrum can be considered a special case of these features.
A

Furthermore, it should be noted that the Fourier magnitude spectrum contains LBPriu2 features
as a subset, since
p 1
H ( n , 0 )   h I (U P ( n , r ))  h LBP riu 2 ( n )
r 0

This histogram characterizes occurrence statistics of simple texture primitives, as each

local pattern describes a distinct texture primitive.

5
2.2 Modeling using SVM

Support vector machine (SVM), a supervised machine learning algorithm, classifies the
image by training and testing of data, plays a vital role in image classification. A classification
task involves training and testing of data which contain some data instances [18]. Each instance
in the training set contains one target value and several features. The objective of SVM is to
produce a model which forecast the target value of data instances in the testing set, with features
alone as input [19]. Target values or known labels indicate whether the system is performing

T
satisfactorily or not, which points to a desired response, validating the accuracy of the system,

IP
or be used to help the system learn to act in a desired way.

3. Experiments

R
SC
3.1 Image Database

The optical microstructure of various grade cast iron, were determined following

U
standard metallurgical practices viz., grinding, polishing and etching in an optical microscope
N
(VERSAMAT–3) equipped with Clemex image analyzing system and stored as an image
unbiased database. The optical microstructures of a gray, malleable and white cast iron are
A
shown in Fig. 3 (a-c). Preprocessing is performed by accomplishing suitable brightness and
M

contrast adjustments. For each grade of cast iron 100 images are collected as unbiased
database, totally 300 images (gray, malleable and white) are used for training and testing
D

purpose.
TE

3.2 Pre-processing

The preliminary pre-processing stage converts the input microstructural image into
EP

a gray scale image, removing noise and resizing (240 x 240) it, for the same feature
dimension as shown in Fig. 4 (a-c). Similar procedure is performed for all the images (300
CC

images) stored in the database and recorded for further processing. The pre-processed
images having gray textures are subjected to extraction of features and it is described in
the next section.
A

3.3 Feature Extraction

The proposed method employs two salient steps viz., feature extraction and feature
classification to classify the grades of cast iron. Features are extracted based on two methods-
gray level co-occurrence matrix (GLCM) and local binary pattern (LBP).

6
3.3.1 Feature Extraction using GLCM
Feature characterize textures in a matrix form as an overall or average spatial
relationship between two pixels of gray tones in an image at a specific distance and orientation
angle (θ = 0, 45, 90, 135) as recommended by earlier researchers [11-13]. Feature is computed
by re-quantizing the original texture image into a gray image having 32 levels. GLCM is
computed by a gray co-matrix function based on horizontal proximity of the pixels: [0 1] i.e.,
the pixel on left column of the pixel of interest. Similarly, thirteen features Eqs. (1-13) are

T
extracted from each gray image ‘I’ for each image (three grades of cast iron) stored in the

IP
database.

R
Fig. 5 shows the 5 x 5 sample matrix [21:25 1:5] for each images (Fig.5 (a-c)) and the
salient four features (energy, correlation, entropy, contrast) alone are shown. From Fig.5, it is

SC
observed that the correlation, energy, entropy, dissimilarity functions have minor variations in
different grades of cast iron microstructural images. More features are required to represent the

U
image textures of different grades. Features like normalization energy, correlation, sum of
N
variance, inverse different moment, entropy, information measure of correlation, sum variance,
A
sum average, different variance, contrast and entropy are computed by Eqs. (1 – 13), make a
feature set of an image.
M

3.3.2 Feature Extraction using LBP

To extract features using LBP, initially the input gray image is divided into circles and
D

further into 240 cells. Further, a circle is placed on the image with neighbourhood denoted by
TE

(8, 1). Where ‘8’ is the number of sampling points and ‘1’ is the radius of the neighbourhood
Eq. (14). The value of the central pixel is compared with the eight neighbouring pixels
positioned at 45o (on its left-top, left-middle, left-bottom, right-top, etc) in clock wise direction.
EP

If the value of the centre pixel is greater than the neighbour, ‘1’ is recorded, else ‘0’ Eq. (15).
The eight bit binary digit obtained from LBP is converted into a decimal value. Similarly, 238
CC

X 238 values are extracted from each image. By utilizing the bit shift Eq. (16) function, the
generated circle is moved from one region to another. Then, the histogram of the entire 238 X
A

238 pixels is computed for determining the frequency of each number (pixel value) prevailing
in an image. However, features having uniform pattern is employed for describing the textures
in order to reduce the dimensionality and to improve the speed and performance.
LBP is called uniform, if it contains at most two 0-1 or 1-0 transitions. Uniform
patterns are less prone to noise and thereby, stability is more, which provides a reliable
estimation from few samples [14]. Uniform patterns are useful extension of the original

7
operator, reduce the length of the feature vector and implement a simple rotation invariant
descriptor. Ojala et al. [15] from their attempt of texture classification using LBP reported 90.0
% and 70.0 % of uniform patterns respectively in (8, 1) and (16, 2) neighbourhood. The non-
uniform patterns (more than two transitions) are grouped under a single bin for every region,
which is not considered in this study due to 1. LBP in natural images is uniform in nature. 2.
To attain statistical robustness [16]. Every regional histogram consists of p (p − 1) + 3 bins: p
(p − 1) bins for the patterns with two transitions, two bins for the patterns with zero transitions
and one bin for all non-uniform patterns.

T
In this proposed work, 59 features (p=8, 8(7) + 3=59) are generated from each image.

IP
In the computation of LBP histogram, a separate bin for every uniform pattern, (1,1) 0 to 7,

R
(2,1), 8 to 15………..(8,1) 56 (9,1) 57 (10,1) 58 patterns are stored as features. Using uniform
patterns, the length of the feature vector of an image reduces from 256 to 59 [17]. Even then,

SC
some insignificant information was also present, removed by DFT, which sharpens the edge
points, in the reduced 59 features. After execution, the same image is rotated by 90 degree and

U
same steps are repeated to obtain another 59 image features, and stored. Fast Fourier Transform
N
(FFT) derived from DFT, determines the most significant features from the available (59
A
features), using equation given as
(p-1) * (floor (p/2) +1) + 3 (18)
M

Where, floor is a function used to round the value of ‘p’ to the nearest integer. Ignoring

all repeated and non influential features, 38 significant features were determined to represent
D

the image texture. The feature vector can now be processed using the support vector machine to
TE

classify the images.

EP

3.4 Training and Testing

In this study, microstructural images of gray, white and malleable cast iron were
CC

classified through the SVM classifier. The features extracted by both GLCM and LBP
techniques are given as input for the training. On the other hand, the result of the manual cast
iron categorization by a human expert is employed as grade labels viz., Grade 1 for gray cast
A

iron (thread shape grains), Grade 2 for malleable cast iron (sphere shaped grains), Grade 3 for
white cast iron (gray combinations spread on the surface). The features extracted by GLCM
and LBP techniques are trained separately by SVM following ‘leave-one-out’ (LOO) algorithm
in ‘K-fold’ method using the SVM torch tool. In this algorithm, equal size of sample is selected
leaving one fold in each training set, for testing. Using three fold methods, three different

8
combinations of training and testing were performed using SVM Torch as recommended by
Tang et al. [20].
Jaochims [21] employed conventional SVM for binary classification, however in this
study, a multiclass classifier is attempted to classify three categories. Out of three hundred
(300) images in the database, hundred images are used for testing and the remaining is used for
training. Post training and testing, the performance of the system is evaluated by determining
the accuracy. Experiments are conducted in a Pentium Intel Core-i5 PC and the necessary
coding is developed using Matlab-2013b.

T
4. Experimental results

IP
4.1 Performance measurement

R
Earlier researchers [22, 23] reported classification of two categories viz., positive (P)
or negative (N). To distinguish between the actual class and the predicted grade labels {G1

SC
(gray cast iron), G2 (malleable cast iron), G3 (White cast iron)} is adopted for the grade
predictions developed by this model. Given a classifier and an instance, there are four possible

U
outcomes. If the instance is positive and it is classified as positive, it is counted as a true positive
N
(TP); if it is classified as negative, it is counted as a false negative (FN). If the instance is
A
negative and it is classified as negative, it is counted as a true negative (TN); if it is classified
as positive, it is counted as a false positive (FP).
M

Precision = (23)
D

Recall = (24)
TE

Accuracy = (25)
Specificity = ( ) (26)
EP

2
F − score = (27)
1⁄precision + 1⁄recall
CC

FP, TP, TN and FN values were extracted from the confusion matrix .The evaluation
metrics like precision, recall, accuracy, specificity and F-score are calculated from above
equation to predict the performance of the developed system Eqs. (23-27). The confusion
A

matrix require two dimensions, one is indexed by the actual grade (manual inspected) of an
object, while the other is indexed by the grade that the classifier predicts [18, 19]. The
performance measure of the classifier system employ a test set of features excluded from the
training set. Thus, each time 100 images in the database is used for testing and the classifier
works three times using 3- fold method. The sample of accurate classification and

9
misclassification in each grade of cast iron attempted based on microstructure is summarized
in Fig.7. The first row shows the gray cast iron microstructures, which reveal four visually
similar microstructures and the misclassified image. While, the malleable cast iron
microstructure shown in the second row appears to have similar visual texture but few are
misclassified due to variation in grains. Whereas the microstructure shown in the third row
(white cast iron) contains visually similar spherical features, entirely different from the earlier
two grades.
The overall accuracy is the average accuracy of the attempted three cross-validations.

T
The confusion matrix shown in Table 1, show the classification performance of the developed

IP
system with respect to each individual grade. The correct classifications lie along the diagonal

R
of the confusion matrix. In this study, two methods with three grades of classification are
attempted.

SC
U
From Table 1, it is observed that, out of 300 images categorized into three grades (G1,
N
G2 and G3), 89.0, 88.0 and 98.0 percent images were accurately classified by GLCM
A
technique. 11.0% of the gray cast iron micrographs are misclassified as malleable or white,
while 12.0 % of malleable cast irons are misclassified as the gray or white. The
M

misclassification between gray and malleable iron is significant due to the prevalence of gray
features in both microstructural classes. In contrast, majority of the white cast iron were
D

classified accurately, except a few. In case of LBP (Table 2), the classification is more precise
TE

as it results in 92.0, 96.0 and 99.0 percent exact predictions. The misclassification of grade 1
and grade 2 reduces to 8.0 and 4.0 percent, whereas, the predication of the white cast iron is
EP

more accurate (99.0 %) and is consistent with Sivakumar et al. [24].

CC

After generating confusion matrix, performance is measured through accuracy,

precision, recall, specificity and F-score [25]. Performance of cast iron classification using
A

GLCM and LBP are shown in Table 3 and Table 4 respectively. The accuracy of grade
classification based on GLCM is 91.6 %, whereas, LBP technique provides 96.3 % accuracy
(Fig. 8) along with a shorter processing time (< 5 ms). The closer prediction makes it suitable
for utilization during classification of large volume of cast iron microstructures.

5. Conclusions

10
 The classification of cast iron grades based on their microstructure employing support
vector machine is proposed in this study. The following salient conclusions were drawn from
this experimental study. Support vector machine utilizing GLCM and LBP techniques can
effectively be employed in classifying gray, malleable and white cast iron based on their optical
microstructure. The cast iron classification employing GLCM and LBP technique results in
91.6 % to 96.3 % accuracy, respectively.

T
R IP
SC
U
N
A
M
D
TE
EP
CC
A

11
References
[1] R. Singh, Applied Welding Engineering Processes, Codes and Standard, Second ed.
Elsevier (2015), London.
[2] S.G. Lee, Y. Mao, A.M. Gokhale, J. Harris, M.F. Horstemeyer, Application of digital
image processing for automatic detection and characterization of cracked constituent
particles/inclusions in wrought aluminium alloys, Materials Characterization,60(9)
(2009) 964-970.
[3] C.E. Honeycutt., R. Plotnick. Image analysis techniques and gray-level co-occurrence
matrices (GLCM) for calculating bioturbation indices and characterizing biogenic
sedimentary structures, Computational Geosciences. 34 (11) (2008)1461–1472.
[4] B. Bataineh, S.N.H.S. Abdullah, K. Omar, A novel statistical feature extraction method

T
for textual images: Optical font recognition, Expert Systems with Applications, 39 (5)

IP
(2012)5470-5477.
[5] J. Ohser, F. Muecklich. Statistical Analysis of Microstructures in Materials Science.
John Wiley & Sons, Chichester (2000).

R
[6] E.S. Rebouças, A.M. Braga, R.C.P. Marques, P.P.R Filho, A new approach to calculate
the nodule density of ductile cast iron graphite using a Level Set, Measurement,

SC
89(2016), 316–321.
[7] K. Roberts, G. Weikum, F. Mucklich. Examinations on the Automatic Classification of
Lamellar Graphite Using the Support Vector Machine. Praktische Metallographie, 42
(8) (2005)396-410.
U
[8] A Velichko, F. Mücklich, Shape Analysis and Classification of Irregular Graphite
N
Morphology in Cast Iron, Praktische Metallographie. 43 (4) (2006)192-207.
[9] V. Karathanassi, C.H Iossifidis, D. Rokos, A texture-based classification method for
A
classifying built areas according to their density, international journal remote sensing
application, 21 (9) (2000) 1807-1823.
M

[10] B. Julesz, Visual Pattern Discrimination, IRE Transactions on Information Theory IT-
8(2) (1962) 84-92.
[11] V. Karathanassi, C.H Iossifidis, D Rokos, A texture-based classification method for
D

classifying built areas according to their density, International Journal of Remote

Sensing Application, 21 (9) (2000)1807-1823.
TE

[12] K. Darabi, G. Ghinea. User-centred personalised video abstraction approach adopting

SIFT features, Multimedia Tools and Applications, 76 (2) (2017) 2353–2378.
[13] R.W. Conners, M. Trivedi, C.A. Harlow, Segmentation of a High Resolution Urban
EP

Scene using Texture Operators Computer Vision, Graphics and Image Processing.
25(1984)273-310.
[14] R.M. Haralick, K. Shanmugam, I. Dinstein, Textural features for image classification
IEEE Transactions on Systems Man and Cybernetics - Part A (1973) 610-621.
CC

[15] T. Ojala, M. Pietikainen, T. Maenpaa, Multi resolution gray-scale and rotation invariant
texture classification with local binary patterns. IEEE transactions on pattern analysis
and machine intelligence, 24(7) (2002)971–987.
A

[16] T. Ahonen, J. Matas, C .He, M. Pietikainen, Rotation Invariant Image Description with
Local Binary Pattern Histogram Fourier Features, A.-B. Salberg, J.Y. Hardeberg, and
R. Jenssen (Eds.): SCIA 2009, LNCS 5575 61–70(2009).Springer-Verlag Berlin
Heidelberg.
[17] G. Zhao, T. Ahonen, J. Matas, M. Pietikainen, Rotation Invariant Image and Video
Description with Local Binary Pattern Features IEEE Transcation Image Process, 21
(4) (2010)1465-1477.
[18] N. Cristianini, J.S. Taylor, An Introduction to Support Vector Machines and Other
Kernel-based Learning Methods (2000) Cambridge University Press, London.

12
[19] K.B. Duan, J.C. Rajapakse, H .Wang, F. Azuaje, Multiple SVM-RFE for Gene
Selection in Cancer Classification with Expression Data, IEEE Transactions on Nano
Bioscience, 4(3) (2005) 228 – 234.
[20] F. Tang, I. You, C. Tang, M. Guo. An efficient classification approach for large-scale
mobile ubiquitous computing. Information Sciences, 232(2013) 419-436.
[21] T. Joachims. Text Categorization with Support Vector Machines: Learning with Many
Relevant Features, ECML, Springer-Verlag (1998), London.
[22] X. Deng, Q. Liu, Y. Deng, S. Mahadeven. An improved method to construct basic
probability assignment based on the confusion matrix for classification problem,
Information Sciences, 340–341(2016)250–261.
[23] T.S. Subashini, V. Ramalingam, S. Palanivel. Breast mass classification based on

T
cytological patterns using RBFNN and SVM, Expert System with Application, 36(3)
(2009)5284–5290.

IP
[24] G. Sivakumar, S. Saravanan, K. Raghukandan. Investigation of microstructure and
mechanical properties of Nd: YAG laser welded lean duplex stainless steel joints, Optik

R
- International Journal for Light and Electron Optics,131(2017) 1-10.
[25] Chan B.S, Sia, C.L, Wong F, Chin, K.Y, Dargham, J.A., Yang S.S. Analysis of Surface

SC
Electromyography for On-Off Control. Advanced Materials Research. 701 (2013) 435-
439.

U
N
A
M
D
TE
EP
CC
A

13
 GLCM
Input
Feature SVM Cast iron
Preprocessing
Extraction Classifier Grade
image LBP

Fig. 1. Proposed work.

T
R IP
SC
U
N
Fig. 2. SVM classification.
A
M
D
TE
EP
CC

Fig. 3. Cast iron microstructure. (a) Gray. (b) Malleable. (c) White.
A

14
 T
Fig. 4. Gray input images. (a) Gray. (b) Malleable. (c) White.

R IP
SC
U
N
(a) (b)
(c)
A
M
D

Fig. 5. Sample GLCM features of (a) Gray. (b) Malleable. (c) White.
TE
EP
CC

(a) (b) (c)

A

Fig. 6 Sample LBP features of (a) Gray. (b) Malleable. (c) White.

15
Grades Classified images Misclassified images

Gray cast iron (G1)

Malleable cast iron (G2)

White cast iron (G3)

T
IP
Fig. 7. Classified and misclassified images.

R
SC
100 96.3
91.6
Accuracy (%)

75

U
N
50
A
25
M

0
GLCM LBP
D

Fig. 8. Performance of cast iron classification.

TE
EP
CC
A

16
 Table 1. Confusion matrix for GLCM

. Predicted grade
(%)
Actual
G1 G2 G3
grade
G1 89.0 10.0 1.0
G2 11.0 88.0 1.0
G3 1.0 1.0 98.0

T
R IP
SC
Table 2. Confusion matrix for LBP.

Predicated
(%)
grade
U
N
Actual
G1 G2 G3
grade
A
G1 92.0 8.0 0.0
G2 3.0 96.0 1.0
M

G3 0.0 1.0 99.0

D
TE
EP

Table 3. Performance of cast iron classification using GLCM

CC

Grade Precision Recall Specificity F-score

G1 89.0 89.0 90.0 88.0

A

G2 88.0 88.0 94.0 88.0

G3 98.0 97.0 99.0 97.0

17
Table 4. Performance of cast iron classification using LBP

Grade Precision Recall Specificity F-score

G1 92.0 91.0 96.0 90.0

G2 96.0 95.0 97.0 95.0

G3 99.0 99.0 99.0 98.0

T
R IP
SC
U
N
A
M
D
TE
EP
CC
A

18