Development of welding defects identifier application on radiographic film

using gray level co-occurrence matrix and backpropagation

Zaenal Abidin, Muhammad Angger Anompa, and Muhtadan

Citation: AIP Conf. Proc. 1555, 70 (2013); doi: 10.1063/1.4820996

View online: http://dx.doi.org/10.1063/1.4820996
View Table of Contents: http://proceedings.aip.org/dbt/dbt.jsp?KEY=APCPCS&Volume=1555&Issue=1
Published by the AIP Publishing LLC.

Additional information on AIP Conf. Proc.

Journal Homepage: http://proceedings.aip.org/
Journal Information: http://proceedings.aip.org/about/about_the_proceedings
Top downloads: http://proceedings.aip.org/dbt/most_downloaded.jsp?KEY=APCPCS
Information for Authors: http://proceedings.aip.org/authors/information_for_authors

Downloaded 08 Sep 2013 to 68.68.96.145. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://proceedings.aip.org/about/rights_permissions
 Development of Welding Defects Identifier Application
on Radiographic Film using Gray LevelCo-Occurrence
Matrix and Backpropagation
Zaenal Abidin*, Muhammad Angger Anompa and Muhtadan

Sekolah Tinggi Teknologi Nuklir - BATAN

Jln. Babarsari Kotak Pos 6101 YKBB Yogyakarta 55281
Telp. : (0274) 48085, 489716 ; Fax : (0274) 489715
*Email: zaenala6@gmail.com

Abstract. Development of Welding Defect Identifiers for application in Radiographic Film by using Gray Level Co-
Occurrence Matrix and Back-Propagation.A research on the application development to interpret the welding defects in
industrial radiographic films by using neural networks has been conducted. This research is aimed to produce an
application that implement the digital image processing, feature extraction and pattern recognition using artificial
neural networks. Digital image processing applied in the development is the technique of noise removal using median
filter, contrast stretching and image sharpening by Laplacian filter. Method of Grey level co-occurrence matrix (GLCM)
is applied to extract features from digital images radiographic films. Back-propagation artificial neural network method
is used for defect classification and interpretation of welding defect in radiographic films. The result of this research is
an application of back-propagation neural networks with classification results for60 simulated data with 95% of
classification successful rate.

Keywords: Radiography film, GLCM, back-propagation

PACS: 81.20.Vj

Downloaded 08 Sep 2013 to 68.68.96.145. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://proceedings.aip.org/about/rights_permissions
 INTRODUCTION system are the film digitization stage, pre-processing of
images, and identification of defects. These
Radiographic Testing (RT) is one of the most developments mostly rely on techniques such as image
important nondestructive testing techniques for welding processing, feature extraction, and pattern recognition.
inspection. It is based on the ability of X-rays or gamma The pattern classification stage is one of the most
rays to pass through metal and other materials opaque important steps in the implementation of an automated
to ordinary light, and produce photographic records by radiographic inspection system [2]. Digital image
the transmitted radiant energy [1.2]. Since different processing applied in the development is technique of
materials absorb either X-ray or gamma rays to noise removal using median filter, contrast stretching
different extents, penetrated rays show variations in and image sharpening by Laplacian filter. Method of
intensity on the receiving films. RT can examine the Grey level co-occurrence matrix (GLCM) is applied to
internal structure of a welding. Traditionally, extract features from digital images radiographic films.
experienced interpreters evaluate the welding quality Method of Back-propagation artificially neural network
based on radiography. It is time and manpower is used for defect classification and interpretation of
consuming work. In addition, human interpretation of welding defect in radiographic films [4].
welding quality based on film radiography is very
subjective, inconsistent and sometimes biased.It is METHOD
desirable to develop a computer-aided system to assist
interpretation of radiographic images to increase the
There are a lot of methods of texture extraction
objectivity, accuracy and efficiency of radiographic
including statistical texture, Gabor filter, and wavelet
inspection[1,3]. Therefore, the objectives in paper are to
analysis. In this paper, gray level co-occurrence matrix
develop an automated radiographic weld defect
(GLCM) was used. GLCM is constructed with matrix
interpreter based on Gray Level Co-Occurrence method
of joint probability density among some gray levels of
and Back-propagation neural network.
image. GLCM represents spatial relationship of any two
Currently, there are a great deal of works and
points in the image.
researches on the development of automated systems
The other texture-based feature extraction method
for inspection and analysis of radiographs. In our point
is a statistical texture method. This method performs
of view, the major steps of an automatic detection
International Conference on Theoretical and Applied Physics (lCTAP 2012)
AIP Conf. Proc. 1555, 70-74 (2013); doi: 10.1063/1.4820996
© 2013 AIP Publishing LLC 978-0-7354-1181-4/$30.00

70

Downloaded 08 Sep 2013 to 68.68.96.145. This article is copyrighted as indicated in the abstract. Reuse of AIP content is subject to the terms at: http://proceedings.aip.org/about/rights_permissions
texture analyzing based on histogram statistical moment
of image. Statistical texture does not measure image
texture based on the relationship among the intensity
values of adjacent pixels, thereby potentially give a
texture that is not appropriate because the value of the
histogram only know the intensity probability in all A statistical approach that can well describe
image. While GLCM is local analyzer, this method is second-order statistics of a texture image is a co-
constructed based on relationship between the intensity occurrence matrix. A gray-level co-occurrence matrix
o
value of adjacent pixels with certain distance of (0 , (GLCM) is essentially a two-dimensional histogram in
o o o
45 , 90 , and 135 ) GLCM produce a matrix that states which the (i,j) the element is the frequency of event i
the number of adjacent pixels within the intensity range co-occurs in event j. A co-occurrence matrix is
of the image, so that the method produces a matrix that specified by the relative frequencies P(i, j, d, ø) in
shows the image texture. Feature value can be extracted which two pixels, separated by distance d, occur in a
from GLCM by using statistical value from its matrix. direction specified by the angle ø, one with gray level i
The formula of computing several statics are as follows: and the other with gray level j. A co-occurrence matrix
Contrast: is therefore a function of distance r, angle ø and
grayscales i and j.[5,6]

PREPROCESSING OF THE IMAGES

The radiographic image has low contrast and high
Correlation: noise and the image background is not uniform. Noise
on digitized radiographic images usually appears as
randomly dispersed pixels having different values of
intensity in relation to their neighbors [6]. Adaptive
wavelet threshold and adaptive histogram equalization
techniques are used to remove the noise and improve
Energy: the contrast of the radiographic image[7]. Then
theradiographic image is segmented using multi-level
threshold based on maximum fuzzy entropy as shown
in Fig. 1.

Homogeneity:

FIGURE 1. Pixelneighborsin eightdirections.

Back-propagation
where L is gray level value, 1 and 2 are variances
that can be calculated by this formula: An Artificially Neural Network (ANN) is a
simplified simulation of biological neural networks in
human brains. ANN is capable of “learning”; that is, it
can be trained to improve its performance by either
supervised or unsupervised learning. The back-
propagation network (BPN) and the supervised
wherev is mean that can be calculated using this learning, i.e., learned by samples, are chosen in this
formula: study. After learning (ortraining), the trained weight
can be used for future prediction of debris flow
occurrence. The BPN is an ANN using back-
propagation algorithm and is one of the popular ANNs,

71
which has been widely applied to many scientific and The application program is to read bmp format
commercial fields for non-linear analysis and and jpeg image for the digitization of radiographic
prediction. The structure of BPN contains three layers, films using digital cameras are common, and because
such as, input, hidden, and output layers as shown in digital cameras generally output image format jpeg or
Fig. 2. Each layer contains I, J, and K nodes denoted bmp.
respectively by circles. The node is also called neuron Selection of Region of Interest (ROI)reduces
or unit. The circles are connected by links, denoted by operator in treating the parts that are notuseful inthe
arrows in Fig. 2, each of which represents a numerical image, thus reducing the computational time. The
weight. The wij is denoted as numerical weights second advantage is to reduce the processing based on
between input and hidden layers and so is wjk between a global approach to using an irrelevant region in the
hidden and output layers as also shown in Fig. 2. The image, which can have a negative effect on the output.
processing or the computation is performed in each Cropping is used to determine the areas that are
node in the hidden and output layers. The back- considered to have defects. In general, the size of the
propagation learning algorithm is composed of two image of radiographic films has a width (width)
procedures: (a) feed-forward and (b) back-propagation greater than the height, so cropping the image to size
weight training [8]. 300x100pixels.The cropping images done by
determining the point or suspected of having defective
pixels welding, then the program will automatically
determine the coordinates of its midpoint, the
algorithms for cropping image f are:
1. Find the coordinates (x, y) of the pixel in the image f
which is suspected disability.
2. Find the coordinates of the start cutting the steps:
a. Xmin = x - 150
b. Ymin = y– 50
c. If Xmin<1, then Xmin = 1
d. If Ymin<1, then Ymin = 1
e. If Xmin> (image width f - 300), then Xmin = width
ofthe image of f – 300
FIGURE 2. Structure of back-propagation neural network. f. Ymin<(width of the image f - 100), then Ymin =
width of the image of f - 100
Digital image radiographic films used are a film 3. Cut the image into the image of 300 x 100 with
that has a defective weld radiographs crack, porosity a. For i = 1 to 300, do
and worm hole. Furthermore, do the following steps. b. For j = 1 to 100 g(m, n) = f (i, j), for m = 1,2,3 ...
1. Create a program to read digital image radiographic 300, and n = 1,2,3 ... 100
films, both in format bmp or jpg. Then cut the Image enhancement using noise reduction surgery
image (cropping) to localize the defects contained with median filter, contrast stretching and image
image. enhancement using Laplacian filters. Feature extraction
2. Creating image improvement program. performed using Gray Level Co-occurrence Matrix.
3. Creating a program to calculate the GLCM matrix Some stages will be passed to get the feature vector of
and extracting statistical characteristics. an image in the form of a matrix. Of the image file in
4. Creating a training program with the back- the form of a matrix, Finding co occurrence matrix size
propagation method. 16x16, with 4 different angles direction so in one image,
5. Test the program using the results of previous there are 4 matrix will be sought to limit co-occurrence
training on the test file. matrix 16x16, with 4 different angles so that the
6. Creating GUI. direction of the image to be obtained 4 matrix.
After a matrix co-occurrence extraction of
Digital image used radiographic films have a statistical information from the image file is obtained,
format. Bmp or Jpg, for a program that can read digital some information is taken into extracting digital image,
image radiographic film designed to read both formats. they a recontrast, correlation, energy and homogeneity.
Bmp or bitmap files area common format for storing Eight features vector will be obtained by measuring
bitmap images developed by Microsoft and is used as mean and range of contrast, correlation, energy, and
 o o o
an uncompressed image. Meanwhile, the jpeg (Joint homogeneity in four angles that are 0 , 45 , 90 , and
o
Photographic Experts Group) format is a standard 135 .
photographic image quality. Extracting characteristic defects with Gray Level
Co-occurrence Matrix was performed on all the training

72
images of 480 training data, resulting in the
RESULTS AND DISCUSSIONS
characteristic each feature vector combined from all
training data to characterize the training data matrix
In this study, 6 radiographic weld images are
measuring 8x480.Types of defects are labeled 1 for the
used (Fig 3),image a and b are crack defect types,
crack, 2 for porosity, and 3 for wormhole. The training
images c and d are the kind of flawed porosity, as well
data matrix and labels are then stored in a file
as images e and f which are worm hole defect types.
latih_GLCM_480_data.mat
Initial weights for the training of tissue are taken
from the training data for each defect at random, and
then the final weights is determined and used as a
constant in the simulation. The expected outputs of the
classification made by the back-propagation method I
sas much as 3 classes of defects classified as grade 1
crack, Grade 2 classify defects as porosity, and grade
3classify defects as worm holes. So, designing back-
propagation method used in this study is the number of
inputs and outputs for8 piecesis3 pieces for a single
piece of digital imagery.
The testing process is the interpretation of the FIGURE 3. Image data.
test method training back-propagation. Process
simulation using the input image data from the Image cropping is done to get the partial image,
research materials used, by determining the section which contains defects that are suspected by operator;
where the image will be processed. From the testing there by it can reduce the computation load in further
process can be known whether the results of the image processing operations. Image cropping generates
training are able to recognize the pattern of defects of images with size of 300x100 pixels for further
the film radiography. processing as shown in Fig. 4.
In general, test interpretation radiographic films
defect similar to the process in the training and
simulation, starting from the test image acquisition,
image cutting, image enhancement, feature extraction
by using the method of Gray Level Co-occurrence
Matrix to get the feature vector of the test images as
input. The difference with the training process, the
classification of the testing process using the final
FIGURE 4. Digital image (size 300x100).
weight value results from previous back-propagation
training method that has been stored in
The image before and after process of image
afilemodel_BP_480_011.mat)
enhancement process include reduction of noise,
Development of applications for disability
contrast stretching, and sharpening are shown in Fig.5.
identification radiographic films using Gray Level Co-
occurrence Matrix and Back-propagation is an
application-based Graphical User Interface(GUI).The
use of GUI is intended that the resulting applications
to be user friendly and easy to use, because the FIGURE 5. Image before and after process enhancement.
application is expected to be a supporter of the
decision to an officer in interpreting radiographic The results of extraction using GLCM method is 8
defect types on film radiography. feature vector of each image, the sample results from
Making an application is done by creating the extraction characteristics to an image, statistical
functions to perform image processing, featuring characteristics are shown in Table 1.
extraction with Gray Level Co-occurrence Matrix, Back-propagation neural network method using
training with back-propagation method and the input feature vector extraction from the GLCM
interpreting of test types of defects in the film method, which has been stored in a file
radiographic images. The next step is to program a model_BP_480_011.mat. The training process includes
GUI application in Matlab using GUIDE facilities. several tests, the variation in the number of hidden layer
and the number of iterations of training, the end result
of training the neural network is in the form of weights.
Testing is done by simulating the image of training with
73
back-propagation training weighting factor of the By the result that shown in Figure 6, we conclude
previous. that generally this application can interprets weld defect
type with slightly fault interpret, and GLCM method
TABLE 1. Sample values of feature extraction result
that combined with ANN-BP method can give best
Weld defect types result for identifying or interpreting weld defect.
 Features
Crack Porosity Wormholes
Mean of 3.0758
1.4661 0.5478 CONCLUSION
contrast
Mean of 0.9546 0.9905 This application development is aimed to produce
0.8388
correlation an out put that has better quality with a little noise, un
Mean of 0.0205 0.0564 even contrast and sharper than the digital image input
0.0345 radiographic films. Image processing techniques include
energy
Mean of 0.6285 0.6964 0.8295
the use of a median filter to reduce noise, contrast
homogeneity stretching to enhance the image intensity and image
Range of enhancement to provide a sharper image, especially at
1.0196 0.1936 0.2041
contrast the edges of objects is considered defects in digital
Range of image radiographic films.
0.0528 0.0061 0.0036
correlation GLCM is capable of being used as a feature
Range of extraction method in this study (to distinguish the
0.0061 0.0023 0.0050
energy different types of disabilities). Through the resulting
Range of GLCM feature vector which will be used as input for
0.0589 0.0282 0.0382
homogenity the training program. By using GLCM ,it will be
grouped based on the value of statistical characteristics
The testing has been conducted to find ANN-BP of the distribution relationship among pixels. The
capability in classifying of weld defect type. This training is done by using back-propagation neural
testing use two digital image of radiographic film network. From the simulation results obtained that
samples in each weld defect type, so that from three back-propagation method is able to perform
types of weld defect, this testing use 6 digital image classification and interpretation with a
samples. In each sample, it has been taken 10 locations success rate of 95% of the 60test data.
of suspected weld defect. All suspected weld defect is
cropped and extracted its GLCM feature, furthermore it REFERENCES
is used as ANN-BP input.
Figure 6 shows the result of testing that represent 1. C. Hayes, Weld J 76(5), 46-51 (1997).
classifying capability of ANN-BP. Figure 6 shows that 2. X. Wang, B. S. Wong and C. S. Tan, J. Appl. Scie. Eng.
ANN-BP is capable to classify with 100%, 90%, and Technol 2(3), 295-301 (2010).
95% success rate for crack, porosity, and wormhole 3. Z. Abidin, Nuklir&Aplikasinya, STTN-BATAN, 2010
defect respectively.The "true" condition is mean that 4. Muhtadan and D. Harsono, “Pengembangan Aplikasi
untuk Perbaikan Citra Digital Film Radiografi”,
this application successfully interpret, while the "false"
Prosiding Seminar Nasional IV SDM Teknologi Nuklir
condition is mean that it interpret with the result of STTN-BATAN, Yogyakarta,2008, pp:467-478.
weld defect type is not match with the actual weld 5. L. G. Shapiro and G. C. Stockman, Computer Vision,
defect type. Upper Saddle River, N.J: Prentice Hall, 2001, p.752.
6. A. Kehoe and G. Parker, Br. J. NDT 32(4), 183-190
(1990).
7. X. Wang and B. S. Won g.
Res.Nondestruct.Eval
16(3), 132-142(2005).
8. T. C. Chang, R. J. Chao, Engineering Geology
85, 270-280(2006).

FIGURE 6. The result of simulation testing.

74