Original Paper

Measurement and Control

2019, Vol. 52(7-8) 804–813
Automatic visual inspection system for Ó The Author(s) 2019
Article reuse guidelines:
quality control of the sandwich panel sagepub.com/journals-permissions
DOI: 10.1177/0020294019847706

and detecting the dipping and buckling journals.sagepub.com/home/mac

of the surfaces

Vahid Torkzadeh and Saeed Toosizadeh

Abstract
In this paper, an automated inspection system is proposed for detecting the location and measuring the size of existing
dipping or buckling on the sandwich panel surface using an RGB camera and an inexpensive linear laser. The proposed
method, by observing the radiated laser beams on the sandwich panel surface, can localize and calculate the level of dip-
ping and buckling with acceptable accuracy while being robust to vibrations of moving sandwich panel on the production
line conveyor. After a complete processing of the panel by the system, a three-dimensional (surface plot) or two-
dimensional (heat map) output is produced to assist the production line supervisor to easily inspect the surface quality
of the sandwich panels. Our experimental results show that the proposed system can detect and measure the surface
defects including dipping and buckling with a high accuracy and performance.

Keywords
Calibration, defect understanding, digital image processing, industrial inspection, visible lasers

Date received: 22 October 2018; accepted: 21 March 2019

Introduction inspection of painting and thickness, metal sheet bezel

and fracturing inspection before manufacturing, cello-
The sandwich panel manufacturing industry is of cru- phane cover on sheet as ordered, correct shaping of
cial importance around the world. The sandwich panels sheets in forming station, adapting the molds width
are used in production of the fixed and portable Conex with the specified width in the customer order, double
boxes, sustainable prefabricated buildings and prefabri- belt alignment with forming and coil holders, and lack
cated sanitary units in the world. Compared to other of dipping and buckling on the sandwich panel surface.
industries, designing an automated inspection system Also, the examination of the thickness, width, angle,
for this industry has many unique challenges. and density of the produced foam to meet customers’
Considering the production of the sandwich panels in demands is of crucial importance. These items are per-
custom dimensions and characteristics, the parameters formed in five stations of forming, injection, double
and setting of the line production are changing con- belt, cutting, and packaging.
tinuously. Hence, the definition of a fixed set of Diagnosing the physical problems of foam plates
instructions and constant settings that the production (Figure 1), such as bending, buckling or dipping, is of
line can work optimally and with a desirable quality crucial economical importance in the initial stages of
are not practically possible.
production. The precision tool systems are traditionally
Also, unlike most of the automated visual inspection
costly and using them involves modifying the produc-
systems that control the quality of discrete productions
tion line. In this work, we try to solve these challenges
(such as auto production line and food industry), the
production line of sandwich panel is continuous, and
the products with variable dimensions and specifica- Department of Computer Engineering, Islamic Azad University,
tions (depending on the customer order type) are manu- Neyshabur, Iran
factured continuously involving various operations at
Corresponding author:
separate stations (such as forming and injection). Some Saeed Toosizadeh, Department of Computer Engineering, Islamic Azad
instances of inspection and quality control of the sand- University, Neyshabur Branch, Neyshabur, Iran, 9319975853.
wich panel production process are the quality Email: [email protected]

Creative Commons CC BY: This article is distributed under the terms of the Creative Commons Attribution 4.0 License
(http://www.creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work without
further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/
open-access-at-sage).
Torkzadeh and Toosizadeh 805

Figure 1. Samples of sandwich panels.

by inexpensive hardware and employing techniques of and cracks on the surface.14–17 The main drawback of
image processing, computer vision, and machine learn- the methods in this category is that they often need to
ing. Therefore, in this paper, our main goal is to design perform complex transformations to frequency domain
a system that has acceptable precision as well as afford- and therefore may have low performance.
ability. Furthermore, we aim to provide the possibility Another category of the existing methods is model-
of checking the sandwich panel’s quality with no or lit- based which in particular uses fractal geometry.18–21 In
tle modifications in the production line. these methods, due to the lack of a standard modeling
Since an extensive set of variables affects the produc- process, model design and construction are rather
tion of sandwich panels and, to the best of our knowl- difficult.
edge, there is no comprehensive study about the There is also another kind of the proposed methods
usability of computer vision and artificial intelligence which attempts to solve this problem by employing
methods for quality control of the sandwich panels, this machine learning and artificial intelligence techniques
paper only focuses on detection of the dipping and alone, or sometimes by combining them with previous
buckling on the surface of sandwich panels as an initial methods.22 Jiuliang et al.23 proposed a fuzzy expert sys-
phase of designing and implementing a comprehensive tem to assess the steel strips surface quality which con-
automated visual inspection system. sists of six components. A monitoring system was
proposed by Zheng et al.24 for detecting the metal sur-
face defections, which in particular could detect the
Literature review cracks and existing gaps on aluminum surfaces. The
In recent years, some methods, which employ computer accuracy of their solution, which is based on morpholo-
vision and image processing techniques, have been pro- gical operations and genetic algorithms, is reported to
posed for detection of surface defects of products in be near 91%. Jia et al.25 proposed another solution for
industrial and commercial applications. A group of the detecting the defects on steel sheet surface. Their real-
proposed methods is based on statistical characteristics. time solution employs support vector machines and is
For example, properties of statistical histogram are capable of detecting the defects with an accuracy of
used to diagnose the surface defects of steel strips.1 94.4%. In another approach, Lee et al.,26 by employing
Among other statistical methods, gray level co- a neural network and the energy and entropy character-
occurrence matrix2,3 and morphological operations4–7 istics calculated using adaptive wavelet packet expan-
have been used for quality inspection of surfaces. sion, designed a system for detection and classification
Recently, a method based on multivariate image analy- of various defects on steel sheet strips. Their solution
sis, which is based on multivariate statistical analysis, performs with a reported classification accuracy of
has been proposed for detecting the metal surface 99% and is capable of detecting several different defects
defects.8 In statistical methods, usually the existence or such as scratches.
absence of a defect is determined and no further infor- Surface defect detection is also used in other indus-
mation is given about the location and status of the tries such as textile and is not limited to metal indus-
defect. try.27 For example, Kumar and Shen28 proposed a
On the other hand, there is another category of system that can detect and classify fabric defections
methods which are based on frequency domain analysis using support vector machines and principal compo-
such as Fourier analysis9,10 and Gabor transform.11–13 nent analysis. They claimed that their solution can also
There is also a subset of frequency-based methods, be used in ceramic industry. Another proposed solution
which uses the wavelet analysis to identify the defects by Elbehiery et al.29 employed simple image processing
806 Measurement and Control 52(7-8)

method, the curvature of the laser beam is mapped to

the size of the dipping or buckling in millimeters in
order to give a better understanding of the inspection
result to the human supervisor. As can be seen in
Figure 2, the physical structure of our proposed system
could be installed on the existing production line so
there is no need to apply any modifications in the pro-
duction line.
The proposed method consists of four main compo-
nents as follows: (1) camera calibration for eliminating
the distortion in the images taken by camera, (2) locat-
ing the laser beam in the image, (3) localization of the
curvatures of laser beam on the surface of sandwich
panel, and (4) perform a mapping from the curvatures
to the size of dipping or buckling in millimeters. In the
following subsections, first the physical structure of the
proposed system is described, and then each compo-
nent of the proposed method is briefly explained.

Mechanical structure of proposed system. To prevent the

Figure 2. Top: The complete designed system, which consist of vibrations of the production line disturbing the vision
the camera and the laser, after installing it on the ceiling at the component of our system, and also for the purpose of
production line. Bottom: The design of camera station with two easy installation and reduction of occupied space, the
degrees of freedom of movement eliminates the installation camera and laser are installed on the ceiling. The dis-
error using oval bolt. tance between the camera and the laser is determined
based on several factors including the desirable accu-
racy of defect detection, the distance of the production
techniques such as edge detection and morphological line to the ceiling, the initial angle of the camera, the
operations to detect the existence of defects on ceramic field of view and resolution of the camera, and the
tile surfaces. thickness of the panels. Small angle of view in the cam-
era results in a more accurate detection of the surface
Method defects.
The designed structure has six degrees of freedom
Detection of abnormal dipping and buckling on the for free adjustments. This feature helps to adjust the
sandwich panel surface camera alignment with the production line with a high
precision. Figure 3 shows the degrees of freedom of the
Since the aim is to scan the surface of the panel and
camera structure. In addition, creating perpendicular
extract the three-dimensional information for detecting
gaps, instead of holes, in adjacent surfaces of the struc-
the dipping and buckling, using the portable laser scan-
ture results in the ability of structure adjustment.
ners and RGBD cameras such as Microsoft Kinect are
In addition, the adjustability of the camera’s angle
proposed as a first option to solve this problem.
of view helps to provide the possibility of positioning a
However, due to the fact that sandwich panel on the
laser beam in the middle of the camera’s field of view
production line conveyor is moving during inspection,
according to the thickness of the manufactured sand-
our proposed system employs a simpler and more
wich panels (Figure 4).
affordable technology that checks only one horizontal
band of the panel at any moment and gives the result
of inspection as a surface image to the production line Camera calibration. Due to distortion of the lenses used
supervisor after the full scan of sandwich panel. In this in the cameras, the images produced by most of the
system, a combination of a simple RGB camera (with a existing cameras have abnormal curvatures and distor-
4-mm focal length) and a linear laser (MW Red5) is tions. This means that a straight line in the real world
used (Figure 2, bottom). The advantage of using RGB is seen as a curved line at the camera’s viewpoint. Since
camera as opposed to gray-scale camera in the pro- our proposed method works based on the curvature
posed setup is that locating laser beams, as described in measurement of the laser beams radiated on the surface
the following sections, can be facilitated by incorporat- of sandwich panel, this problem should be solved
ing color information. The linear beam of laser is verti- before any further processing of the images. To solve
cally radiated on the surface of sandwich panel from the problem, first the employed camera is calibrated to
above, and the image of the laser beam which is in the determine the camera’s intrinsic parameters. For this
field of view of the camera is analyzed for any curva- purpose, various images from different angles are taken
ture and irregularity. Finally, by using a calibration from a standard object with a chessboard pattern, and
Torkzadeh and Toosizadeh 807

Figure 5. All extrinsic and intrinsic parameters of the camera

are determined in the process of camera calibration based on
the images of a standard object.

Figure 3. The degrees of freedom of the camera structure.

Figure 6. A sample image taken by camera with low exposure

time.

then the camera parameters are determined using cali-

bration methods30,31 (Figure 5). Using the specified
parameters, which include two parameters (k1, k2) of
the distortion model of camera, the distortion of the
images produced by the camera is corrected according
to the following equations

x ¼ xd ð1 þ k1 r2 þ k2 r4 Þ ð1Þ
y ¼ yd ð1 þ k1 r2 þ k2 r4 Þ ð2Þ

where ðx; yÞ is the corrected location of the pixel,

ðxd ; yd Þ is the original image, and r could be computed
using the following equation

r2 ¼ x2d þ y2d ð3Þ

Locating the laser beam in the image. After the initial cali-
bration of the camera, the system should be able to
detect the laser beam in the images captured by the
camera. The beam of the laser is very bright, and there-
fore, in the normal working mode of the camera it
causes the full saturation of the light-sensitive sensor of
the camera. By setting the light exposure time of the
Figure 4. Top: The angle of 52° between the camera direction optical sensor of the camera, we make the image so
and the laser line for the panel of 7 cm thickness. Bottom: The dim such that the sensor is no longer saturated. As
angle of 59° between the camera direction and the laser line for shown in Figure 6, increasing exposure value (accord-
the panel of 20 cm thickness. ing to equation (4)) and consequently decreasing
808 Measurement and Control 52(7-8)

could create significant challenges for the system

(Figure 7).
Considering the various sandwich panels of different
thicknesses, since the laser beams in the image are
located in a bounded area in the image, it is not neces-
sary to process the whole image for finding the location
of the laser beams and therefore only a pre-specified
portion of the image which includes the laser beams
can be analyzed. Applying a threshold ðT1 Þ on the
image, considering the high intensity of laser light, and
finding the approximate location of the laser beam is
the easiest way to locate the laser beams. To determine
the appropriate value of this threshold, we average the
red components of those pixels whose red values are at
least twice the values of their green and blue compo-
nents. The threshold is then applied to the red channel
of the given frame to binarize the image. This operation
together with applying the morphological opening
operation can properly approximate the location of the
laser beams in the image.
We intend to design a very high precision method
that can detect the slightest curvatures of the laser
beams to detect the dipping and buckling on the surface
of sandwich panel. Therefore, the method used in our
proposed system goes beyond the simple solution above
and tries to find the location of laser beam with a sub-
pixel precision in the image. Since, the thickness of the
laser beam in the image is larger than one pixel and the
vertical slices of the laser beams usually have local max-
imum, the proposed method attempts to obtain a con-
tinuous and sub-pixel precision for localizing the laser
beams by analyzing the intensity of the laser beam’s
neighborhood pixels. For this purpose, for each pixel in
the laser beam image (obtained from applying the
threshold T1 and morphology operations), we extract a
small neighborhood, fðxÞ, which is a vertical slice, per-
pendicular to the laser beam direction in that pixel and
attempt to fit a Gaussian function, yðxÞ, to it. The cen-
ter of this Gaussian function (which, due to the conti-
Figure 7. The laser beams from various viewpoints under nuity of the function, can be a decimal number and on
various lightening conditions. a sub-pixel scale) is considered as laser’s precise posi-
tions in that neighborhood. Looking at the images of
the laser beams, one would observe the Gaussian-like
exposure time, the extra amount of light in the environ- shape of the vertical slices of the laser beams, with the
ment and noise are significantly reduced while laser peak being at the center of such slices and tails fading
beam becomes brighter than the image out at their two ends. Using a Gaussian function, aside
from helping to identify the precise location of the laser
N2 beams, also helps to achieve sub-pixel accuracy. Thus,
EV ¼ log2 ð4Þ
t Gaussian function is an appropriate choice for locating
where N is the relative aperture (f-number) and t is the the laser beam in the image.
exposure time in seconds. In the optimal fitting of Gaussian function, since
It is worth mentioning that in an environment where gradient descent (or ascent) method is susceptible to be
the ambient light is high, considering the fact that the stuck in the local maxima, selecting a proper initial
camera functions in the visible light spectrum, physical guess for the parameters of Gaussian function could
cautions should be exercised to control the ambient significantly result in better fitting accuracy and speed.
light since the reflection of the ambient light from the In our proposed system, the initial value of Gaussian
surface of the sandwich panel that is being analyzed center ðmÞ for each section is the detected pixel in the
Torkzadeh and Toosizadeh 809

thresholding operation. In addition, the Gaussian range

ðAÞ and sigma ðsÞ are initialized using the light inten-
sity (i.e. the mean of RGB value) of that pixel and the
number of laser beam pixels in the examined section,
respectively.
In our method, instead of directly using the differ-
Figure 8. An example of the line segments extracted from a
ence between the Gaussian function (equation (5)) and
processed strip of one of the lasers beams. Every line segment is
the laser data as our loss function, we use the exponen-
shown surrounded by a red rectangular bounding box. In this
tial function with a negative sign (equation (6)). This image, for clarity, the overlapped rectangles (i.e. line segments)
negative sign is responsible for reducing the effect of have not been displayed.
high values of the difference in finding the optimal val-
ues of parameters
ðxmÞ2 At the end, the value of m indicates the location of
yðxÞ ¼ Ae 2s2 ð5Þ the laser beam in the analyzed section. We only need to
X 
ðxmÞ2
2 repeat the same process for each of the remaining sec-
½fðxÞAe 2s2 
E¼ e ð6Þ tions to find the precise location of laser beam.
x

By differentiating the objective function E with Detecting the laser beam curvatures. We can determine,
respect to unknown variables (m, A, s) and utilizing with a high precision, the location of the existing dip-
the gradient ascent, we can reach the desirable result ping and buckling on the surface of the panel by ana-
starting from the initial values lyzing the laser strip, given its precise location found on
 
∂E X 2 ðxmÞ2 the image in the previous step. For doing so, it is suffi-
¼ 2ðfðxÞ  yðxÞÞe ðfðxÞyðxÞÞ  2s2 ð7Þ cient to first extract several line segments with short
∂A x
  lengths (the values from 10 to 100 pixels are appropri-
ðxmÞ2
X ðfðxÞyðxÞÞ2  2 ate) from the laser strip using the sliding window
∂E 2Aðx  mÞðfðxÞ  yðxÞÞe 2s
¼ method where each window has a pre-specified overlap
∂m x
s2
with the previous window (Figure 8). In the second
ð8Þ step, by using linear regression, a straight line is fitted
 ðxmÞ2

ðfðxÞyðxÞÞ2  on each extracted line segment independently. It is
∂E X 2Aðx  mÞ2 ðfðxÞ  yðxÞÞe 2s2
¼ expected that the gradients of the extracted line seg-
∂s x
s3 ments from laser strip follow a regular pattern on the
ð9Þ surface of a sandwich panel with no defects (i.e. any
anomaly in the values of gradients is probably an indi-
Choosing a constant value for the learning rate of cator of a defect). Therefore, in the third and final step,
the gradient ascent extends the fitting time. This prob- for each laser strip being analyzed independently, we
lem occurs, especially, when the data have long tails.
can precisely specify the defect locations of the panel
This situation relatively increases the total value of
by comparing the gradients of the segments with the
objective function, which decreases the contribution of
average of the gradients of all line segments extracted
other parts of data, and therefore causes a relatively
from that laser strip. Those line segments whose gradi-
slow progression to an often low quality solution. In
ents differ from the average gradients above a certain
order to solve the problem, the learning rate of the gra-
threshold ðT2 Þ would be marked as candidates of defect
dient ascent is increased and consequently the progres-
locations. Considering the movement of the sandwich
sion will be accelerated toward the goal initially.
panel on the production line conveyor, the laser beams
However, it bounces several times over the local maxi-
covers the whole surface of the sandwich panel as it
mum, when reaching near it (the overshoot phenom-
moves and therefore, according to our proposed
enon). A proper learning rate coefficient should have a
method, it is precisely scanned to detect any existing
high initial value, and overtime approaches to zero,
dipping and buckling points. Furthermore, to prevent
while the optimization approaches a solution. Hence,
false positives and negatives in our results, our method
every n iterations, the value of the learning rate coeffi-
further checks and confirms the existence and nonexis-
cient of the gradient descent, denoted by g in equation
tence of defects in a specific location not only by one,
(10), is multiplied by a positive number a, which is less
than 1, in order to prevent overshooting phenomenon but also by four laser beams (which can be of different
intensity and width) radiated on the surface of sand-
2 ∂E 3
wich panel. The advantage of utilizing four laser lines is
2 3 2 3 ∂s
snew sold 6 7 to ensure that the time lag between the speed of data
6 ∂E 7
4 mnew 5 ¼ 4 mold 5 þ g 6 7 ð10Þ collection and processing stages in our system, with the
6 ∂m 7
Anew Aold 4 5 conveyor speed, never results in any dipping or buck-
∂E ling being missed. It means that if one beam ignores a
∂A
810 Measurement and Control 52(7-8)

Figure 9. The standard panel for calibration of the measuring unit of the system which is located at the location of each of the
lasers beams.

defect due to various reasons such as speed, the other

ones would capture it.
In addition, our method could be extended to be
applicable on the sandwich panels without a flat sur-
face as well. It should be noted that if the surface of the
sandwich panels examined is not flat and have regular
patterns (for decorative purposes), for example, regular
up and down patterns of surface, the line segments
extracted may have very different gradients values. In
this case, first the values of gradients are clustered in
two or more clusters using a simple clustering algo-
rithm. Then, using the previous method for sandwich
panels with flat surfaces, in each cluster the segment
lines whose gradients are considered anomaly are
detected and therefore the corresponding location of
those line segments on the sandwich panel is considered
Figure 10. Detected points of checkered pattern that are used
as a defective location.
as training data to learn neural network parameters.

Measuring the size of dipping and buckling. So far, our pur-

posed system is able to locate the surface defects with a which determine the height of each image pixel located
high accuracy, but it could be extended to measure the in the area of the laser beam in millimeters. Using these
size (i.e. the height or depth) of dipping and buckling data, we train a feedforward artificial neural network,
points with a standard unit (in accordance with the which has two neurons (which represents pixel coordi-
International System of Units). To resolve this prob- nates in the image) in the input layer and one neuron
lem, a standard panel for calibration of the measure- (height in millimeters) in the output layer (Figure 11).
ment unit of system, similar to Figure 9, is designed Since, the feedforward artificial neural networks with a
which covers the whole laser beam in the image and hidden layer are used as a global estimator,32 a hidden
has a checkered pattern with precise and pre-specified layer with four neurons has also been considered in the
sizes (in millimeters) in vertical direction. As shown in architecture of our neural network which uses the sig-
Figure 10, the checkered pattern points are detected moid function as the activation function. In
Torkzadeh and Toosizadeh 811

Figure 11. The architecture of the neural network for

estimation of the height of each point of the laser beams.

deployment of our system, after locating the defective

points on the surface, the coordinates of those points
are given as input to the neural network and this net-
work predicts the height of laser beam at those points.
This feature further enables the human supervisor and
the quality control specialist to assess and ensure the
quality of productions and take any discretionary
actions (i.e. stopping the production line to inspect the
machinery if the size of defects exceeds a specified
threshold T3 ).

Experiments and results

Due to the discrete manufacturing of the sandwich
panels (e.g. 10 or 20 m in length), to continuously and
automatically utilize the proposed system in the pro-
duction line, a module is required to detect the begin-
ning and ending of a sandwich panel to provide the
Figure 12. The effect of the thinning process that takes
processing result of each panel separately to the super-
advantage of the Gaussian assumption of the laser beams. Top: A
visor. For this purpose, the thickness of the line is esti- sample panel surface with multiple defects. Middle:
mated using the average of the standard deviations of Corresponding obtained binary image of laser beam’s location
the vertical lines. Since, when a panel enters and leaves before applying the thinning process. Bottom: After applying the
the camera field of view, the thickness of the line would thinning process.
increase, the value of the standard deviation would also
become larger. Using this assumption, by passing the
average value of the standard deviations of the vertical strip which deviates from the fitted straight line, that is,
lines through a certain threshold ðT4 Þ, the beginning their distance to the fitted straight line is above a certain
and the end of a sandwich panel are easily determined. threshold ðT5 Þ, as anomalies and probable surface
Experiments were carried out on 35 sandwich panels defects. Figure 12 illustrates a sample panel surface with
with lengths 10 and 20 m and thicknesses 7 and 20 cm multiple defects and its corresponding binary image of
to assess the accuracy and efficiency of the proposed laser beams’ locations before and after applying the
system. A number of dippings and bucklings of various thinning process, taking advantage of the Gaussian
sizes and depths were manually added to the surface of assumption of the laser beams. As can be seen, the
each sandwich panel. Unfortunately, the unavailability defects (with heights/depths greater than T3 ) can be eas-
of a public benchmark dataset for this specific problem ily observed in the line segments obtained from laser
prevents us from being able to compare our results with beams.
the existing solutions. However, to demonstrate the As previously mentioned, this research entails five
effectiveness of our proposed method, we compare it different thresholds, T1 to T5 , which are used in various
with a baseline method. The baseline method is very steps of the proposed system. The value of T1 is speci-
similar to our proposed method and it only differs in fied adaptively based on each specific image and the
that it attempts to find the surface defects by fitting a values of other parameters are specified experimentally
straight line to the whole laser strip, instead of laser line and are stated in Table 1.
segments like performed in our proposed method. The The results obtained from our experiments show
baseline method considers those points on the laser that our proposed system achieves an accuracy of 93%
812 Measurement and Control 52(7-8)

Table 1. The estimated values for different thresholds in proposed system helps detecting the potential defects in
method. the surface of sandwich panels in real-time. Future
research in this field could focus on discovering other
Threshold T2 T3 T4 T5
production defects, including the misalignment of the
Value 5° 3 mm 2.5 3 mm
top and bottom sheets of sandwich panels, toward
designing a comprehensive automated visual inspection
system for sandwich panel production lines.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with
respect to the research, authorship, and/or publication of this
article.

Funding
The author(s) received no financial support for the research,
authorship, and/or publication of this article.

Figure 13. An example of how the system works. The image ORCID iD
on the right have been taken with the camera and analyzed, Saeed Toosizadeh https://orcid.org/0000-0002-3605-6095
according to our proposed method, to detect the surface
defects. In this image, some of the dipping and buckling points
have been annotated with numbers. The image on the left is the
References
output of our system as a heat map which represents the 1. Weiwei L, Yunhui Y, Jun L, et al. Automated on-line
surface of the analyzed sandwich panel. This image shows the fast detection for surface defect of steel strip based on
corresponding points with dipping and buckling (with heights/ multivariate discriminant function. In: Proceedings of the
depths above 3 mm) which have been accurately detected. second international symposium on intelligent information
technology application, Shanghai, China, 20–22 Decem-
ber 2008, vol. 2, pp. 493–497. New York: IEEE.
in detection of the dipping and buckling of sandwich 2. Chao Z, Desen Z and Li X. Textural defect detection
panels, whereas the base line method has an accuracy based on label co-occurrence matrix. J Huazhong Univ
Sci Technol (Nat Sci Ed) 2006; 34(6): 25–28.
of 76%. The experiments also show that only 1% of
3. Caleb P and Steuer M. Classification of surface defects
positive detections in our method were errors (i.e. 1%
on hot rolled steel using adaptive learning methods. In:
false positive), whereas the baseline method has 9% Proceedings of the fourth international conference on
errors. Figure 13 shows the system output in one of the knowledge-based intelligent engineering systems and allied
experiments (note that several dipping and buckling technologies, Brighton, 30 August–1 September 2000,
have been created manually on the sandwich panel). As vol. 1, pp. 103–108. New York: IEEE.
can be seen, the location of most of the surface defects 4. Mak KL, Peng P and Yiu K. Fabric defect detection
has been correctly detected. using morphological filters. Image Vis Comput 2009;
Furthermore, our experiments show that our pro- 27(10): 1585–1592.
posed method achieves an average error of 2.7 mm in 5. Cord A, Bach F and Jeulin D. Texture classification by
measuring the amount of each detected dipping and statistical learning from morphological image processing:
application to metallic surfaces. J Microsc 2010; 239(2):
buckling, which further proves the effectiveness of our
159–166.
proposed method. In terms of throughput, in average
6. Shi DH, Gang T, Yang SY, et al. Research on segmenta-
the system can perform 50 cm/s which, in our field tion and distribution features of small defects in precision
experimental setup, was well above the speed of the weldments with complex structure. NDT E Int 2007;
conveyor. 40(5): 397–404.
7. Alaknanda, Anand R and Kumar P. Flaw detection in
radiographic weldment images using morphological
Conclusion watershed segmentation technique. NDT E Int 2009;
In this paper, we designed and presented the implemen- 42(1): 2–8.
8. Zhou S, Liang D and Wei Y. Automatic detection of
tation of an inspection system for the surface of sand-
metal surface defects using multi-angle lighting multivari-
wich panels to detect abnormal dipping and buckling. ate image analysis. In: Proceedings of the international
It was shown that by using inexpensive equipments and conference on information and automation, Ningbo,
without human intervention, the proposed method can China, 1–3 August 2016, pp. 1588–1593. New York:
precisely detect most of the dipping and buckling on the IEEE.
surface of sandwich panels. Considering the relatively 9. Chan CH and Pang G. Fabric defect detection by Four-
high cost of the sandwich panel manufacturing, the ier analysis. In: Proceedings of the industry applications
Torkzadeh and Toosizadeh 813

conference on thirty-fourth IAS annual meeting, Phoenix, case of inspection system for specular surfaces. Sensors
AZ, 3–7 October 1999, vol. 3, pp. 1743–1750. New York: 2017; 17(7): 1466.
IEEE. 22. Cui Z, Lu W and Liu J. Real-time industrial vision system
10. Wang F and Zuo B. Detection of surface cutting defect for automatic product surface inspection. In: Proceedings
on magnet using Fourier image reconstruction. J Cent of the 8th international conference on information manage-
South Univ 2016; 23(5): 1123–1131. ment and engineering, Istanbul, Turkey, 2–5 November
11. Raheja JL, Kumar S and Chaudhary A. Fabric defect 2016, pp. 93–97. New York: IEEE.
detection based on GLCM and Gabor filter: a compari- 23. Jiuliang Z, Weiwei L, Feng Y, et al. Research on surface
son. Optik 2013; 124(23): 6469–6474. quality evaluation system of steel strip based on com-
12. Choi D, Jeon YJ, Yun JP, et al. An algorithm for detect- puter vision. In: Proceedings of the third international
ing seam cracks in steel plates. World Acad Sci Eng Tech- symposium on intelligent information technology applica-
nol 2012; 6: 1456–1459. tion, Shanghai, China, 21–22 November 2009, pp. 32–35.
13. Jeon YJ, Choi DC, Yun JP, et al. Detection of scratch New York: IEEE.
defects on slab surface. In: Proceedings of the 11th inter- 24. Zheng H, Kong LX and Nahavandi S. Automatic inspec-
national conference on control, automation and systems, tion of metallic surface defects using genetic algorithms. J
Gyeonggi-do, South Korea, 26–29 October 2011, pp. Mater Process Technol 2002; 125–126: 427–433.
1275–1279. New York: IEEE. 25. Jia H, Murphey YL, Shi J, et al. An intelligent real-time
14. Tsai DM and Hsiao B. Automatic surface inspection vision system for surface defect detection. In: Proceedings
using wavelet reconstruction. Pattern Recogn 2001; 34(6): of the 17th international conference on pattern recognition,
1285–1305. Cambridge, 26 August 2004, vol. 3, pp. 239–242.
15. Jeon YJ, Yun JP, Choi DC, et al. Defect detection algo- New York: IEEE.
rithm for corner cracks in steel billet using discrete wave- 26. Lee C, Choi CH, Choi JY, et al. Feature extraction algo-
let transform. In: Proceedings of the 2009 ICCAS-SICE, rithm based on adaptive wavelet packet for surface defect
Fukuoka, Japan, 18–21 August 2009, pp. 2769–2773. classification. In: Proceedings of the international confer-
New York: IEEE. ence on image processing, Lausanne, 19 September 1996,
16. Jeon YJ, Choi DC, Lee SJ, et al. Defect detection for cor- vol. 2, pp. 673–676. New York: IEEE.
ner cracks in steel billets using a wavelet reconstruction 27. Chu HH and Wang ZY. A vision-based system for post-
method. JOSA A 2014; 31(2): 227–237. welding quality measurement and defect detection. Int J
17. Yun JP, Choi SH, Jeon YJ, et al. Detection of line defects Adv Manuf Technol 2016; 86(9–12): 3007–3014.
in steel billets using undecimated wavelet transform. In: 28. Kumar A and Shen HC. Texture inspection for defects
Proceedings of the international conference on control, using neural networks and support vector machines.
automation and systems, Seoul, South Korea, 14–17 Octo- In: Proceedings of the international conference on
ber 2008, pp. 1725–1728. New York: IEEE. image processing, Rochester, NY, 22–25 September 2002,
18. Zheng H, Jiang B and Lu H. An adaptive neural-fuzzy Vol. 3, p. III. New York: IEEE.
inference system (ANFIS) for detection of bruises on Chi- 29. Elbehiery H, Hefnawy A and Elewa M. Surface defects
nese bayberry (myrica rubra) based on fractal dimension detection for ceramic tiles using image processing and
and RGB intensity color. J Food Eng 2011; 104(4): 663– morphological techniques. In: Proceedings of the third
667. world enformatika conference, Istanbul, Turkey, 27–29
19. Yazdchi M, Yazdi M and Mahyari AG. Steel surface April 2005.
defect detection using texture segmentation based on mul- 30. Zhang Z. A flexible new technique for camera calibration.
tifractal dimension. In: Proceedings of the international IEEE T Pattern Anal Mach Intell 2000; 22: 1330–1334.
conference on digital image processing, Bangkok, Thai- 31. Elbehiery H, Hefnawy A and Elewa M. Surface defects
land, 7–9 March 2009, pp. 346–350. New York: IEEE. detection for ceramic tiles using image processing and
20. Blackledge J and Dubovitskiy D. A surface inspection morphological techniques. In: Proceedings of the third
machine vision system that includes fractal texture analy- world enformatika conference, Istanbul, Turkey, 27–29
sis. Int Soc Adv Sci Technol 2008; 3: 76–89. April 2005.
21. Azorin-Lopez J, Fuster-Guillo A, Saval-Calvo M, et al. 32. Hornik K. Approximation capabilities of multilayer feed-
A novel active imaging model to design visual systems: a forward networks. Neural Netw 1991; 4(2): 251–257.