ISIJ International, Vol.

57 (2017),
ISIJ International,
No. 6 Vol. 57 (2017), No. 6, pp. 1045–1053

Detection of Pinholes in Steel Slabs Using Gabor Filter

Combination and Morphological Features

Doo-chul CHOI,1) Yong-Ju JEON,1) Seung Hun KIM,1) Seokbae MOON,1) Jong Pil YUN2) and Sang Woo KIM1,3)*

1) Department of Electrical Engineering, Pohang University of Science and Technology (POSTECH), Pohang, 37673 South
Korea. 2) Aircraft System Technology Group, Korea Institute of Industrial Technology (KITECH), Daegu, 42994 South
Korea. 3) Department of Creative IT Engineering, Pohang University of Science and Technology (POSTECH), Pohang,
37673 South Korea.
(Received on April 6, 2016; accepted on March 9, 2017)

In this paper, an algorithm for detecting pinholes was proposed. Pinholes are very tiny holes generated
in the surfaces of scarfed slabs. To detect various sizes of pinholes, the Gabor filter combination was
applied. A new image segmentation method was proposed in pre-processing. To reduce the number of
pseudo-defects, dual thresholding method was used. We have increased the performance of classification
by adding new morphological features to general texture based features. To evaluate the performance of
proposed algorithm, an images obtained in real production line were used.

KEY WORDS: quality control; machine vision; surface inspection; defect detection; Gabor filter.

In this paper, we focus on the detection of defects that

1. Introduction
exist in the surface of steel slabs. Steel slabs are semi-man-
Product inspection to guarantee the quality of products is ufactured products. The slabs tested in this study are used
an important element in the steel-making process. Although to produce steel sheets for automobiles. Because these steel
most elements of the steel-making process are currently sheets must meet high quality specifications, any defects in
automated, the inspection of steel products is still manually the steel slabs can degrade the quality of the end product.
operated.1) Manual inspection methods that show a high per- Therefore, defect detection should be performed to increase
formance are human resource and time consuming, and also the quality of the steel sheets and reduce manufacturing
inefficient. Furthermore, because the inspectors examine the costs. In particular, we address a defect called a pinhole,
product according to different criteria, manual inspection which is a very diminutive hole in the surface of scarfed
cannot guarantee reliability and accuracy. For example, the slabs. Scarfing is a process in which defective areas in the
reliability of the manually-operated quality control process surface of ingots or semi-finished products are burned out
for textiles is low and the defect detection rate by highly to a depth of about 2 mm so that the product is suitable
trained inspectors is only about 70%.2) Therefore, it is essen- for subsequent rolling or forging. Pinholes are generated
tial to develope an automated inspection system that yields by inclusions that are injected in the process of melting,
very reliable and accurate results. solidification, and casting. Because the surface of slabs is
Vision-based inspection systems are widely used for not uniform, and the size of a pinhole is relatively small, it is
controlling the quality of products in various manufac- difficult to distinguish pinholes from the surface. Therefore,
turing industries, such as that of LCD panels,3) fabric,4) many noise components, as well pinholes, are detected in a
food,5) and float glass.6) In the steel manufacturing industry, scarfed slab surface. To overcome the difficulties involved
various vision-based defect inspection systems have been in pinhole detection, a new vision-based defect detection
introduced.7,8) An automatic defect detection system using method for steel slab surfaces was developed in this study.
Gabor filters optimized by a univariate dynamic encoding In order to detect pinholes of various sizes, a combination
algorithm for searches has been developed for detecting method in which two Gabor filters are applied according to
cracks in raw steel blocks.9) A real-time vision-based defect the size of the pinhole in an image was designed.
inspection system for coiled steel bars for high-speed appli- The Gabor filter as applied in visual inspection has been
cations has been proposed.10) Although these systems were extensively studied. Since the Fourier coefficients are deter-
developed for steel products, it is difficult to apply them mined by the entire image, Fourier analysis cannot localize
directly to pinhole detection in steel slabs, because they defective areas in the spatial domain. To overcome this
are optimized for a specific steel type, defect type, lighting limitation, the windowed Fourier transform was introduced.
condition, and so on. If the window function is Gaussian, the windowed Fourier
transform becomes the well-known Gabor transform. Since
* Corresponding author: E-mail: swkim@postech.edu Turner11) and Bovik et al.12) first introduced Gabor filters
DOI: http://dx.doi.org/10.2355/isijinternational.ISIJINT-2016-160 for the purpose of texture analysis, the Gabor filter has

1045 © 2017 ISIJ

 ISIJ International, Vol. 57 (2017), No. 6

been extensively studied in the field of visual inspection. system. To develop an algorithm for detecting defects in
Kumar and Pang13) developed a defect detection algorithm the surface of steel products, it is very important to acquire
for fabric using real Gabor functions. Later, in their study,14) high-quality images of the surface. Therefore, to obtain an
the same authors used Gabor wavelet features to classify image of the surface, four camera/lighting modules were set
fabric defects. They also developed a defect detection sys- up on each side of a slab to capture images of its upper and
tem using only the imaginary part of the Gabor function as lower sides, so that eight images, four per side, are acquired
an edge detector. Bodnarova et al.15) proposed applying a from the frame grabbers. Four dual-core PCs operate the
Fisher cost function to select a subset of Gabor functions detection algorithm for two neighboring images, one image
based on the mean and standard deviation of the template per core. The surface images and detection results are trans-
feature images to perform textile flaw detection. Tsai et mitted to the server/monitoring system. The information of
al.16) used a single Gabor filter for multiple texture seg- the defect’s size and location is displayed on the monitor
mentation using a systematic optimization algorithm. Later in real time, and the images of the steel slabs and detection
Tsai and Lin17) developed a fast defect detection algorithm results are stored on the server.
for textured surfaces using 1D Gabor filters. Further, Tsai It is important to acquire the uniform-quality images of
and Wu18) also used Gabor filter selection so that the filter the slabs surfaces. Because it is difficult to achieve uniform
response energy of the normal texture was close to zero. illumination of the slab surface, a line-scan camera was
Yun et al.9) proposed an algorithm using a Gabor filter for used instead of an area-scan camera, which has disadvan-
detecting cracks in raw steel block. The parameters of the tages, such as blurring, distortion, lighting problems, and
Gabor filter were optimized by using an univariate dynamic the restriction of the inspection range.20) Line-scan cameras
encoding algorithm for searches (uDEAS). scan one line each time the slab is moved 0.25 mm along
In this paper, a new approach using Gabor filters is a direction. Therefore, 40 000–60 000 lines are obtained
introduced. In order to detect various sizes of defects in from a 10–15 m long steel slab. One frame image from the
the surfaces of steel products, filter combination method is frame grabber consists of 1 000 lines and each line consists
developed. Gabor filter combination method combines two of 4 096 pixels. The vertical resolution of one frame image
Gabor filtered-image based on the responses of two single is about 0.25 mm/pixel. Typically, images of a slab are
Gabor filters that are designed to detect various sizes of composed of 40–60 frames with a resolution of 4 096 pix-
defects.19) We have proposed dual thresholding method to els × 1 000 pixels. To obtain an image of the surface, the
reduce the number of defect candidates. To verify the per- lighting is directed such that it illuminates the slab surface
formance of the proposed method, it was evaluated using at an angle of 45 degrees. Moreover, we used a horizontal
artificial pinhole defects. array of LEDs as the lighting source to provide uniform
Using the proposed Gabor filter combination, a defect brightness of one scan line. One frame image of a slab is
detection algorithm for pinholes, which are very fine holes shown in Fig. 2.
in the surfaces of slabs, was developed. A Gabor filter com-
bination method is required because the sizes of the pinholes
vary between 1 mm and 5 mm. The experimental results
show that the Gabor filter is suitable for detecting pinholes.

2.  System Configuration and Image Analysis
2.1.  System Configuration
We developed an automatic inspection system (Fig.  1)
that consists of four components: camera/lighting modules,
Fig. 2. One frame image of a scarfed slab with an enlarged pin-
frame grabbers, dual-core PCs, and a server/monitoring hole. (Online version in color.)

Fig. 1. Automated inspection system.

© 2017 ISIJ 1046

 ISIJ International, Vol. 57 (2017), No. 6

Fig. 4. Artificial pinholes of different size.

3.  Gabor Function and Filter Combination
3.1.  Gabor Function
Fig. 3. Difference sizes of pinholes. The Gabor function has been widely used in image
processing, in particular, for texture analysis and defect
detection, owing to its ability to extract characteristics in a
specific scale and orientation. Since Gabor21) proposed the
2.2.  Image Analysis 1-D Gabor function in 1946, many studies in which it was
In this study, the detection targets are pinholes that utilized were reported. The 2-D Gabor function was pro-
exist in the surface of scarfed slabs. As mentioned, steel posed by Daugman, based on our understanding of the scale
slabs are half-finished products from which steel sheets for and orientation-selective properties of neurons in the brain’s
automobiles are produced. These steel sheets should meet visual cortex, and was mathematically completed in.22)
high-quality requirements; otherwise, defects in the surface The 2-D Gabor function is a sinusoidal plane wave of
of slabs degrade the quality of the end products. Steel slabs specific frequency and orientation with a Gaussian enve-
used to produce steel sheets are passed through a scarfing lope. It consists of a complex-valued sinusoidal function
process. Defect detection in steel slabs should be performed and a Gaussian function. The frequency spectrum of the
to guarantee the quality of the steel sheets and to decrease Gabor function is the Gaussian function shifted by a given
manufacturing costs. frequency and orientation.23) The 2-D Gabor function is
A pinhole is a tiny circular defect that exists in the surface
of slabs after the scarfing process. The diameter of a pinhole 1  1  x2 y2 
f ( x, y ) = exp  −  2 + 2   exp ( 2π juo x ) ... (1)
is typically 1–5 mm. Pinholes are formed in the surface of 2πσ xσ y  2  σ x σ y  
slabs because of gas bubbles generated by impurities that
are injected during the steel making process. Because the The 2-D Gabor function is divided into real and imagi-
scarfed slab’s surface is not uniform and pinholes are very nary parts. The real part of the Gabor function is well known
small, it is difficult to detect them. Moreover, many noise as a blob detector, and the imaginary part is well known as
components are also present on a scarfed slab surface; an edge detector.24) In this study, because the intensity char-
hence, it is hard to identify pinholes, even with the naked acteristics of pinholes are similar to those of edge elements,
eye. the imaginary part of the Gabor function is used to detect
The feature of a defect in an image is determined by pinholes in the surface of the steel slab. The general equa-
two elements: its geometrical shape, and the angle between tion of the imaginary part of the Gabor function is given as
the lighting source and the camera. Because the surface is
illuminated by light directed at an angle of 45 degrees, a  1  x ′ 2  y′ 2  
1   
shaded region appears at the floor of a pinhole, whereas, g ( x, y ) = exp  −   +    sin ( 2π fx ′ )
because the light is projected and reflected at the upper edge 2πσ xσ y  2  σ x   σ y  
  
of a pinhole, a bright area appears in that region. Thus, the .......................................... (2)
image of a pinhole has a vertically bright and dark intensity
profile, as shown in Fig. 3. where
As mentioned above, pinholes are very small defects, and
x ′ = xcosθ + ysinθ
many noise components which have similar characteristics
to pinholes exist on the surface of slabs. However, a pinhole
y′ = − xsinθ + ycosθ
has a vertically bright and dark profile and its shape is cir-
cular, and thus, these are the major features that are used to The parameters of the 2-D Gabor function are f, θ, σx and
distinguish pinholes from noise components. We designed σy. Parameter f represents the frequency of the sinusoidal
a Gabor filter using a filter combination method based on function and θ denotes the orientation of the sinusoidal
these characteristics. After Gabor filtering, an adaptive function. σx and σy are the constants of the Gaussian enve-
thresholding method using a vertical standard deviation lopes along the x and y-axis, respectively.
profile is applied. To identify defects among the defect can-
didates, morphological features are extracted and an support 3.2.  Filter Combination
vector machine (SVM) is applied for classification. The parameters of the Gabor filter are determined by the
size and direction of a defect. Therefore, a single Gabor
filter cannot guarantee a good performance in the case of
defects with various sizes. In this paper, a Gabor filter
combination method is proposed to detect various sizes of

1047 © 2017 ISIJ

 ISIJ International, Vol. 57 (2017), No. 6

defects in steel surfaces. In the Gabor filter combination, a defects. Figure 4 shows artificial pinholes of different sizes.
convex combination of two Gabor filtered images is applied We compared a single Gabor filter and the filter combination
as follows. in the case of a pinhole. Different sizes of pinholes can exist
in one frame image, and therefore, a selection method is
G ( x, y ) = α ( x, y ) × G1 ( x, y ) + (1 − α ( x, y ) ) × G2 ( x, y ) ... (3)
ineffective. The resultant images of the single Gabor filters
To evaluate the performance of the proposed method, and the filter combination are shown in Fig. 5. The images
we compared three methods: combination, averaging, and shown in Figs. 5(a) and 5(b) are the responses of two Gabor
selection. First, the combination method is the proposed filters designed to detect small and large pinholes, respec-
method. Second, averaging the responses of Gabor-filtered tively. In Fig. 5(a), it can be seen that the response of a large
images is not expected to yield a good performance as pinhole is as high as that of a small pinhole; however, the
compared to that of a single Gabor filter with appropriate response area of the Gabor filter is small as compared with
parameters. Third, a problem arises selection occurs in case pinhole’s size. For this reason, this Gabor filter is designed
of various sizes defect in one frame image. In this study, to detect small pinholes. On the other hand, the response
we evaluated the Gabor filter combination method using for a small pinhole is low in the case of the second Gabor
pinholes of various sizes. filter, as shown in Fig. 5(b). These results show that the
We tested the proposed algorithm using artificial pinhole Gabor filter combination effectively detects small and large
pinholes, as shown in Fig. 5(c). High responses and proper
response areas are obtained.
To evaluate the performance of the Gabor filter combina-
tion in detail, we obtained the vertical and horizontal pro-
files, marked as dotted lines in Figs. 5(a), 5(b), and 5(c). The
vertical and horizontal profiles are shown in Fig. 6. Gabor
1 and Gabor 2 denote the Gabor filters for small and large
pinholes, respectively. In the profiles of the horizontal direc-
tion, as shown in Fig. 6(a), and vertical profiles, as shown in
Fig. 6(b), the profiles of Gabor 1 and the filter combination
are suitable for detecting pinholes. In the vertical profiles as
shown in Fig. 6(c), however, the filter combination shows a
better performance than Gabor 1 in terms of region detec-
tion. Because the profile width of the filter combination is
wider than that of Gabor 1, the filter combination can detect
a closer size of a defect compared to Gabor 1 when binariz-
ing with same value.
We evaluated the performance of the Gabor filter combi-
Fig. 5. Resultant images of (a) the Gabor filter designed to detect
nation method using pinholes. To evaluate the performance
small pinhole, (b) the Gabor filter designed detect large
pinhole, and (c) the filter combination. of the proposed method, artificial defects were used. The
results verify that the proposed method is effective for
detecting various sizes of pinholes.

4.  Proposed Algorithm
4.1.  Preposcessing
In this inspection system, four images of each side to the
slab were obtained. In the images of the left and right edge
of a slab, there exists a background region that has no gray-
level information. To reduce unnecessary computational
cost, the background region should be removed by using
a segmentation process. Owing to the lighting condition in
the real field, the gray-values of the background are not near
zero, as shown in the left region in Fig. 7(a). Therefore, the
left part in the vertical projection profile, as shown in the
blue profile (Fig. 7(b)), is not zero. These non-zero gray
values lead to segmentation failure. To solve this problem,
we focused on using line-scan cameras. A line-scan camera
scans one line of a slab’s surface at every movement of the
slab, whereas the background region captured by the cam-
era remains fixed. Therefore, the standard deviation along
the vertical direction should be low, as shown in the red
profile in Fig. 7(b). In the case of uniform surface with low
Fig. 6. Profiles of (a) horizontal, (b) vertical 1, and (c) vertical 2. gray values or when some particles exist in the background
(Online version in color.) region, the standard deviation becomes large. To take

© 2017 ISIJ 1048

 ISIJ International, Vol. 57 (2017), No. 6

Fig. 7. (a) Slab image with background; (b) vertical projection profile and vertical standard deviation profile; (c) product
of the two profiles. (Online version in color.)

advantage of both profiles, we applied the product form of Table 1. Parameters of the two Gabor filters.
two profiles, as shown in Fig. 7(c). The profile is defined as Filter f θ σx σy
the product of the vertical projection profile M(x) and verti-
g1(x,y) 1/8 π/2 4 8
cal standard deviation profile S(x) as follows.
g 2(x,y) 1/16 π/2 8 16
P ( x ) = M ( x ) S ( x ) .......................... (4)
where
1
M ( x) = ∑ I ( x, y ) ....................... (5)
N

N y =1

1
S ( x) = ∑ I 2 ( x, y ) − M 2 ( x, y ) ............ (6)
N

N y =1

x = 1,2,..., M, y = 1,2,...,N, and M and N are the horizontal

and vertical sizes of a slab’s image, respectively. The value
of P(x) is almost zero for the background region, whereas
its value for the slab’s surface is significantly higher; hence,
we can remove the background region using a simple thresh-
olding method.

4.2.  Gabor Filter Combination
The pinhole has a bright and dark intensity profile along
the vertical direction. Therefore, Gabor sine filters were
used based on the characteristic of a pinhole. As mentioned,
to detect various sizes of pinhole, a combination of two Fig. 8. (a) Response image of Gabor filter combination; (b) a 3-D
Gabor filters is applied. Because various sizes of pinholes profile of (a). (Online version in color.)
can exist in one frame image, a filter selection method is
ineffective, while averaging the responses of Gabor filtered
images provides a low performance. Gabor-filtered output
images were obtained by following a convolution operation. the Gabor-filtered image is determined by four parameters.
Therefore, the parameters should be selected on the basis of
Gi ( x, y ) = I seg ( x, y ) * gi ( x, y ) the diameter and orientation of the pinholes. The parameters
...... (7)
= ∑ m = 0 ∑ n = 0 gi ( x, y ) T ( x − m, y − n )
M −1 N −1 of the Gabor filters are listed in Table  1. Gi(x,y) are the
Gabor-filtered images using gi(x,y), and Iseg(x,y) is an image
after segmentation.
where i = 1, 2, and gi(x,y) are Gabor filters. The response of The response of the Gabor filter combination is obtained

1049 © 2017 ISIJ

 ISIJ International, Vol. 57 (2017), No. 6

by factor is melted steel powder. Scarfing is the process that

grinds the surface of a steel slab using gas cutting. In the
G ( x, y ) = α ( x, y ) × G1 ( x, y ) + (1 − α ( x, y ) ) × G2 ( x, y ) ... (8)
area in which the gas pressure is low, a powder of melted
where steel remains along the vertical direction as shown in Fig. 
9(a). Because the melted steel powder is prominent on the
G12 ( x, y ) surface, the intensity profile is that of a pinhole in reverse;
α ( x, y ) = ................. (9)
G12 ( x, y ) + G2 2 ( x, y ) hence, it is dark and bright along the vertical direction. The
melted steel powder appears continuously along the vertical
The image after Gabor filtering is shown in Fig. 8(a). direction. In the region where two powders of melted steel
are connected, the intensity profile is similar to that of a
4.3.  Defect Candidate Extraction pinhole. The second factor is stab marks caused by a roller
After Gabor filtering using the filter combination method, to which foreign substances adhered, as shown in Fig. 9(b).
the response of a pinhole in the Gabor-filtered image is high. Their intensity profile is similar to that of a pinhole. While
However, the responses of pseudo-defects are also high, as the shape of a pinhole is circular, however, the shape of a
shown in Fig. 8(b). For the next process which distinguishes stab mark is not fixed. Moreover, the depths of stab marks
pinholes among high responses in an image, we have to are much shallower than those of pinholes. The third factor
perform a binarization process. is white spots as shown, in Fig. 9(c).
There are three kinds of pseudo-defects that make it dif- The important characteristic of these factors is that they
ficult to detect pinholes in the surface of a slab. The first appear continuously along the vertical direction. Using this

Fig. 9. Pseudo-defects in the surface of the slab (a) powders of Fig. 11. (a) Response image after diving the Sg (x,y); (b) a 3-D
melted steel, (b) stab marks, and (c) white spots. profile of (a). (Online version in color.)

Fig. 10. A vertical standard deviation profile and the profile after smoothing. (Online version in color.)

© 2017 ISIJ 1050

 ISIJ International, Vol. 57 (2017), No. 6

Fig. 13. (a) Defect candidate; (b) defect model; (c) orientation and
distance.

threshold is proposed. Because the intensity characteristics

of images are diverse, a fixed-value threshold is ineffective.
In general, an adaptive threshold value Tada is widely used
in defect detection of steel products. When a pinhole exists
in the surface, however, the response of the pinhole is much
higher than that of others, as shown in Fig.  12(a). There-
fore, the performance of a maximum-value-based threshold
Fig. 12. (a) 3-D profiles of an image with a pinhole; (b) 3-D pro-
Tmax is better than that of an adaptive threshold, while the
files of a defectfree image. (Online version in color.) performance of an adaptive threshold is better than that of
a maximum-value-based threshold in defect-free images,
as shown in Fig. 12(b). In this paper, we propose a new
dual thresholding method, where the higher of these two
characteristic, we applied a vertical standard deviation pro- threshold values is taken. The threshold value T is obtained
file before the thresholding process. Because pseudo-defects as follows.
appear in a group along the vertical direction, the standard
T = max {Tada , Tmax } ........................ (11)
deviation in the vertical direction is high. A vertical standard
deviation profile calculated using the Gabor filtered image where
G(x,y) is shown in Fig.  10. In the region of the pseudo-
Tada = mean {Gs ( x, y )} + β ada × std {Gs ( x, y )} .... (12)
defects group, as shown in this figure, the values of the
vertical standard deviation profile are high and spread along
Tmax = β max × std {Gs ( x, y )} .................. (13)
the horizontal direction. In the defect region, however, the
values of the profiles are narrow in the horizontal direction. βada and βmax are empirically determined coefficients.
Therefore, we performed a smoothing process to maximize To remove small noise components, size filtering was
the effect of the pseudo-defects group. Using this smoothed applied to a binarized image. Blobs having an area of less
version of the vertical standard deviation profile, the Gabor than sixteen pixels were removed. After size filtering, we
filtered image is divided as identified defect candidates C(x,y) (Fig.  13(a)) from the
original image based on the size and position information
G ( x, y ) of the binarized image.
Gs ( x, y ) = ........................ (10)
S g ( x, y )
4.4.  Feature Extraction
where Sg(x,y) is the smoothed version of the vertical stan- After the defect candidates extraction, feature extraction
dard deviation profile of G(x,y). The image Gs(x,y) is shown was performed to classify defects and pseudo-defects among
in Fig.  11(a). After this process, many noise components the defect candidates. In this paper, we have applied a sta-
reduced, as shown in Fig. 11(b). tistics based on the histogram, gray Level co-occurrence
To extract defect candidates, binarization process is matrix (GLCM), and geometric moment invariant.25)
conducted. In this paper, a new dual thresholding method To improve the performance of the proposed algorithm,
using an adaptive threshold and a maximum-value-based we propose new features based on the morphological shape

1051 © 2017 ISIJ

 ISIJ International, Vol. 57 (2017), No. 6

of a pinhole. Pinholes have intensity profiles that are verti- Table 2. Experimental results using SVM.
cally bright and dark; hence, it is essential to determine the Training Validation
locations of the bright and dark regions. In the gray image
defect pesudo-defect defect pseudo-defect
of a defect candidate, we note five features: the score of the
mask, orientation, relative distance, cohesiveness, and gray- General
120/120 5 408/5 527 42/52 2 306/2 369
features
value distance. To calculate these features, the gray image
should be divided into bright and dark regions. First, a (100.00%) (97.85%) (80.77%) (97.34%)
normalized defect candidate image Cnor(x,y) is calculated as + morpological
120/120 5 461/5 527 50/52 2 338/2 369
features
Cnor ( x, y ) = C ( c, y ) − mean {I seg ( x, y )} ......... (14)
(100.00%) (98.81%) (96.15%) (98.69%)
A bright region is defined as a region where the gray val-
ues of a normalized defect candidate image are greater than
half its maximum value. Likewise, a dark region is defined
as a region where the gray values of a normalized defect
( xhigh − xlow ) + ( yhigh − ylow )
2 2
candidate image are lower than half its minimum value.
dr = ........... (19)
Using Cnor(x,y), the bright and dark regions are obtained as
WH
Cnor ( x, y ) , if Cnor ( x, y ) > 0.5max {Cnor ( x, y )}
Cb ( x, y ) = 
4.4.4. Cohesiveness
 0, else
Cohesiveness refers to the level of cohesiveness of the
........................................ (15)
bright and dark region.

Cnor ( x, y ) , if Cnor ( x, y ) < 0.5min {Cnor ( x, y )} W −1 H −1 Cb ( x, y )

C d ( x, y ) =  ∑ x =0 ∑ y =0
( xhigh − xlow ) + ( yhigh − ylow )
2 2
 0, else ... (20)
........................................ (16) Chigh = W −1 H −1
∑ x = 0 ∑ y = 0 Cb ( x, y )
where Cb(x,y) and Cd(x,y) are bright and dark regions,
respectively.
W −1 H −1 C d ( x, y )
4.4.1. Score ∑ x =0 ∑ y =0
( xhigh − xlow ) + ( yhigh − ylow )
2 2
... (21)
To evaluate the distribution of the bright and dark Clow = W −1 H −1
regions, a mask K(x,y) was designed, as shown in Fig. 13(b). ∑ x = 0 ∑ y = 0 C d ( x, y )
The size of the mask was the same as that of the defect
candidate. The score was calculated as
4.4.5. Gray Value Distance
1
∑ ∑ Cnor ( x, y ) K ( x, y ) ... (17)
W −1 H −1
Score = The gray value distance is the difference in the aver-
W × H x =0 y =0 age gray values of the bright and dark region. Owing to
where, W and H are the width and height of a defect can- the lighting condition of the inspection system, a pinhole
didate, respectively. has a bright and dark intensity profile, because the light is
reflected in the upper part of a pinhole and cannot reach the
4.4.2. Orientation bottom of a pinhole. Therefore, the difference in the gray-
The intensity feature of a pinhole is bright and dark, values of the reflected region and bottom are high in the
vertically. Therefore, the angle between the bright and dark case of a pinhole.
regions can provide good information for classifying pin-
holes and pseudo-defects. The orientation is defined in Fig.
5.  Experimental Results
13(c) and is determined as
The evaluation of the performance of the proposed algo-
 yhigh − ylow  rithm is presented in this section. In previous study,26) the
θ = arctan   .................... (18) performance of algorithm was evaluated using sub-images
 xhigh − xlow 
because only small data sets of slab images. Moreover, the
where (xhigh, yhigh) and (xlow, ylow) are the centers of gravity performance of the algorithm was not satisfactory.
of Cb(x,y) and Cd(x,y), respectively. We tested the proposed algorithm using slab images
directly obtained from a product line. As shown in Fig. 1,
4.4.3. Relative Distance the slab images were obtained using a line scan camera and
As shown in Fig. 13(c), this feature defines the distance their resolution was 4 096 by 1 000 pixels. The diameter
between the bright and dark region and can be used effec- of a pinhole is typically 1–5 mm and the resolutions in
tively for distinguishing between pinholes and pseudo- the horizontal and vertical axes are 0.2 mm/pixel and 0.25
defects. Owing to the various sizes of the candidates, the mm/pixel, respectively. Therefore, the minimum size of a
relative distance is obtained by dividing the size of the pinhole detected in this experiment was 4 pixels × 4 pixels.
candidate as The proposed algorithm was tested on 958 slab images, of
which 165 images included pinholes and 803 were defect-

© 2017 ISIJ 1052

 ISIJ International, Vol. 57 (2017), No. 6

free. The experimental results are summarized in Table  2.

REFERENCES
In this study, we employed an SVM27) to classify defects
and pseudo-defects. 8 068 candidates were extracted from 1) X. Li, S. K. Tso, X. P. Guan and Q. Huang: IEEE Trans. Ind. Elec-
tron., 53 (2006), 1927.
the 968 slab images, of which 172 candidates were pinholes. 2) C. H. Chan and G. K. Pang: IEEE Trans. Ind. Appl., 36 (2000), 1267.
The average number of candidates per frame image was 3) O. Jong-Hwan, Y. Byoung-Ju, K. Se-Yun and P. Kil-Houm: IEICE
Trans. Fundam. Electron. Commun. Comput. Sci., 91 (2008), 1400.
approximately 8.3. The proposed algorithm in which the 4) H. Y. Ngan, G. K. Pang, S. Yung and M. K. Ng: Pattern Recognit.,
size of the pinhole is considered and local information and 38 (2005), 559.
dual thresholding are applied was proved to be effective. 5) O. Kleynen, V. Leemans and M. F. Destain: J. Food Eng., 69 (2005),
41.
To improve the classification performance, we defined new 6) X. Peng, Y. Chen, W. Yu, Z. Zhou and G. Sun: Int. J. Adv. Manuf.
morphological features. By adding these proposed morpho- Technol., 39 (2008), 1180.
7) J. P. Yun, Y. J. Jeon, D. C. Choi and S. W. Kim: J. Opt. Soc. Am. A,
logical features to general texture features, the classification 29 (2012), 797.
performance was improved, as summarized in Table 2. Con- 8) Y. J. Jeon, D. C. Choi, S. J. Lee, J. P. Yun and S. W. Kim: J. Opt.
Soc. Am. A, 31 (2014), 227.
sequently, proposed algorithm using morphological features 9) J. P. Yun, S. Choi, J. W. Kim and S. W. Kim: NDT&E Int., 42 (2009),
was effective to detect pinholes in the steel slab images. 389.
10) J. P. Yun, S. Choi, B. Seo and S. W. Kim: Opt. Eng., 47 (2008),
077204.
6.  Conclusion 11) M. R. Turner: Biol. Cybern., 55 (1986), 71.
12) M. Clark, A. C. Bovik and W. S. Geisler: Pattern Recognit. Lett., 6
(1987), 261.
In this paper, we proposed a new defect detection algo- 13) A. Kumar and G. Pang: Opt. Eng., 39 (2000), 3176.
rithm using the Gabor filter combination. Because pinholes 14) A. Kumar and G. K. Pang: IEEE Trans. Ind. Appl., 38 (2002), 425.
are very small in size, it is not easy to detect them to slab 15) A. Bodnarova, M. Bennamoun and S. Latham: Pattern Recognit., 35
(2002), 2973.
images. To overcome these difficulties, in our method we 16) D. M. Tsai, S. K. Wu and M. C. Chen: Image Vis. Comput., 19
first applied a Gabor filter combination to extract defect (2001), 299.
17) D. M. Tsai and C. P. Lin: Int. J. Adv. Manuf. Technol., 20 (2002),
candidates. Next, to analyze the pinhole in detail, we defined 664.
morphological features: the score of the mask, orientation, 18) D. M. Tsa and S. K. Wu: Int. J. Adv. Manuf. Technol., 16 (2000),
relative distance, cohesiveness, and gray-value distance. 474.
19) D. C. Choi, Y. J. Jeon, S. J. Lee, J. P. Yun and S. W. Kim: Appl.
Images of steel slab surfaces were used evaluate the perfor- Opt., 53 (2014), 4865.
mance of the proposed algorithm. The experimental results 20) C. S. Cho, B. M. Chung and M. J. Park: IEEE Trans. Ind. Electron.,
52 (2005), 1073.
verified that the proposed defect detection algorithm effec- 21) D. Gabor: J. Inst. Electr. Eng. III Radio Commun. Eng., 93 (1946),
tively detects pinholes in steel slab surfaces. 429.
22) J. G. Daugman: J. Opt. Soc. Am. A, 2 (1985), 1160.
23) J. Weng, J. Zhong and C. Hu: Appl. Opt., 48 (2009), 3308.
Acknowledgments 24) D. P. Casasent and J. S. Smokelin: Opt. Eng., 33 (1994), 2255.
This research was supported by the MSIP (Ministry of 25) R. C. Gonzalez: Digital Image Processing, Pearson Education India,
Delhi, (2009), 1.
Science, ICT and Future Planning), Korea, under the “ICT 26) D. C. Choi, Y. J. Jeon, J. P. Yun and S. W. Kim: Appl. Opt., 50
Consilience Creative Program” (IITP-2015-R0346-15-1007) (2011), 5122.
27) C. Cortes and V. Vapnik: Mach. Learn., 20 (1995), 273.
supervised by the IITP (Institute for Information & Com-
munications Technology Promotion).

1053 © 2017 ISIJ