Advances in Engineering Software 115 (2018) 17–25

Contents lists available at ScienceDirect

Advances in Engineering Software

journal homepage: www.elsevier.com/locate/advengsoft

Intelligent assessment of subsurface cracks in optical glass generated

in mechanical grinding process
Yong Jie Zhao, Yun Hui Yan, Ke Chen Song, Hao Nan Li∗
School of Mechanical Engineering and Automation, Northeastern University, Shenyang 110819, China

a r t i c l e i n f o a b s t r a c t

Article history: Grinding process of optical glass has been reported to be related with the creation of subsurface cracks.
Received 20 May 2017 However, for the time being, most measurement methods have been depended on human operations. In
Revised 31 July 2017
this paper, an intelligent assessment method based on image processing technique is proposed. Grind-
Accepted 13 August 2017
ing trials proved that, the proposed method can accurately (with the biggest relative error of 3.53%) and
Available online 17 August 2017
quickly (nearly 1.6 seconds per micrographs) recognize and measure the subsurface crack depths. More
Keywords: importantly, the proposed method has good robustness to different-sized images. Besides, the method
Automatic detection does not require any input parameters or any adjustment of thresholds, therefore the method does not
Optical glass require any prior knowledge of either mechanical grinding process or brittle material behaviors relat-
Image processing ing with subsurface cracks. Based on above, the proposed method is expected to be meaningful to both
Subsurface damage metrology equipment companies and optical glass manufacturers.
Grinding
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction
As a key kind of optical materials, manufacturing of optical
glass is considered as the foundation of many industries includ-
ing optics, telecommunication, electronics, and mechanotronics [1].
The mechanical grinding process of optical glass, however, has
been reported closely related with the generation of subsurface
cracks [2–4] due to the unstable grain-workpiece interactions in-
duced by stochastic grinding wheel topography [5–7], although
more advanced grinding tools with uniformly-protruded mono-
layer abrasives were produced [8–10].
With this, substantial studies relating with the quantification
of subsurface crack depths have been reported so far. Although
many measurement methods have been provided, nearly all of
them have been suggested based on the cross-section micrographs
observed by an optical microscope, where many human involve-
ment/operations have been required, because, as seen in Fig. 1,
• In most measurement cases, the ground surfaces of optical glass
have been obliquely placed, therefore the subsurface damage
(SSD) depth should be the tilted distance labeled by “SSD” in
Fig. 1;
• Due to the polishing operation during the creation of sample
cross sections, many noise points have been introduced (see
“noise” in Fig. 1);
∗
Corresponding author.
E-mail address: lhnlwfb@163.com (H.N. Li).
• Material flaws might also interfere with the SSD measurement.
However, it should also bear in mind that, human measurement
of SSDs might be only feasible when a small quantity of operations
are required. Therefore a more advanced automated measurement
of SSDs is highly in-demand.
On the other hand, intelligent recognition and measurement
based on image processing technique has been recently found as a
powerful tool in many fields containing mechanical, civil and agri-
culture engineering. This technique can be used for not only detec-
tion of a certain macro/micro feature, but also quantification of a
certain process to provide a more in-depth understanding.
In mechanical and manufacturing engineering, both Zatočilová
et al. [11] and Du et al. [12] proposed an image-based method
to on-line measure both the dimension and the axis straightness
of the hot forgings. Experiments for both unheated and heated
samples showed the method accuracy of up to 97%. Zhao et al.
[13] used an image-processing-based method to detect the defects
in the cold rolling process with the consideration of the influence
of industrial liquids and surface textures. The defect detection ac-
curacy achieved 91% although some defects were covered by in-
dustrial liquids. A similar method was also reported in the mon-
itoring of the brittle material grinding process [14], enabling the
automation of observations of process details. The method was
also employed to in-process and in-situ monitor ground surface
morphologies [15], where the brittle and ductile regions can be
achieved during the machining process, providing the reference
of the process statuses. Besides, Gadelmawla [16] employed the
image-processing method to measure the screw thread features.

http://dx.doi.org/10.1016/j.advengsoft.2017.08.008
0965-9978/© 2017 Elsevier Ltd. All rights reserved.
18 Y.J. Zhao et al. / Advances in Engineering Software 115 (2018) 17–25
Nomenclature
i, j example pixel
Vi,j 3∗ 3 pixel matrix around pixel i and j
i L∗ L pixel region around pixel i
Ii grey value of pixel i
Oi grey value of pixel i after denoising
wij similarity between pixel i and j (see Eq. (2))
zi normalized coefficient (see Eq. (3))
h smoothing factor
μ randomly selecting three clustering centers
P ( μ| j ) degree of membership between the center μ and
the pixel j
drj grey value difference between the center μ and the
pixel j
N total pixel number in the image
Jfuz objective function during clustering
B cross-shaped structure matrix
I matrix containing the color information of the input
image
a a certain element in the matrix I
B shrinkage matrix
xmax ,
( ) coordinate of the deepest pixel of the detected sub-
ymax
surface cracks
Slength scale bar length (with the unit of μm)
Spixel scale bar length (with the unit of pixels)
The method followed the standard ISO metric thread plug gage and
the maximum difference between the standard and measured val-
ues was found to be smaller than ±5.4 μm.
In civil engineering, Valença et al. [17] and Dogan et al. [18] as-
sessed cracks on the concrete bridge by using images captured
by terrestrial laser scanning technology. The method can achieve
highly-localized results where the crack width, length and orienta-
tion can be accurately recognized and measured.
An interesting application of the image-based measurement in
agriculture engineering was reported in Ref. [19]. The authors used
image processing technology to monitor the automatic drying sys-
tem of rough rice. The system sent the acquired images of rough
rice to a computer for image processing so that the distribution of
stalks and the moisture content of rough rice could be obtained.
Besides, various image-based measurement methods were used
to measure plasma temperature [20], thin oil film thickness [21],
particle velocity [22], 3D deformation [23,24], bubble [25] and
multiphase flow [26] behaviors, showing good robustness, high de-
tection and measurement speed, and high accuracy.
Based on above, an image-based intelligent assessment method
is proposed in this paper to measure SSDs in optical glass gen-
erated in mechanical grinding process. The method aims to en-
able the automated recognition and quantification of SSDs, so that
Fig. 1. Cross section image showing the problems during the measurement of sub-
surface cracks.
the quick, accurate, and large quantities of measurements can be
performed by computers. The method therefore is expected to be
promising to facilitate the industrial manufacturing of optical glass.
2. Method details
The proposed methodology mainly includes three steps (see
Fig. 2): (1) detection of the subsurface cracks, (2) recognition and
reconstruction of the ground surface, and (3) calculation of the SSD
depths.
2.1. Detection of the subsurface cracks
The detection of the subsurface cracks is one of the most im-
portant steps in this study. It mainly includes four substeps: (1)
non-local means denoising, (2) fuzzy c-means clustering, (3) mor-
phological dilation of subsurface crack edges, and (4) morphologi-
cal erosion and reconstruction of subsurface crack edges.
2.1.1. Non-local means denoising
As illustrated in Fig. 1, both noise pixels and material flaws
would interfere with the detection and measurement of SSDs,
therefore the first substep is to denoise the input image. Here the
non-local means denoising [27] is used because, in comparison
with other denoising methods, this method was proved to be very
effective to denoise randomly-scattered noise pixels whilst retain-
ing all the edge details in the micrographs [27] so that subsurface
cracks would not be influenced.
As seen in Fig. 3(a), the basic principle of this method is as fol-
lows: for a certain pixel i in the image, each 3∗ 3 pixel matrix (with
the center pixel j, symbolised by “matrix Vj ” in Fig. 3a) within the
neighboring range L∗ L pixel region (denoted as the region i ) is
compared with the pixel i neighboring matrix (symbolized by “ma-
trix Vi ” in Fig. 3a).
Assuming the original grey value of pixel i is Ii , then the de-
noised grey value of pixel i (denoted as Oi ) would be

Oi = wi j · I j (1)
j∈i
where wij refers to the similarity between pixel i and j, and could
be calculated by
   2
1 Vi − V j 
wi j = exp − (2)
zi h2
where zi and h are separately the normalized coefficient (calcu-
lated based on Eq. (3)) and the smoothing factor (set to be 2 ac-
cording to Ref. [27]), and the operator  •  denotes the norm.
   
 Vi − V j 2
zi = exp − (3)
h2
j
Fig. 3(b) and (c) separately show the original and the denoised
micrographs containing subsurface cracks.
2.1.2. Fuzzy c-means clustering
For most micrographs, they generally have three different kinds
of regions: (1) background (called dark region), (2) crack region
(the target region in this study), and (3) bulk material region. Here
the fuzzy c-means clustering [28] is employed in this study to di-
vide the whole image into three regions.
The basic procedures of the method are: (1) randomly selecting
three clustering centers (denoted as μ) for the three regions, and
then calculating the Degree of Membership (DM) of each pixel (as-
suming the pixel j) in reference to the centers P(μ|j) according to
 Y.J. Zhao et al. / Advances in Engineering Software 115 (2018) 17–25 19

Fig. 2. The flowchart of the proposed methodology.

Fig. 3. (a) Illustration of the Non-Local Means Denoising (NLMD), (b) the original and (c) denoised micrographs containing subsurface cracks.

2.1.3. Morphological dilation of subsurface crack edges

After the recognition of the three regions, the morphological
dilation [29] is employed to allocate which region is the subsur-
face crack region. As illustrated in Fig. 5(a), for a certain pixel (see
pixel i in Fig. 5a), the Boolean intersection between a cross-shaped
structure matrix B and pixel i is performed, after which the grey
value of pixel i is set to be the biggest value within the matrix B.
As illustrated in Fig. 5(b), the edges of a target region would be
thickened (see the purple pixels in Fig. 5b). Fig. 5(d) presents the
dilated image.
Based on above, the subsurface crack edges can be recognized
by subtracting the original image from the dilated image, because
only the dilated edges would be left as seen in Fig. 5(c). Fig. 5(e)
Fig. 4. The micrograph after the fuzzy c-means clustering.
shows the example result after the subtraction, where it could ob-
serve that, all the subsurface crack edges are highlighted in com-
parison with the original image.
 
1
dr j
P (μ\ j ) =    (4) 2.1.4. Morphological erosion and reconstruction of subsurface crack
3 1
r=1 dr j
edges
It might also observe from Fig. 6(a) that, except for subsurface
where drj is the grey value difference between the center μ and cracks, there are still several material flaws recognized by mor-
the pixel j, i.e. phological dilation performed in Section 2.1.3. Given that these
flaw sizes are larger than 1∗ 1 pixel thus denoising methods might
d r j =  μ r − j 2 ( r = 1 , 2 , 3 ) (5) not detect them, the morphological erosion-reconstruction method
where [30] is used here.
N Morphological erosion-reconstruction is a method that can
P (μ\ j ) · j
2
eliminate the target boundary points so that the boundary can be
μr = j=1 (6)
P (μ\ j )
N 2 shrunken. This method can be mathematically expressed as
j=1

with the total pixel number in the image N. (I  B ) = {a|(B )a ⊆ I} (8)

The final segmentation result of the three target regions can
where I refers to the matrix containing the color information of
be achieved, when the following objective function Jfuz reaches its
the input image, a is a certain element in the matrix I, and B is
minimum value. Fig. 4 shows the processed result.
the shrinkage matrix and is equal to [1; 1; 1; 1; 1; 1; 1; 1; 1; 1]

3 
N because the material flaws are distributed in a random direction.
P (μ\ j ) · dr j
2
J f uz = (7) Fig. 6(b) shows the result after the morphological erosion-
r=1 j=1 reconstruction, where it could find that, only the subsurface cracks
20 Y.J. Zhao et al. / Advances in Engineering Software 115 (2018) 17–25

Fig. 5. Basic principle of morphological dilation for (a) a certain pixel and (b) a certain region, (c) illustration of the result by subtracting the original image from the dilated
image, and image examples of (d) the dilated image and (e) the final result where the subsurface crack edges are highlighted.

Fig. 6. Image results (a) before and (b) after the morphological erosion-reconstruction is performed.

are properly kept while nearly all the material flaws are detected
and removed.
Here it needs to explain that, although a white layer is pre-
sented on the top of the ground surface (see Fig. 6), it would not
influence the detection of the subsurface cracks, because subsur-
face damage refers to the deepest crack depth therefore only the
deepest pixel within the detected subsurface crack region would
affect the method accuracy.
2.2. Recognition and reconstruction of the ground specimen surface
servation of the ground sample cross sections, therefore the SSD
depth would be an inclined distance from the deepest subsurface
crack tip to the ground surface, rather than the vertical distance
(see Fig. 1).
2.3. Calculation of the SSD depths
When both the subsurface cracks and the ground surface are
detected, the SSD depth can be easily obtained based on the point-
to-line distance in the analytic geometry, i.e.

As stated in the Introduction, even though subsurface cracks |tanβ · xmax − ymax + b|
can be recognized by the above steps, it is also important to recog-
SSD = (with the unit of pixels) (9)
(tanβ )2 + (−1 )2
nize the tilted ground surface, because in most cases, the sample
placement can not be ideally horizontal. where (xmax ,ymax ) is the coordinate of the deepest pixel of the de-
It can also be observed from Fig. 7 that, the detailed recognition tected subsurface cracks.
process is: (1) the Canny edge detection [31] is firstly performed to Please note that, all the above calculation is based on the unit
highlight all the potential edges including the ground surface (see of pixels. Therefore to obtain SSD depths in the length unit, unit
Fig. 7a); (2) the top pixel of the highlighted edges are kept while conversion needs to perform. Assuming the scale bar length is
others are removed (see Fig. 7b); (3) the mathematical express of Slength (with the unit of μm) and is expressed by using Spixel pix-
the ground surface can be obtained by fitting the left pixels with els, the realistic subsurface crack depth SSDlength would be
the basic fitting function form of y = tanβ • x + b (see Fig. 7c).
Here the slope is calculated because the proposed method con-
Slength
SSDlength = · SSD. (10)
siders the tilted sample placement during optical microscope ob- S pixel
 Y.J. Zhao et al. / Advances in Engineering Software 115 (2018) 17–25 21

Fig. 7. The image results after (a) the Canny edge detection is performed, (b) the top pixel of the highlighted edges are kept while others are removed, and (c) the ground
surface fitted by using the left pixels with the basic fitting function form of y = tanβ · x + b.

Fig. 8. The experimental setup of the grinding trials.

Table 1 is induced by the lateral cracks, rather than the median cracks,
Grinding parameters.
meaning the proposed method can also be applicable to various
Parameters Values subsurface damage situations. Besides, Fig. 9(d) and (e) as well as
Depth of cut, ap (μm) 0.1, 0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, Fig. 9(g) and (h) indicate that, even though the subsurface cracks
4.0, 5.5, 6.0, 8.0, 10.0, 12.0, 14.0, induced by the median cracks are shallow, the proposed method
15.0 can also successfully recognize and accurately measure the crack
Grinding wheel speed, vs (m/s) 12.5, 14, 15.5, 17, 18.5, 20, 21.5, 23, depths.
24.5, 26
Workpiece feed rate, vw (mm/min) 2, 5, 10, 15, 20, 25, 30, 40, 50, 60
4.2. Method speed

3. Experimental validation of the methodology
As seen in Fig. 8, grinding trials of optical glass blocks (25 mm ∗
25 mm ∗ 5 mm) have been performed by using a grinder machine
tool in this section to validate the method. The optical glass sam-
ples have been mounted on the jig, which then has been mounted
on the dynamometer. The used grinding parameters are given in
Table 1, and the SSD observation after the grinding trials have been
the same as Ref. [3,4].
4. Validation and discussion
4.1. Method accuracy
To validate the method accuracy, the SSD depths automatically-
assessed by the proposed method are compared with the
manually-measured ones, and the results are presented in Fig. 9.
It can find that, among all the four sets of comparisons, the
biggest relative error is only 3.53% while the smallest one even
reaches 1.09%, indicating the method accuracy to a large extent.
It should note that, in Fig. 9(a) and (b), the subsurface damage
Because the aim of the proposed method is to automatically de-
tect and measure the subsurface crack depths so that when a large
quantity of measurement tasks are required measurement speed
would not be the problem, the proposed method speed is there-
fore evaluated in this section.
Fig. 10 shows the original images and the images processed by
the proposed method. For all the twelve micrographs, the total
subsurface crack recognition and measurement time is 19.23 sec-
onds when a PC is used (InterCore i7-4500U 3.2GHz). This means
the average speed of the proposed method is around 1.6 seconds
per micrographs, which would be hard for manual operators to be
faster, even for the skilled operators. It can also expect that, the
proposed method would largely facilitate the measurement effi-
ciency, and be specially interesting for metrology equipment com-
panies.
4.3. Method robustness to different-sized images
Except for method accuracy and speed, the method robustness
to different-sized images is also evaluated in this section, because
different optical microscopes usually output images with different
22 Y.J. Zhao et al. / Advances in Engineering Software 115 (2018) 17–25

Fig. 9. SSD depth comparison between the automatically-assessed and manually-measured ones.

resolutions. As seen in Fig. 11, although different-sized images are
employed as the input images, the proposed method can success-
fully recognize and accurately measure nearly all the images, prov-
ing the wide method applicability to a large extent.
5. Failure case analysis
Although the proposed method can successfully recognize the
subsurface cracks for most micrographs, there are still some failure
cases and three typical examples are presented in Fig. 12.
For the failure case 1 (see Fig. 12a1 for the input image and
Fig. 12a2 for the failure detection result), the failure detection of
subsurface cracks might probably be because of the overexposure
during the microscope observation. In fact, it might be also hard
for human raw eyes to clearly recognize the subsurface cracks in
Fig. 12(a1), especially the crack edges and tips. Given that this fail-
ure can be avoided by adjusting the microscope observation illumi-
nation, this kind of failure detections is believed to be acceptable.
For the failure case 2 (see Fig. 12b1 for the input image and
Fig. 12b2 for the failure detection result), the scratch marks during
the cross section polishing process interfere with the subsurface
 Y.J. Zhao et al. / Advances in Engineering Software 115 (2018) 17–25 23

crack detection. Although the cracks are successfully recognized,

the scratch marks are also detected by the method, which would
be then mistakenly employed as the deepest cracks. Obviously, this
failure is not mainly induced by the method but the inappropriate
operations before the microscope observation.
For the failure case 3 (see Fig. 12c1 for the input image and
Fig. 12c2 for the failure detection result), optical performance inho-
mogeneity is induced by the large-scaled fracture during the cross
section polishing, therefore the bulk material region presents two
kinds of completely different color features (one tends to be dark
and the other tends to be white). Similarly, this failure is also prob-
ably induced by the inappropriate operation during the polishing
process, and the failure detection can be considered acceptable.

6. Conclusions

In this paper, an intelligent assessment method of subsurface

cracks in optical glass generated in mechanical grinding process is
proposed. The main findings of this paper might include:

• Experimental grinding trials proved that, the method can ac-

curately (with the biggest relative error of 3.53%) and quickly
(nearly 1.6 seconds per micrographs) recognize and measure
the subsurface crack depths;
• The method also has good robustness to different-sized images,
meaning the method can be applicable to most micrographs,
even though they are captured by different microscopes and
different in sizes/resolutions;
• The method does not require any input parameters or any ad-
Fig. 10. Original images and the images processed by the proposed method that justment of thresholds (the method does not have any embed-
are used to evaluate the method speed. ded threshold), because all the crack region segmentation pro-
cess would be based on the color information and morpholog-

Fig. 11. Method robustness to different-sized input images (i.e. images with different resolutions).
24 Y.J. Zhao et al. / Advances in Engineering Software 115 (2018) 17–25
Fig. 12. Failure case analysis: orignal images and failure results because of (a1, a2) overexposure during the observation, (b1, b2) scratch marks during the cross section
polishing process, and (c1, c2) optical performance inhomogeneity due to the large-scaled fracture during the cross section polishing process.
ical features in the original image. This also means that, the
method did not require any prior knowledge of either mechan-
ical grinding process or brittle material behaviors relating with
subsurface cracks;
Based on above, the proposed method is expected to be mean-
ing to both metrology equipment companies and optical glass
manufacturers.
Acknowledgments
The authors acknowledge the supports from both the National
Natural Science Foundation of China in undertaking this research
work under the grant reference number of 51374063 and the
Fundamental Research Funds for the Central Universitiesunder
the grant reference numbers of N140303008, N141008001 and
N150308001.
References
[1] Hartmann P. Optical glass. USA: SPIE; 2014.
[2] King RI, Hahn RS. Handbook of modern grinding technology. USA: Springer;
2012.
[3] Li HN, Yu TB, Da Zhu L, Wang WS. Evaluation of grinding-induced sub-
surface damage in optical glass BK7. J Mater Process Technol 2016;229:
785–794.
[4] Yu TB, Li HN, Wang WS. Experimental investigation on grinding characteris-
tics of optical glass BK7: with special emphasis on the effects of machining
parameters. Int J Adv Manuf Technol 2016;82:1405–19.
[5] Li HN, Yu TB, Wang ZX, Da Zhu L, Wang WS. Detailed modeling of cut-
ting forces in grinding process considering variable stages of grain-workpiece
micro interactions. International Journal of Mechanical Sciences 2017;126:
319–339.
[6] Li HN, Yu TB, Zhu LD, Wang WS. Analysis of loads on grinding wheel binder in
grinding process: insights from discontinuum-hypothesis-based grinding sim-
ulation. Int J Adv Manuf Technol 2015;78:1943–60.
[7] Li HN, Yu TB, Zhu LD, Wang WS. Modeling and simulation of grinding wheel by
discrete element method and experimental validation. Int J Adv Manuf Technol
2015;81(9-12):1921–38.
[8] Ding W, Linke B, Zhu Y, Li Z, Fu Y, Su H, et al. Review on monolayer
CBN superabrasive wheels for grinding metallic materials. Chin J Aeronaut
2017;30(1):109–34.
[9] Ding W, Zhang L, Li Z, Zhu Y, Su H, Xu J. Review on grinding-induced residual
stresses in metallic materials. Int J Adv Manuf Technol 2017;88(9):2939–68.
[10] Xi X, Ding W, Li Z, Xu J. High speed grinding of particulate reinforced titanium
matrix composites using a monolayer brazed cubic boron nitride wheel. Int J
Adv Manuf Technol 2017;90(5):1529–38.
[11] Zatočilová A, Paloušek D, Brandejs J. Image-based measurement of the
dimensions and of the axis straightness of hot forgings. Measurement
2016;94:254–64.
[12] Du Z, Wu Z, Yang J. 3D measuring and segmentation method for hot heavy
forging. Measurement 2016;85:43–53.
[13] Zhao YJ, Yan YH, Song KC. Vision-based automatic detection of steel surface
defects in the cold rolling process: considering the influence of industrial liq-
uids and surface textures. Int J Adv Manuf Technol 2016;90(5):1–14.
[14] Zhao YJ, Li HN, Song KC, Yan YH. In-situ and in-process monitoring of optical
glass grinding process based on image processing technique. Int J Adv Manuf
Technol 2017 in Press. doi:10.10 07/s0 0170-017- 0743- x.
[15] Zhao YJ, Li HN, Da Zhu L, Yan YH. Machined brittle material surface
in grinding: modeling, experimental validation, and image-processing-based
surface analysis. Int J Adv Manuf Technol 2017 In press. doi:10.1007/
s00170-017- 0712- 4.
[16] Gadelmawla ES. Computer vision algorithms for measurement and inspection
of external screw threads. Measurement 2017;100:36–49.
[17] Valença J, Puente I, Júlio E, González-Jorge H, Arias-Sánchez P. Assessment of
cracks on concrete bridges using image processing supported by laser scanning
survey. Constr Build Mater 2017;146:668–78.
[18] Dogan G, Arslan MH, Ceylan M. Concrete compressive strength detection using
image processing based new test method. Measurement 2017;109:137–48.
[19] Shei H-J, Lin C-S. An optical automatic measurement method for the mois-
ture content of rough rice using image processing techniques. Comput Electron
Agric 2012;85:134–9.
[20] Mahdavipour B, Hatami A, Salar Elahi A. Results on plasma temperature mea-
surement using an image processing technique. Results Phys 2016;6:1008–11.
[21] Bonilla Riaño A, Rodriguez IH, Bannwart AC, Rodriguez OMH. Film thickness
measurement in oil–water pipe flow using image processing technique. Exp
Therm Fluid Sci 2015;68:330–8.
[22] Mohmmed AO, Nasif MS, Al-Kayiem HH, Time RW. Measurements of transla-
tional slug velocity and slug length using an image processing technique. Flow
Meas Instrum 2016;50:112–20.
[23] Kim Y-N, Kim JS, Park GC, Cho HK. Measurement of sliding bubble behavior on
a horizontal heated tube using a stereoscopic image processing technique. Int
J Multiphase Flow 2017;94:156–72.
[24] Yu L, Pan B. Full-frame, high-speed 3D shape and deformation measurements
using stereo-digital image correlation and a single color high-speed camera.
Opt Lasers Eng 2017;95:17–25.
[25] Fu Y, Liu Y. Development of a robust image processing technique for bub-
bly flow measurement in a narrow rectangular channel. Int J Multiphase Flow
2016;84:217–28.
 Y.J. Zhao et al. / Advances in Engineering Software 115 (2018) 17–25 25

[26] Fu Y, Liu Y. Experimental study of bubbly flow using image processing tech-
niques. Nucl Eng Des 2016;310:570–9.
[27] Manjón JV, Coupé P, Martí-Bonmatí L, Collins DL, Robles M. Adaptive non-lo-
cal means denoising of MR images with spatially varying noise levels. J Magn
Reson Imaging 2010;31(1):192–203.
[28] Bezdek JC, Ehrlich R, Full W. FCM: the fuzzy c-means clustering algorithm.
Comput Geosci 1984;10(2-3):191–203.
[29] Jackway PT, Deriche M. Scale-space properties of the multiscale morphological
dilation-erosion. IEEE Trans Pattern Anal Mach Intell 1996;18(1):38–51.
[30] Vincent L. Morphological grayscale reconstruction in image analysis: applica-
tions and efficient algorithms. IEEE Trans Image Process 1993;2(2):176–201.
[31] Bao P, Zhang L, Wu X. Canny edge detection enhancement by scale multiplica-
tion. IEEE Trans Pattern Anal Mach Intell 2005;27(9):1485–90.