998 Vol. 35, No.

6 / June 2018 / Journal of the Optical Society of America A Research Article

Nondestructive, fast, and cost-effective image

processing method for roughness measurement
of randomly rough metallic surfaces
SAJJAD GHODRATI, SAEIDEH GORJI KANDI,* AND MOHSEN MOHSENI
Department of Polymer Engineering and Color Technology, Amirkabir University of Technology, Tehran 15875-4413, Iran
*Corresponding author: s.gorji@aut.ac.ir

Received 31 January 2018; revised 17 April 2018; accepted 29 April 2018; posted 30 April 2018 (Doc. ID 321105); published 22 May 2018

In recent years, various surface roughness measurement methods have been proposed as alternatives to the com-
monly used stylus profilometry, which is a low-speed, destructive, expensive but precise method. In this study, a
novel method, called “image profilometry,” has been introduced for nondestructive, fast, and low-cost surface
roughness measurement of randomly rough metallic samples based on image processing and machine vision. The
impacts of influential parameters such as image resolution and filtering approach for elimination of the long
wavelength surface undulations on the accuracy of the image profilometry results have been comprehensively
investigated. Ten surface roughness parameters were measured for the samples using both the stylus and image
profilometry. Based on the results, the best image resolution was 800 dpi, and the most practical filtering method
was Gaussian convolution
cutoff . In these conditions, the best and worst correlation coefficients (R 2 ) between
the stylus and image profilometry results were 0.9892 and 0.9313, respectively. Our results indicated that the
image profilometry predicted the stylus profilometry results with high accuracy. Consequently, it could be a
viable alternative to the stylus profilometry, particularly in online applications. © 2018 Optical Society of America
OCIS codes: (100.2000) Digital image processing; (100.2960) Image analysis; (120.6660) Surface measurements, roughness;
(240.5770) Roughness; (120.4290) Nondestructive testing; (150.3040) Industrial inspection.

https://doi.org/10.1364/JOSAA.35.000998

1. INTRODUCTION standard method (especially in industrial applications), it suf-

Surface roughness is a quantity that has a direct impact on a
product’s performance and its visual appearance [1–5].
Moreover, it is a key factor in scientific study of the phenomena
such as wetting, facial and interfacial interactions, friction, and
contact mechanics [6,7]. Surface roughness along with surface
chemical nature are two principal parameters that identify the
quality of a surface [8–10]. Thus, precise and quantitative
evaluation of surface roughness is of vital importance in quality
control of the finished products and also in description of the
surface phenomena for engineers and scientists [6,7,11].
Generally, surface roughness measurement methods can be
classified into groups of contact and noncontact approaches
[1,12]. The most common contact method is stylus profilom-
etry [3,13–16]. In the stylus profilometry, a tiny diamond tip
moves along a line on the surface, and, due to vertical displace-
ment of the tip on the bumps and the troughs of the rough
surface, an interpretation of the surface roughness is obtained,
which is called a “surface profile” [17,18]. Although surface
roughness measurement by the stylus profilometry is per-
formed with acceptable precision and is considered as the
fers from substantial drawbacks [2,19], as follows:
– Due to the tip’s direct contact with the surface of the
workpiece, occurrence of scratches and surface damage, particu-
larly in the case of the soft materials, is considerably prob-
able [3,20].
– Measurement precision depends highly on the tip dimen-
sions; as the diameter of the tip increases, detection of narrow
troughs and penetration of the tip into tiny valleys on the sur-
face will not be possible [2,18,21].
– The stylus profilometer devices are sensitive and
fragile, which makes working with them a stringent task.
Additionally, expertise and experience are required to choose
a suitable location and direction for recording of the surface
roughness [3,8].
– In the stylus profilometry, surface roughness measure-
ment is done based on a limited number of surface lines (pro-
files). Although roughness information from a single profile
may sufficiently describe topography, for example, of a compo-
nent’s surface produced through a plain and stable turning pro-
cess, it is not sufficient for topography description of the
recently introduced surfaces in which deterministic structures

1084-7529/18/060998-16 Journal © 2018 Optical Society of America

 Research Article Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A
have been used to control the surface’s functions [22–24].
Examples of such surfaces are nanostructured antireflection
coatings, cell adhesion improved structured surfaces, micro-
optical arrays, and surfaces with deterministic patterns for tri-
bological purposes [24]. Most of these cases require an areal
topography description to control or predict their functionality
[24]. Furthermore, areal topography measurement is a more
practical approach for problem diagnostics compared with a
single profile measurement [24]. If the number of the measured
profiles in stylus profilometry is increased to perform areal map-
ping, the measurement will be time consuming and expen-
sive [11,25].
– The stylus profilometry method cannot be used for
roughness measurement of the moving (in-process) surfaces
[26]. The contact nature of the stylus profilometry makes
the method low-speed and quite useless in automated produc-
tion lines. It is considered the most important drawback of this
method [2,8].
To overcome these restrictions, there is a growing tendency
toward employment of noncontact methods, especially in online
applications [2,18,20]. Numerous noncontact approaches, in-
cluding ultrasonic, pneumatic, and optic methods, have been
proposed for surface roughness measurement [26]. The optic
methods have shown promising potential for online surface
roughness measurement and characterization of surface topogra-
phy [18]. The most popular optic methods for surface roughness
measurement are laser speckle [27–31], light scattering [32],
confocal microscopy, both imaging and chromatic methods
[32], optical interferometry [12,32,33], or more state-of-the-
art methods such as coherence scanning interferometry, phase
shifting interferometry, focus variation method [32], and ma-
chine vision and image processing. All of these mentioned optical
methods have their own advantages and drawbacks. Via imaging
confocal microscopy (ICM), for example, a 3D surface topogra-
phy can be measured. This method possesses the highest lateral
resolution (up to 120 nm) in comparison with the common op-
tical methods. Moreover, ICM can measure topography of the
surfaces with high local surface slopes. One of the main limita-
tions of the method is its high sensitivity to the surfaces with
dissimilar materials. Surface topography measurement of the
transparent surfaces is also done with difficulty by ICM [32].
Roughness of the transparent surfaces can be easily measured
by chromatic confocal microscopy (CCM), if the transparent
layer thickness is not too low. Employing this method, either
surface profile or areal topography measurement could be pos-
sible. CCM is not sensitive to ambient light because, in this
method, a chromatic confocal probe scans the surface points;
thus, it has speed limitations. Additionally, CCM is sensitive
to high local surface slopes. The method has difficulty in rough-
ness measurement of dark surfaces (with low reflection) and also
shiny (highly reflective) surfaces. While spherical pits exist on the
sample surface, roughness measurement by CCM is imple-
mented erroneously. This source of error is called the “ghost foci”
[32]. It is possible to easily measure roughness of the surfaces
with steep local slopes by the focus variation method (FVM).
In addition, it is a fairly robust method for roughness measure-
ment of surfaces with optically dissimilar materials [32]. In
FVM, it is possible to simultaneously measure surface roughness
and surface form even on complex geometries [34,35]. Areal
999
topography measurement is a convenient practice in FVM.
One of the FVM’s restrictions is roughness measurement of
translucent materials. Furthermore, this method is not able to
correctly measure topography of smooth surfaces (Ra less than
15 nm) [32]. Phase shifting interferometry (PSI) is a powerful
and widespread method for evaluation of areal surface topogra-
phy. Analogous to other interferometry methods, this method
also works based on wave properties of light. PSI has nanometer
scale height resolution [32]. Generally, in the interferometry
methods, lateral resolution is quite low (about 500 nm)
[25,36]. Roughness measurement of surfaces with sharp local
height variations is done erroneous in PSI [25]. Moreover, pres-
ence of the optically dissimilar materials on the surface make the
roughness measurement unreliable, which is considered one of
the main limitations of the PSI [32]. Coherence scanning inter-
ferometry (CSI) is a new method for areal surface topography
measurement. Roughness measurement of transparent surfaces
is a practical task by this interferometry method [32]. CSI pos-
sesses a high height resolution (subnanometer scale) [25].
Furthermore, this method is faster in comparison with CCM
because it does not require point-to-point surface scanning
[37]. Compared with PSI, CSI is more flexible to different local
surface height variations. Previous studied showed that CSI over-
estimates roughness measurement results compared with other
methods [38,39]. Some other well-known error types in CSI
are ghost step and batwing (see [32,37,38] for more information).
Light scattering-based methods are categorized as the area-
integrating methods (versus coordinate-based methods, i.e., line
profilometry and areal topography) [40]. In these methods, in-
stead of investigating every surface point, an area of the surface
is sensed as a whole, and usually a single overall parameter is pre-
sented for surface topography characteristics. Because a large
amount of surface topography information is embodied in the
scattered light, light scattering is considered as a promising ap-
proach to be employed as a fast method, which can provide valu-
able information about the overall surface texture. However, these
methods suffer from theoretical and practical restrictions. For ex-
ample, the calculated Rq by light scattering methods differs from
that of the line profilometry and areal topography, and this
deviation is due to use of theoretical models in light scattering
methods. In such models, simplifying assumptions such as
Gaussian height distribution and low height variations are consid-
ered for the surface, which make it different from the reality [32].
As one of the light scattering methods, the laser speckle method
also possesses the aforementioned features [31]. Speckle tech-
niques benefit from some advantages, including cost effective
and simple hardware and fast data recording. However, it should
be mentioned that the speckle techniques, which use more than
one speckle field such as the speckle pattern illumination method,
require expensive and complex equipment [27,30]. Generally, sur-
face roughness evaluation by speckle method is performed
through two major approaches: model-based methods such as
the speckle contrast method [31], which have the aforementioned
light scattering restrictions, and the methods based on texture
analysis of speckle patterns such as GLCM [30], which is consid-
ered an indirect method of surface roughness measurement.
This, among machine vision and image processing, has
attracted researchers’ attention in recent years, owing to its
inexpensive devices and requirements, high precision, and high
 1000 Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A
performance speed [1,6,8,11,41–44]. Roughness evaluation by
this approach seems promising; however, this task is still a prac-
tical and computational challenge. Hladnik and Lazar at-
tempted to evaluate surface roughness of paper samples by
the image processing method because the paper surface rough-
ness is a key factor in properties such as ink absorption and
print quality. The authors used a confocal laser for image ac-
quisition and Gray-level co-occurrence matrix (GLCM) for tex-
ture analysis of the samples’ surface images. They observed
strong correlations between some parameters of GLCM such
as energy and roughness parameters obtained from confocal
laser profilometry such as R a and R q [45]. Moreover, recent
studies have revealed strong correlations between surface tex-
ture parameters obtained from GLCM and subjective and also
objective surface roughness of sandpaper samples [46,47].
Ghodrati et al. also reported that the image texture evaluation
methods, including fractal dimension, edge frequency, and
histogram analysis, are effective tools for surface roughness
assessment of sandpaper [47,48]. Accurate roughness measure-
ment of stone surfaces is also a major issue. As an example,
Stemp and Stemp tried to measure roughness of stone surfaces
for quantitative and nonsubjective archaeological study of an-
cient stone tools [49]. In this research, the authors benefited
from surface texture analysis of laser profilometer images
and showed that fractal dimension is an informative parameter
for quantitative archaeological studies of ancient stone tools
based on surface roughness. Surface roughness of polymeric
components and producing high-quality polymeric products
with respect to surface finish is of great importance in industries
such as aircraft and automotive manufacturing. In this context,
Sarma and coworkers evaluated surface roughness of machined-
reinforced polyester-based composite components by process-
ing the samples’ images [14]. The results of the study indicated
that the average gray-scale value of the images (G a ) suitably
correlates with the stylus profilometry’s R a . Furthermore,
Ghodrati et al. employed an image processing method based
on edge feature extraction for surface roughness quantification
of hot pressed polypropylene-based components [50]. The au-
thors attained promising results for prediction of stylus profilom-
etry’s roughness parameters. In addition to the aforementioned
cases, accurate measurement of surface roughness is a common
practice in so many other fields and materials, including ceramics
[51] and metals [52]. This, among machine vision and image
processing methods, have shown promising potential in rough-
ness measurement of metallic surfaces and has attracted research-
ers’ attention. Priya and Ramamoorthy recorded images of
ground mild steel specimens [2]. Then, they calculated five dif-
ferent parameters from Fourier transformed images of the spec-
imens. Training an artificial neural network, they found
significant relationships between the parameters and the stylus
profilometry’s arithmetic average height (R a ). To find a correla-
tion between the stylus profilometry results and image-based
data, Gadelmawla employed image texture (i.e., GLCM) analysis
[6]. He extracted four parameters from GLCMs of specimens’
images and determined their relationship with the stylus profil-
ometry’s R a . In a study, optical microscope images of machined
aluminum workpieces were captured [8]. Using pixel intensities,
three statistical parameters, i.e., gradient factor, average cycle
Research Article
texture, and average gray level, were introduced. In the next step,
the impact of the microscope magnification on the correlations
between the image-based results and the stylus profilometry re-
sults were investigated. Dhanasekar and Ramamoorthy recorded
the images of ground and milled metallic surfaces, which were
moving on a conveyer of a production line [20]. They trained an
artificial neural network by two statistical parameters obtained
from pixel intensities of the sample images (i.e., average and stan-
dard deviation of pixel intensities) and one parameter, which was
extracted from Fourier transformed images (i.e., major peak fre-
quency), as the input and the stylus profilometry roughness
parameters as the output. Using the trained artificial neural net-
work, the stylus roughness parameters were estimated for new
samples. Ghodrati et al. calculated fractal dimensions of abrasive
blasted steel plates as a surface roughness representative, using a
digital image processing approach [53]. Although the correlations
between the fractal dimensions and the stylus profilometry rough-
ness parameters were not high, the results of the research dem-
onstrated that the fractal dimension is an effective tool for surface
roughness assessment. Al-Kindi and Shirinzadeh prepared 20
metallic samples, which had been roughened by different meth-
ods such as grinding, milling, and lapping [1]. Considering three
different models, i.e., intensity-topography compatibility model,
intensity-topography compatibility model+moving average filter,
and diffuse model, employing image pixel intensities, they con-
structed a surface profile of the samples based on digital images.
Afterward, the authors developed relationships between the image
profile and the stylus profile roughness.
The majority of studies related to evaluation of surface
roughness by means of image processing has been conducted
on the machined, turned, or milled surfaces. Although some
machining process such as grinding and lapping produce a ran-
dom pattern [1,2,3], most of these methods create a periodic
pattern on the surface. Therefore, assessment of a single profile
across the sample surface can adequately characterize the whole
surface roughness [5]. There are a few researches on roughness
evaluation of the surfaces roughened by the methods such as
abrasive blasting. In the abrasive blasting method, the surface
is roughened by jetting the abrasive particles onto it. In such
methods, due to creation of random undulations on the sur-
face, assessment of a single profile across the sample surface can-
not be an adequate representative of the whole surface
roughness. In this research, it has been attempted to evaluate
the surface roughness of abrasive blasted mild steel samples by
means of machine vision and image processing. Because the
surface profile and the roughness parameters obtained from
the surface profile is still one of the most popular quantitative
interpretations of surface roughness, the authors have
attempted to propose a method that can produce a surface pro-
file as well as profile-based roughness parameters while resolv-
ing the stylus profilometry’s restrictions as much as possible.
It is worth noting that mild steel is among the most widely used
alloys in the construction industry and heavy infrastructures,
such as refineries, owing to its high mechanical strength and
low cost [54–56]. One of the most common mild steel surface
preparation methods, before the application of protective coat-
ings, is abrasive blasting [57,58]. Accurate measurement of sur-
face roughness created by abrasive blasting is important because
 Research Article Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A 1001

it strongly affects the performance of the protective coating

[57,59]. For this purpose, mild steel samples were roughened
through the abrasive blasting method by abrasive particles with
different sizes and nature. In the next step, the images of rough
samples were recorded. Employing the stylus profilometry,
quantitative surface roughness parameters, including ampli-
tude, frequency, and hybrid parameters, were measured and
compared with corresponding image-based roughness parame-
ters. The performance of four different filtering methods, in
elimination of long wavelength surface undulations (i.e., form Fig. 1. Schematic view of typical sample dimensions and real surface
and waviness) and improving accuracy of the image-based re- images of the samples with eight different roughness levels.
sults, was investigated. Because the surface roughness measure-
ment depends highly on the measuring scale, the impact of
image resolution on the image-based roughness parameters
was studied. Thus, the main purpose of the present paper is levels were obtained from the abrasive blasting, which, together
to develop a high-speed, low-cost, and precise image-processing with an untreated sample, created eight roughness levels. It
method (image profilometry) for accurate simulation of the should be noted that, for each roughness level, five replicas were
stylus profilometry method’s data. Because the latter is a prepared to ensure roughness measurement repeatability.
well-known and common roughness measurement method, Figure 1 shows schematic dimensions of the samples as well
it has therefore been attempted to unify the units of the image as real surface images of the samples with eight different rough-
profilometry’s results and the results of the stylus profilometry ness levels.
and also to produce surface roughness values as close as possible B. Roughness Parameters
to that of the stylus profilometry’s roughness values by control-
ling image acquisition conditions and equalizing the image pro- Surface roughness parameters obtained from the stylus profile
filometry and the stylus profilometry variables. are divided into three groups: amplitude, frequency, and hybrid
parameters [1,15]. To calculate the value of the parameters, a
reference line is required. The reference line can be the mean of
2. MATERIALS AND METHODS the profile points’ heights, or it can be the best fitted line on the
A. Sample Preparation profile points using the least-squares method [17,60] (in our
research, the latter definition was used for the reference line).
For preparation of samples with different roughness levels, ST-
The amplitude parameters are related to vertical distances of the
12 mild steel plates (Esfahan’s Mobarakeh Steel Company, Iran)
profile points from the reference line, e.g., arithmetical mean
with dimensions of 80 mm × 80 mm × 2 mm were used.
deviation of the profile (R a ), which is the most common surface
Chemical composition of the ST-12 alloy was determined us-
roughness parameter in the industries [4]. The frequency
ing energy dispersive spectroscopy (EDS) (Rontec, Germany).
parameters are related to repetition of the peaks and the valleys
The result is presented in Table 1.
in a specific profile length, such as average width of the profile
The plates were roughened through abrasive blasting by
elements (R Sm ). Finally, the hybrid parameters contain both the
seven different abrasive particles. The abrasive blasting condi-
amplitude and the frequency parameter properties, e.g., root
tions are listed in Table 2. Therefore, seven different roughness
mean square slopes of the profile (R Δq ) [15,61]. It is worth
mentioning that the total measured profile length is called
Table 1. Chemical Composition of Mild Steel (ST-12) “evaluation length.” According to the standard, the evaluation
length is divided to five sections (i.e., sampling length) for fur-
Element Wt. % Element Wt. % Element Wt. %
ther statistical analysis [62–64]. Detailed explanation for the
Fe 98.01 P 0.05 Mn 0.28 surface profile’s substantial concepts and the mathematical def-
C 0.16 Al 0.04 S 0.04
initions of 10 different roughness parameters, which were used
Si 0.30 O 1.12
in this study, is provided in Appendix A. For calculation of the

Table 2. Different Abrasive Blasting Conditions

Blasting Nozzle Angle Average Particle
Sample Pressure (MPa) Abrasive Type Nozzle Diameter (mm) Nozzle Distance (cm) (Relative to Substrate) Diameter (μm)
1 Raw surface, without abrasive blasting surface treatment
2 0.7 Sand 6 30 60° 200
3 0.7 Sand 6 30 60° 350
4 0.7 Sand 6 30 60° 480
5 0.7 Steel grit 8 50 90° 600
6 0.7 Steel grit 8 50 90° 800
7 0.7 Steel shot 8 50 90° 900
8 0.7 Steel grit 8 50 90° 1200
 1002 Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A
surface roughness parameters from the surface profile in the both
methods (i.e., stylus profilometry and image profilometry), the
reference line was determined as the best fitted line on the profile
data using the least-squares method. Subsequently, employing
the mathematical relationships of Table 5, surface roughness
parameters were calculated from the first parameter (R a ) up
to the last parameter (R Δq ) in turn.
C. Stylus Profilometry
For roughness measurement of the steel specimens, a stylus
profilometer (Perthometet M3, Mahr Company, Germany)
was employed. The specifications and the conditions of the sty-
lus profilometry are reported in Table 3. Measurements were
implemented based on the ISO 3274 standard. Roughness
measurement was repeated three times on the different loca-
tions and directions of each sample surface. Moreover, five steel
samples were used for each roughness level. Therefore, a total
of 15 measurements were done for each roughness level.
D. Machine Vision and Image Preprocessing
A digital scanner (V550, Epson Company, Japan) was used for
acquisition of the steel sample images. The main aims of employ-
ment of the scanner for image preparation were attaining uni-
form lighting on the sample surfaces and invariant imaging
distance and direction. For this purpose, a scanner could be a
more suitable and practical choice compared with a camera,
although a camera records the images faster than the scanner.
A machine vision system consisted of a personal computer
(Intel Processor Core i5 2.3 GHz, RAM memory 4 GB), and
the digital scanner was used in this study. Because the surface
roughness is a scale-dependent property, the steel sample images
were captured in eight different resolutions in order to investigate
the effect of resolution on the correlations between the image
processing results and the stylus profilometry measurements.
The resolutions were 96, 150, 200, 240, 300, 600, 800, and
1200 dots per inch (dpi). The images were recorded in gray scale
with 256 gray levels and saved in tagged image file format (TIFF)
to minimize information loss due to compression and/or trans-
formation during saving process. Afterward, a median filter was
used to remove unavoidable device noise contamination during
image acquisition. The median filter is a nonlinear filtering
method, which is appropriate for elimination of instrumental
noises and which are mostly impulse (salt and pepper) type
[65]. Using the median filter, these noises are eliminated, while
image sharpness and image details are preserved [65].
The median filter operates based on the median of pixel
intensities in a predefined neighborhood, which is entitled “win-
dow.” For this target, a 3 × 3 pixel window was used for filtration
of 96 dpi images. As the image resolution increased, the window
Table 3. Specifications and Conditions of Stylus
Profilometry
Sampling length 2.5 mm
Evaluation length 12.5 mm
Stylus radius 2 μm
Stylus cone angle 90°
Stylus load 0.7 mN
Stylus scanning speed 0.5 mm/s
Research Article
Fig. 2. Three-dimensional images of the steel samples with eight
different roughness levels (image resolution: 600 dpi; image dimen-
sions: 64 × 64 pixels). Each image represents a 2.35 mm ×
2.35 mm area of a real sample surface.
dimensions increased as well, with the same ratio. In Fig. 2, 3D
median filtered images (x, y, and I , which represent pixel coor-
dinates and normalized intensity [see Eqs. (2) and (3)], respec-
tively) of the steel samples with eight different surface roughness
levels are shown. The images are in 600 dpi resolution and have
dimensions of 64 × 64 pixels. Each image represents a
2.35 mm × 2.35 mm area of a real sample surface.
E. Principal Concepts of Surface Image
A digital image is composed of minute elements that are called
pixels (picture element). Each pixel in a gray-scale image is
identified by three coordinates, i.e., x, y, and I . The first
two coordinates indicate location of the pixel in the image,
and I specifies the pixel intensity, which can be an integer be-
tween 0 and 255 (i.e., 0 is attributed to a completely black pixel
and 255 is attributed to a completely white pixel) [66]. The
intensity of a pixel in the image of an object is a function
of three variables: incident light (more precisely, incident light
intensity and direction), the object reflection properties (which
are related to surface nature of the object), and the object sur-
face roughness [5]. Thus, if the sample surface nature as well as
the lighting conditions are kept homogeneous in the whole sur-
face of the sample and also between different samples (as were
done in this research), the image pixel intensities can be con-
sidered proportional to the surface roughness [Eq. (1)]. As is
apparent from the sample’s surface images (Fig. 2), pixel inten-
sities are somehow indicative of the sample’s surface roughness.
Obviously, absolute values of the pixel intensities are not nec-
essarily the exact estimation of the sample’s surface roughness:
H
X , Y   k × I
x, y: (1)
In this equation, H
X , Y  is the height of a point on the real
sample surface with (X , Y ) coordinates, and I
x, y is the in-
tensity of the corresponding pixel on its image. k is a scaling
factor. Different criteria can be considered for calculation of k,
which will be discussed in the next section.
 Research Article Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A 1003

F. Image Profilometry
For determination of k, extraction of image profile and calcu-
lation of the roughness parameters from the image profile, in
the first step, image pixel intensities were normalized. For this
purpose, the minimum pixel intensity of the image was sub-
tracted from all the image pixel intensities [Eq. (2)].
Subsequently, new pixel intensity values were divided by the
maximum of pixel intensity of the image [Eq. (3)].
Therefore, all the image pixel intensities will be a real number
in the [0, 1] interval:
I 0
x, y  I
x, y − min
I
x, y, (2)
surface. In the second approach, intensity data of each line were
filtered with GC. For this purpose, a Gaussian window (gw)
was defined according to Eq. (4):



1 n 2
gw
n  exp − α : (4)
2 N ∕2
In this equation, N shows window length (the number of
existing points in the window). In this study, N was considered
equal to that of the stylus profilometry’s cutoff value. Moreover,
α 
N − 1∕2σ and −N ∕2 ≤ n ≤ N ∕2 [67]. Afterward, in-
tensity data of each chosen line were convoluted by the nor-
malized Gaussian window [Eqs. (5) and (6)]:

I 0
x, y gw
n
I Norm
x, y  : gwNorm
n  P , (5)
n gw
n
(3)
max
I 0
x, y
In the next step, three lines were chosen over the sample image I GC  I Norm − I Norn  gwNorm : (6)
in different locations and directions. It is worth noting that the
In the final stage, the filtered data (I GC ) were further filtered
length of the chosen lines was equalized with the evaluation length
by CT similar to the first approach. In the third approach, in-
of the stylus profilometry. The equalization was implemented by
tensity data of each line were filtered by MA according to Eq. (7).
attaching a scale to the samples in the image acquisition process.
It should be noted that the length of the window in moving
For instance, in the images with 1200 dpi resolution, each pixel
average filter (2M  1) was considered equal to the stylus pro-
equals 21.2 μm of the real surface. Thus, for investigation of
filometry’s cutoff value. Thereafter, the filtered data (I MA ) were
12.5 mm of the real sample surface, lines with the length of
filtered again by CT similar to the first approach:
590 pixels were considered in the image. As the image resolution
nX
M
changed, the line lengths changed as well, with the same ratio. 1
I MA
n  I Norm
n − I
i: (7)
Performances of four filtering approaches in elimination of 2M  1 in−M Norm
low-frequency surface undulations (i.e., form and waviness)
were investigated. The main objective of employing the four Finally, in the fourth approach, elimination of the low-
different filtering approaches was to study which method per- frequency undulations of the intensity data was done according
forms better in elimination of low-frequency surface undula- to Al-kindi et al.’s proposed method [15]. In this method, suc-
tions. It is essential to note that low-frequency surface cessive moving average filters (SMA) with a window length of
undulations must not exist in the roughness profile. In the 5 pixels are used. In the first stage, the pixel intensity data were
roughness profile, only high-frequency surface undulations estimated from the mean of neighboring pixels (a symmetric
(i.e., roughness) are allowed to be present. The mentioned four neighborhood with the length of 5 pixels) according to Eq. (8):
filtering methods are from different filtering families. The cut- j2
1X
off filtering method (CT) was chosen because this method was I 1
j  I
i: (8)
also used in the stylus profilometry. If a filtering approach is 5 ij−2 Norm
suitable for the mechanical method (i.e., stylus profilometry), Afterward, the mentioned stage was repeated K ∕4 − 1 more
it will not necessarily work properly for the optical method (i.e., times (K is the total number of pixels in the line) to attain
image profilometry). Therefore, a Gaussian convolution filter- I K ∕4
j, which are the estimations of low-frequency undula-
ing method (GC) was also tested as a powerful filtering method tions of the intensity data. Subtracting I K ∕4 from I Norm , ac-
in the image profilometry. Moving average (MA) and successive cording to Eq. (9), the low-frequency undulations will be
moving average (SMA) filtering methods, which are robust fil- removed from the intensity data:
ters from the averaging method family were employed with the
same reasoning. Accordingly, a variety of filtering approaches I SMA  I Norm − I N ∕4 : (9)
has been investigated in the image profilometry. Finally, the filtered data (I SMA ) were filtered one more time
As noted, in the first approach, analogous to the stylus pro- by CT similar to the first approach.
filometry, only CT was applied. In the CT filtering method, the After elimination of the low-frequency undulations from the
best fitted line on the profile points over each sampling length is pixel intensity data by each of these approaches, R a , R z , and R t
considered as the estimation of the long wavelength surface un- were calculated for the resulting profiles (see Appendix A).
dulations. Therefore, for elimination of the long wavelength Then, the k factor of Eq. (1) is defined using three different
surface undulations in the profile, the height differences be- criteria [Eqs. (10)–(12)]:
tween the profile points and the fitted line are calculated on
R a
stylus
each sampling length. Considering the resolution, the cutoff ka  , (10)
value (i.e., sampling length) in the image was equalized with R a
image
the stylus profilometry’s cutoff value (i.e., 2.5 mm). For exam-
R z
stylus
ple, the cutoff value was considered 118 pixels for the images kz  , (11)
with 1200 dpi resolution, which equals 2.5 mm of the real R z
image
 1004 Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A
Fig. 3. Flow chart of the image processing stages for calculation of the surface roughness parameters using the steel samples images.
R t
stylus
kt  : (12)
R t
image
The image profiles were rescaled by each of these scaling
factors to simulate the stylus profiles. Afterward, the roughness
parameters were calculated for the rescaled profiles. All these
stages are illustrated in a flow chart in Fig. 3.
3. RESULTS AND DISCUSSION
A. Stylus Profilometry Results
The average values of 10 roughness parameters for the steel
samples, which were measured by the stylus profilometer,
are graphically displayed in Fig. 4. The standard deviations
of measurement repetitions are also presented in the figure.
Moreover, in Fig. 5, typical examples of surface profiles of
the steel samples with the lowest (sample 1), medium (sam-
ple 4), and the highest (sample 8) surface roughness levels
are presented. According to Fig. 4, it can be seen that the mini-
mum R a value is for sample 1 (0.54 μm), and the largest R a is
attributed to sample 8 (12.13 μm), as expected. This trend can
Research Article
be observed in most of the amplitude parameters (i.e., the mini-
mum parameter value is for sample 1 and the maximum param-
eter value is for sample 8). This, among R sk and R ku , does not
follow an ordered trend. R ku for most of the samples is greater
than 3, which means that the profile points are mostly distrib-
uted near the reference line. Figure 5 also confirms this result. It
is obvious from Fig. 5 that the surface treatment by the abrasive
blasting method creates an irregular/random pattern contrary
to the processes such as turning and milling, which produce
repeating patterns. The small values of R sk also indicate that
the roughness pattern on the sample surface is random, and
the height distribution of the surface points is close to the nor-
mal distribution. The greatest value of the frequency parameter
(R Sm ) is for sample 7, and the smallest value is for sample 1.
This result shows that surface roughening with the steel shots
produce low-frequency fluctuations.
The smallest value of the hybrid parameter (R Δq ) belongs to
sample 1 and sample 7, and the largest value of this parameter is
for sample 8. The other noteworthy point about the stylus pro-
filometry results is small values of the standard deviations,
which were obtained from repetition of measurements.
 Research Article Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A
Fig. 4. Average values of 10 roughness parameters for the steel sam-
ples, which were measured by the stylus profilometer. Error bars show
the standard deviations of measurement repetitions.
In the worst case, the standard deviation value reaches to 15%
of average value (it belongs to R sk of sample 4, excluding sample
1 data because sample 1 was not roughened by the abrasive
blasting method). Thus, it can be concluded that the abrasive
1005
Fig. 5. Steel samples surface profiles with eight different roughness
levels (recorded by the stylus profilometer).
blasting method produces a uniform and reproducible surface
texture. It also confirms the stylus profilometry method’s pre-
cision and repeatability.
B. Image Profilometry Results
According to Section 2.F (image profilometry) instructions, im-
age profiles were constructed using the steel sample images. In
Fig. 6, examples of image profiles of sample 4 are shown. These
profiles were obtained from the image with 800 dpi resolution.
As displayed in Fig. 6, a line with the length of 395 pixels (this
equals the evaluation length of the stylus profilometry in terms
of pixels) was chosen in the image. The image profile is con-
sidered as the curve of the pixel intensities versus pixel num-
bers. Subsequently, the long wavelength undulations of the
profile (i.e., form and waviness) were estimated using four fil-
tering approaches, including CT (with the cutoff value of 79
pixels, which equals the cutoff value of the stylus profilometry
in terms of pixels), GC  CT, MA  CT, and SMA  CT
approaches. The resulting estimations of the long wavelength
undulations of the profile, from each of these approaches, are
presented in Fig. 6. Thereafter, the long wavelength undula-
tions of the profile were subtracted from the profile data. In
this manner, form and waviness were eliminated from the sur-
face profile data. In the next step, R a , R z , and R t were calculated
for the filtered image profiles. Using these roughness parame-
ters and the corresponding stylus-based roughness parameters,
ka , kz , and k t were calculated. It should be noted that, although
R a , R z , and R t values differed significantly from each other, k a ,
kz , and k t values were not different.
Afterward, the amplitude values of the filtered image profiles
were multiplied by ka, kz , and k t , and three differently rescaled
profiles were obtained for each of the filtering approaches
(Fig. 6). Finally, the surface roughness parameters (see
Appendix A) were calculated for each of the rescaled profiles
and were compared with those of the stylus profilometry’s
parameters. According to Fig. 6, it is revealed that there are
 1006 Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A Research Article

Fig. 6. Image profiles of Sample 4, obtained from the image with 800 dpi resolution. At first, a line with the length of 395 pixels was chosen in the
image. Then, the image profile was plotted as the curve of the line pixel intensities versus pixel numbers. Subsequently the long wavelength un-
dulations of the profile were estimated using four filtering approaches, i.e., CT, GC  CT, MA  CT, and SMA  CT. Afterward, the long
wavelength undulations of the profile were subtracted from the profile data. In this manner, form and waviness were eliminated from the surface
profile data. Using k a , k z , and kt scaling factors, the image profiles resulted from each filtering approaches rescaled in terms of micrometers.

no evident differences between the image profiles obtained

from the different filtering approaches. Furthermore, the re-
scaled profiles using different scaling factors are quite similar.
For the other samples and image resolutions, the same pro-
cedure was followed.
In Fig. 7, typical image profiles of the steel samples with the
lowest (sample 1), medium (sample 4), and the highest (sample 8)
surface roughness levels have been illustrated. These profiles have
been obtained from the images with 800 dpi resolution. For these
profiles, the GC  CT filtering approach and k a scaling factor
were used. It is worth mentioning that 800 dpi resolution, rescal-
ing using ka scaling factor as well as GC  CT filtering approach,
created results that correlated the best with the stylus profilometry
results. Comparing Fig. 7 with Fig. 5, it is noted that the image
profiles correctly have demonstrated surface roughness variations
among the eight different samples. Considering the image profiles
of Fig. 7, it can be asserted that the stylus profiles have been prop-
erly simulated by the image profiles.
For more accurate investigation of image resolution, filtering
method, and scaling factor impacts on the correlations between
the image profilometry results and the stylus profilometry re- Fig. 7. Image profiles of the steel samples with eight different rough-
sults, error percent was calculated for all the image profilometry ness levels: 800 dpi resolution, GC  CT filtering approach, and k a
results according to Eq. (13): scaling factor were employed to construct the profiles.
 Research Article Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A
jR stylus − R image j
Error percent  × 100: (13)
R stylus
In this equation, R stylus shows the stylus-based roughness
parameter, while R image indicates the image-based roughness
parameter.
The error percent values are presented in Fig. 8 for eight
image resolutions, three scaling factors, and four filtering
approaches. In this figure, the error percent values are pre-
sented in the form of colored boxes in which each color is
representative of an error percent value. According to Fig. 8,
it can be concluded that the image resolution has a dramatic
impact on the image profilometry accuracy. The best results
belong to 800 dpi resolution, and the most erroneous results
are for 96 dpi resolution. In CT, GC  CT, MA  CT, and
SMA  CT filtering approaches, changing resolution from
96 to 800 dpi caused a 76%, 83%, 77%, and 68% reduction
in mean error values, respectively [total average in one resolu-
tion (TAOR) boxes in Fig. 8].
It should be considered that the colored boxes, which are
attributed TAOR in Fig. 8, show the average error of 10 param-
eters obtained from three differently rescaled profiles [i.e., it is an
average of 30 numbers (boxes) in a specific image resolution].
As with the reduction of tip radius in the stylus profilometry,
which makes it possible to detect more surface details (i.e., tiny
peaks and valleys), by increasing image resolution, more surface
detailed information can be recorded. It appears that an in-
crease of image resolution simultaneously improves both the
lateral and height resolutions in the image profilometry.
Because, by increasing the image resolution, the number of re-
corded surface points per specific length, which is the definition
of lateral resolution, increases as well. Moreover, in low reso-
lutions, reflection intensity of a large number of the surface
points are averaged and recorded as a single pixel intensity
in the image. Therefore, heights of all these surface points
are considered the same in the image. As the image resolution
increases, the number of the recorded surface points per image
pixel decreases. Accordingly, height differences between differ-
ent surfaces points will be recorded more accurately in the im-
age, which causes improvement of the height resolution.
The inaccurate results from the images with 96 dpi resolu-
tion can be ascribed to their low quality. Surface details have
been poorly captured in the 96 dpi images and a great deal
of surface reflection information has been eliminated in these
images. As the image resolution increases, more surface details
are recorded; therefore, the image profilometry results correlate
better with the stylus profilometry results. But, as the image
resolution goes beyond a threshold (800 dpi, in this research),
the accuracy of the image profilometry falls again. This drop in
accuracy for high resolutions (more than 800 dpi) can be attrib-
uted to the inevitable increase of instrumental noises (i.e., the
noise contamination, which is introduced by digital scanner’s
CCDs) during the image acquisition process. It should be men-
tioned that, although more surface details can be recorded as
the capturing resolution increases, the probability of instru-
ment noise intrusion increases as well [61]. Furthermore, in
800 dpi resolution images, the numbers of sampling points
per unit length (i.e., lateral resolution, 316 points/cm) are clos-
est to that of the stylus profilometry’s lateral resolution.
1007
It appears that these aforementioned factors synergistically
caused the results from the 800 dpi resolution to possess the
optimum accuracy.
The other important point, concerning Fig. 8 data, is that
rescaling of image profiles by ka leads to more accurate results
compared with two other scaling factors [10 roughness param-
eters’ average (10RPA) boxes in Fig. 8]. It should be noted that
the colored boxes, which have been attributed as 10RPA in
Fig. 8, show the average of 10 roughness parameters obtained
from a rescaled image profile by a specific scaling factor in a spe-
cific image resolution [i.e., it is the average of 10 numbers (boxes)
in a specific resolution and for a specific rescaling factor]. It
should be considered that profile rescaling by ka not only min-
imizes R a error percent, it also causes R q , which has a fairly sim-
ilar definition to R a , to become close to that of the stylus-based
value. Therefore, employing this scaling factor minimizes the
error percent values of two roughness parameters, while using
two other scaling factors (k z and kt ) simply minimizes one
roughness parameter’s error percent (R z and R t , respectively).
Rescaling of the image profiles by kt led to the most inaccurate
results in comparison with the two other scaling factors. This
observation can be attributed to the fact that R t depends only
on two points of the profile (i.e., the highest peak and the deepest
valley). Therefore, the magnitude of R t is vulnerable to be
affected by different sources of errors during the measurement.
Moreover, R t depends highly on the location of the recorded
profile on the sample’s surface. It is worth noting that, generally,
the impact of scaling factor changes on the image profilometry
accuracy is minor compared with that of the resolution effect.
The highest improvements resulting from the scaling factor
changes were 26.04%, 29.36%, 26.71%, and 15.81%, which
were for CT, GC  CT, MA  CT, and SMA  CT filtering
approaches, respectively. The different rescale average (DRA)
boxes in Fig. 8 indicate the average value of a roughness param-
eter obtained from three differently rescaled image profiles by k a,
kz , and kt scaling factors (i.e., it is the average of three values of a
specific roughness parameter in a specific image resolution, while
the rescaling factor changes). As illustrated in Fig. 8, in almost all
resolutions and filtration approaches, the smallest error percent is
for R q, and the greatest error percent belongs to R sk .
Among the different filtering approaches, GC  CT was
the best approach, and SMA  CT was the most inaccurate.
If only the results of the best resolution (800 dpi) are consid-
ered (total average in 800 dpi boxes), it will be revealed that
GC  CT and MA  CT filtering approaches performed
33.7% and 6.7% more accurately than the single CT filtering
method, respectively. On the other hand, the SMA  CT fil-
tering approach worked 31.1% less accurately than the single
CT. Generally, it can be concluded that using images with
800 dpi resolution, ka as the scaling factor, and GC  CT fil-
tering approach, the image profilometry results strongly corre-
late with the stylus profilometry results.
It should be mentioned that removal of the instrumental
noises employing the median filter improved the image-based
results. On average, the median filtered images showed 2.67%
less error compared with the case of unfiltered images.
To develop mathematical relations between the image-based
roughness parameters (without rescaling) and the stylus-based
roughness parameters, the stylus profilometry results were
 1008 Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A Research Article

Fig. 8. Error percent results for 10 image-based roughness parameters in different conditions: eight image resolutions (the bottom axis), three scaling
factors (the top axis), and four filtering approaches (the right-hand axis). Error percent values are presented in the colored boxes (see color bar).
 Research Article Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A 1009

plotted against the image profilometry results. Using the least-

squares method, the best line was fitted on the experimental
results of each roughness parameter (Fig. 9). The equations
for prediction of the stylus profilometry roughness parameters
using the corresponding image profilometry parameters are pre-
sented in Table 4. It must be mentioned that, for this purpose,
images with 800 dpi resolution were used, and the GC  CT
filtering approach was employed. It is worth considering that
the k scaling factor is used just to rescale the image-based sur-
face profile in order to generate a surface profile with the iden-
tical dimensional units to that of the stylus-based surface profile
and with the shape as similar as possible to the stylus-based
surface profile. Thus, if a user only needs roughness parameters
and does not require the surface profile, there will be no need
for the k scaling factor, and, accordingly, the equations shown
in Table 4 will be sufficient.
According to Fig. 9, the strongest and the weakest correlations
between the results of the stylus profilometry and the image pro-
filometry belong to R t
R 2  0.9892 and R sk
R 2  0.9313,
respectively. To verify accuracy of the mathematical equations,
three differently roughened samples were prepared using abrasive
blasting with different conditions compared with that reported
in Table 2 (nozzle distance and angle were different). Roughness
parameters of these new samples were measured using the stylus
profilometer. The stylus-based R a of the samples were obtained
in the range of 2–5 μm. Employing 800 dpi images and GC 
CT filtering approach, the image-based roughness parameters
were calculated for the samples. Afterward, the mathematical
equations in Table 4 were used to predict the stylus-based rough-
ness parameters. The predicted parameters were compared with
the actual corresponding results. The highest deviation of pre-
dicted roughness from the actual one was for R sk with the value
of 41.68%, and the lowest deviation belonged to R z with the
value of 4.27%. It should be noted that, once the calibration
curves are generated using the stylus profilometry in the first
step, the image profilometry can predict the stylus profilometry’s
roughness parameters independently, with no need to employ
the stylus profilometry again.
In comparison with methods such as an artificial neural net-
work (ANN) [3,20], the curve fitting method and development
of the mathematical equations between the image- and stylus-
based results require a substantially smaller number of samples.
It is considered as one of the major virtues of the curve fitting

Table 4. Mathematical Equations for Prediction

of the Stylus-Based Roughness Parameters Using
Image-Based Roughness Parameters
Parameter Equation
Ra R stylus  93.7468 × R image − 2.0241
Rq R stylus  93.4609 × R image − 1.8956
Rp R stylus  118.7374 × R image − 12.7631
Fig. 9. Plots of the stylus profilometry roughness parameters against Rv R stylus  108.5161 × R image − 10.1516
the image profilometry roughness parameters. The best lines were fit- Rz R stylus  119.8179 × R image − 25.5574
ted on the experimental data using the least-squares method. The lin- Rt R stylus  152.3519 × R image − 51.6884
ear correlation coefficient (R 2 ) is reported in each plot. The horizontal R sk R stylus  1.0038 × R image − 0.1123
error bars are for the standard deviations of the image profilometry R ku R stylus  0.6758 × R image  1.2620
repetition, and the vertical error bars are for the standard deviations R Sm R stylus  31.4235 × R image  34.0790
of the stylus profilometry repetition. R Δq R stylus  18.7558 × R image − 0.5010
 1010 Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A
method. It is worth noting that the effective learning method of
support vector machine (SVM), as introduced by Vapnik [68], is
also a powerful method for classification and regression prob-
lems. In comparison with ANN, SVM requires a smaller number
of data, shorter training time, and showed stronger generalization
[69–72]. However, the curve fitting’s mathematical calculations
are remarkably simpler than that of the learning methods.
Therefore, generally it is preferable to develop a simple math-
ematical equation for the description of the experimental data
rather than using complex deep learning methods, if possible.
Moreover, using classification methods such as calculation of
Euclidian or Hamming distances between reference samples with
known roughness values and test samples [18], estimation of
roughness values for the test samples with the roughness values
different from the reference samples would be inaccurate,
whereas by determination of calibration curves and mathematical
equations, roughness values of the samples can be estimated as
fairly accurate regardless of how close the test sample’s roughness
is to one of the reference samples.
As already mentioned (see introduction), in the stylus profil-
ometry, the diamond tip moves slowly on the sample surface. For
instance, the tip of the profilometer used in this research moved
along the surface with a speed of about 0.5 mm/s. Therefore, to
traverse just an evaluation length, about half a minute is required;
together with calculation process and extraction of roughness
parameters, this time reaches 3 min.
If at least five repetitions of surface roughness measurement
are required in different locations on the surface for the sake of
accuracy and reproducibility, the measurement time will reach
about 15 min. On the other hand, in the proposed image-based
method, the whole process from image acquisition to extraction
of surface roughness parameters takes less than a minute.
Additionally, it should be considered that the image analysis
process required time; using a personal computer, it is recorded
to be less than 10 s. Moreover, once a sample image is captured,
several profiles can be chosen on the surface image to repeat the
measurement, and there will be no need to recapture the surface
image for every measurement. For example, if five profile mea-
surements are required on the sample surface, the total duration
of measurement time will reach 100 s (60 s for image acquis-
ition and first profile measurement and 40 s for the other four
profile measurements). For the given examples, the image pro-
filometry preforms nine times faster than the stylus profilom-
etry. Thus, it can be claimed that the image profilometry is a
more effective method in online application compared with the
stylus profilometry.
Fig. 10. Example of surface profile together with fundamental profile’s concepts.
Research Article
4. CONCLUSIONS
In the present study, carbon steel specimens were roughened by
the abrasive blasting method, and an image processing ap-
proach was proposed to evaluate their surface roughness using
a machine vision system. The impacts of different variables
such as image resolution and filtering method on the perfor-
mance of the approach were investigated. The technical
achievements of the research can be summarized as follows:
– Employing the suggested novel image processing method,
the stylus-based roughness parameters of nonperiodic (ran-
domly) rough surfaces could be predicted as fairly accurate
based on the image data. Moreover, profiting from machine
vision advantages, including online applicability, low cost, high
speed, noncontact, and nondestructive nature of the method,
would be possible especially in production lines.
– To a certain extent, the approach accuracy in the estima-
tion of the roughness parameters owes to controlled conditions,
including uniform lighting, noise elimination, and equalization
of the evaluation profile length in both the stylus profilometry
and image profilometry.
– Image resolution has a great influence on the accuracy of the
image processing results. Generally, by increasing the image resolu-
tion, the accuracy of the results was improved. However, it should
be noted that the image resolution had an optimum value (800 dpi).
– The best and the worst filtering approaches for elimina-
tion of the low-frequency surface undulations were GC  CT
and SMA  CT, respectively.
– Rescaling of the image profiles with k a scaling factor (i.e.,
the ratio of the stylus-based R a to the image-based R a ) led to
most reliable results in the simulation of the stylus-based sur-
face profiles, compared with k z and kt scaling factors.
– The most accurate prediction was made for R t and R z
parameters, and the least accurate prediction was made for
R sk and R Δq parameters.
– A possible restriction on the image profilometry method
could be the dependence of the method on the surface nature.
Because pixel intensity in an image is a function of surface rough-
ness as well as surface reflection nature, roughness of surfaces with
different reflection properties cannot be expressed by a single
equation in image profilometry. Thus, the under-investigation
surfaces must have the same reflection nature in order to carry
out precise roughness evaluations.
APPENDIX A
In this section, substantial concepts of the profilometry method
have been graphically presented. In Fig. 10, a typical surface
 Research Article Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A 1011

Table 5. Mathematical Definitions of 10 Profile Roughness Parameters

Roughness Parameter Description Mathematical Definitiona
Amplitude parameters P P
Arithmetic average height or This parameter shows the average of absolute profile points’ R a  15 5j1 R aj , R aj  1n ni1 jZ i j
central line average (R a ) distances from the reference line in a sampling length P qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P ffi
Root mean square roughness (R q ) This parameter shows the standard deviation of profile points’ R q  15 5j1 R qj , R qj  1n ni1 Z 2i
distances from the reference line in a sampling length P max
Z pi 
The maximum peak height (R p ) This parameter shows distance of the highest peak from the R p  15 5j1 R pj , R pj  i1, 2, …, n
reference line in a sampling length P5
1 max
Z vi 
The maximum valley depth (R v ) This parameter shows distance of the deepest valley from the R v  5 j1 R vj , R vj  i1,2, …, n
reference line in a sampling length P max
Z  max
Z vi 
The maximum peak to valley This parameter shows vertical distance between the highest R z  15 5j1 R z j R z j  i1,2, …,pin  i1,2, …, n
height (R z ) peak and the deepest valley in a sampling length
max
Z pi  max
Z vi 
The absolute profile height (R t ) This parameter shows vertical distance between the highest R t  i1,2, …, N  i1,2, …, N
peak and the deepest valley in the evaluation length PN
1 3
Skewness (R sk ) This parameter shows symmetry of probability density R sk  N ×R 3 i1 Z i
q
distribution of the profile points’ distances from the reference
line in the evaluation length. If the distribution is symmetric,
R sk will be zero. Otherwise, depends on direction of
asymmetry, R sk can be a positive or negative value. PN
1 4
Kurtosis (R ku ) This parameter shows sharpness of probability density R ku  N ×R 4 i1 Z i
q
distribution of the profile points’ distance from the reference
line in the evaluation length. If the distribution is normal, R ku
will be equal to 3. If the distribution is sharper/broader than
the normal distribution, R ku will be bigger/smaller than 3.
Frequency (spacing) parameters P5 Pm
Arithmetic average width of profile This parameter shows the average of profile elements’ R Sm  15 j1 R Smj , R Smj  m1 i1 X si
elements or mean spacing (R Sm ) horizontal lengths in a sampling length
Hybrid parameters qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P
Root mean square slopes of This parameter shows root mean square slopes of the profile R Δq  N1 ni1
θi − θ2
profile (R Δq ) points in the evaluation length. P Z i −Z i−1
θ  N1 Ni1 θi , θi  Δx
a
N : number of the profile points over the evaluation length. n  N ∕5: number of the profile points over a sampling length. m: number of the elements over a sampling
length.

profile is presented. In this figure, the concepts of evaluation
length, sampling length, peak, valley, element, reference line,
and some others have been shown. Furthermore, in
Table 5, the mathematical definitions of 10 different profiles’
roughness parameters used in this study have been presented.
REFERENCES
1. G. A. Al-Kindi and B. Shirinzadeh, “Feasibility assessment of vision-
based surface roughness parameters acquisition for different types of
machined specimens,” Image Vis. Comput. 27, 444–458 (2009).
2. P. Priya and B. Ramamoorthy, “The influence of component inclina-
tion on surface finish evaluation using digital image processing,” Int. J.
Mach. Tools Manuf. 47, 570–579 (2007).
3. G. Samtaş, “Measurement and evaluation of surface roughness
based on optic system using image processing and artificial neural
network,” Int. J. Adv. Manuf. Technol. 73, 353–364 (2014).
4. E. S. Gadelmawla, “Estimation of surface roughness for turning oper-
ations using image texture features,” Proc. Inst. Mech. Eng. Part B J.
Eng. Manuf. 225, 1281–1292 (2011).
5. G. A. Al-Kindi and B. Shirinzadeh, “An evaluation of surface rough-
ness parameters measurement using vision-based data,” Int. J.
Mach. Tools Manuf. 47, 697–708 (2007).
6. E. S. Gadelmawla, “A vision system for surface roughness characteri-
zation using the gray level co-occurrence matrix,” NDT&E Int. 37,
577–588 (2004).
7. D. E. Packham, “The mechanical theory of adhesion,” in Handbook of
Adhesive Technology, Revised and Expanded, A. Pizzi and K. L.
Mittal, eds. (CRC Press, 2003), pp. 69–93.
8. R. Kamguem, S. A. Tahan, and V. Songmene, “Evaluation of ma-
chined part surface roughness using image texture gradient factor,”
Int. J. Precis. Eng. Manuf. 14, 183–190 (2013).
9. H. J. Butt, K. Graf, and M. Kappl, Physics and Chemistry of Interfaces
(Wiley, 2003).
10. H. J. Butt and M. Kappl, Surface and Interfacial Forces (Wiley, 2009).
11. R. Kumar, P. Kulashekar, B. Dhanasekar, and B. Ramamoorthy,
“Application of digital image magnification for surface roughness evaluation
using machine vision,” Int. J. Mach. Tools Manuf. 45, 228–234 (2005).
12. C. Y. Poon and B. Bhushan, “Comparison of surface roughness mea-
surements by stylus profiler, AFM and noncontact optical profiler,”
Wear 190, 76–88 (1995).
13. S. Soleimani, J. Sukumaran, A. Kumcu, P. De Baets, and W. Philips,
“Quantifying abrasion and micro-pits in polymer wear using image
processing techniques,” Wear 319, 123–137 (2014).
14. P. Sarma, L. Karunamoorthy, and K. Palanikumar, “Surface rough-
ness parameters evaluation in machining GFRP composites by
PCD tool using digital image processing,” J. Reinf. Plast. Compos.
28, 1567–1585 (2009).
15. G. A. Al-Kindi, R. M. Baul, and K. F. Gill, “An application of machine
vision in the automated inspection of engineering surfaces,” Int. J.
Prod. Res. 30, 241–253 (1992).
16. A. K. Bewoor and V. A. Kulkarni, Metrology and Measurement
(McGraw-Hill Education, 2009).
17. D. J. Whitehouse, Handbook of Surface and Nanometrology (CRC
Press, 2010).
18. T. Jeyapoovan and M. Murugan, “Surface roughness classification
using image processing,” Measurement 46, 2065–2072 (2013).
19. S. Jetley and D. Selven, “Applying machining vision to surface texture
analysis,” in IEEE 36th Midwest Symposium on Circuits and Systems
(IEEE, 1993), pp. 1456–1459.
 1012 Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A
20. B. Dhanasekar and B. Ramamoorthy, “Restoration of blurred images
for surface roughness evaluation using machine vision,” Tribol. Int. 43,
268–276 (2010).
21. B. Dhanasekar, N. K. Mohan, B. Bhaduri, and B. Ramamoorthy,
“Evaluation of surface roughness based on monochromatic speckle
correlation using image processing,” Precis. Eng. 32, 196–206
(2008).
22. L. De Chiffre, H. Kunzmann, G. N. Peggs, and D. A. Lucca, “Surfaces
in precision engineering, microengineering and nanotechnology,”
CIRP Ann. 52, 561–577 (2003).
23. A. A. G. Bruzzone, H. L. Costa, P. M. Lonardo, and D. A. Lucca,
“Advances in engineered surfaces for functional performance,”
CIRP Ann. 57, 750–769 (2008).
24. R. Leach, Characterisation of Areal Surface Texture (Springer-Verlag,
2013).
25. M. Conroy and J. Armstrong, “A comparison of surface metrology
techniques,” J. Phys. Conf. Ser. 13, 458–465 (2005).
26. M. K. Balasundaram and M. M. Ratnam, “In-process measurement
of surface roughness using machine vision with sub-pixel edge detec-
tion in finish turning,” Int. J. Precis. Eng. Manuf. 15, 2239–2249
(2014).
27. L. Tchvialeva, I. Markhvida, H. Zeng, D. I. McLean, H. Lui, and T. K.
Lee, “Surface roughness measurement by speckle contrast under the
illumination of light with arbitrary spectral profile,” Opt. Lasers Eng. 48,
774–778 (2010).
28. U. Persson, “Real time measurement of surface roughness on ground
surfaces using speckle-contrast technique,” Opt. Lasers Eng. 17,
61–67 (1992).
29. T. Fricke-Begemann and K. D. Hinsch, “Measurement of random
processes at rough surfaces with digital speckle correlation,” J.
Opt. Soc. Am. A 21, 252–262 (2004).
30. R.-S. Lu, G.-Y. Tian, D. Gledhill, and S. Ward, “Grinding surface
roughness measurement based on the co-occurrence matrix of
speckle pattern texture,” Appl. Opt. 45, 8839–8847 (2006).
31. L. C. Leonard and V. Toal, “Roughness measurement of metallic sur-
faces based on the laser speckle contrast method,” Opt. Lasers Eng.
30, 433–440 (1998).
32. R. Leach, Optical Measurement of Surface Topography (Springer-
Verlag, 2011).
33. J. C. Wyant, C. L. Koliopoulos, B. Bhushan, and O. E. George, “An
optical profilometer for surface characterization of magnetic media,”
ASLE Trans. 27, 101–113 (1984).
34. R. Danzl, F. Helmli, and S. Scherer, “Automatic measurement of cal-
ibration standards with arrays of hemi-spherical calottes,” in 11th
International Conference on Metrology and Properties of
Engineering Surfaces, Huddersfield, UK, July 17–20, 2007.
35. R. Danzl and F. Helmli, “Form measurement of engineering parts
using an optical measurement system based on focus variation,” in
7th European Society for Precision Engineering and Nanotechnology
International Conference, Bremen, Germany, May 20–24, 2007.
36. R. K. Leach, C. L. Giusca, H. Haitjema, C. Evans, and X. Jiang,
“Calibration and verification of areal surface texture measuring instru-
ments,” CIRP Ann. 64, 797–813 (2015).
37. R. Su, Y. Wang, J. Coupland, and R. Leach, “On tilt and curvature
dependent errors and the calibration of coherence scanning interfer-
ometry,” Opt. Express 25, 3297–3310 (2017).
38. R. Mandal, J. Coupland, R. Leach, and D. Mansfield, “Coherence
scanning interferometry: measurement and correction of three-
dimensional transfer and point-spread characteristics,” Appl. Opt.
53, 1554–1563 (2014).
39. A. J. Henning, J. M. Huntley, and C. L. Giusca, “Obtaining the transfer
function of optical instruments using large calibrated reference ob-
jects,” Opt. Express 23, 16617–16627 (2015).
40. T. V. Vorburger, H.-G. Rhee, T. B. Renegar, J.-F. Song, and A. Zheng,
“Comparison of optical and stylus methods for measurement of sur-
face texture,” Int. J. Adv. Manuf. Technol. 33, 110–118 (2007).
41. F. Luk, V. Huynh, and W. North, “Measurement of surface roughness
by a machine vision system,” J. Phys. E 22, 977–980 (1989).
42. X. Li, L. Wang, and N. Cai, “Machine-vision-based surface finish in-
spection for cutting tool replacement in production,” Int. J. Prod. Res.
42, 2279–2287 (2004).
Research Article
43. B. Y. Lee, S. F. Yu, and H. Juan, “The model of surface roughness
inspection by vision system in turning,” Mechatronics 14, 129–141
(2004).
44. K.-C. Lee, S.-J. Ho, and S.-Y. Ho, “Accurate estimation of surface
roughness from texture features of the surface image using an adap-
tive neuro-fuzzy inference system,” Precis. Eng. 29, 95–100 (2005).
45. A. Hladnik and M. Lazar, “Paper and board surface roughness char-
acterization using laser profilometry and gray level cooccurrence
matrix,” Nord. Pulp Pap. Res. J. 26, 99–105 (2011).
46. S. Ghodrati, S. G. Kandi, and M. Mohseni, “Mathematical modeling of
visual and actual surface roughness of real materials by gray level co-
occurrence matrix texture analysis method,” in 7th International
Congress on Color and Coatings, Tehran, Iran, December 19–21, 2017.
47. S. Ghodrati, S. G. Kandi, and M. Mohseni, “How accurately do differ-
ent computer-based texture characterization methods predict material
surface coarseness? A guideline for effective online inspection,”
J. Opt. Soc. Am. A 35, 712–725 (2018).
48. S. Ghodrati, S. G. Kandi, and M. Mohseni, “A histogram-based image
processing method for visual and actual roughness prediction of sand-
papers,” in 6th International Congress on Color and Coatings, Tehran,
Iran, November 10–12, 2015.
49. W. J. Stemp and M. Stemp, “UBM laser profilometry and lithic use-
wear analysis: a variable length scale investigation of surface topog-
raphy,” J. Archaeol. Sci. 28, 81–88 (2001).
50. S. Ghodrati, M. Mohseni, and S. G. Kandi, “Polymeric blends surface
roughness evaluation based on a digital image edge detection algo-
rithm,” in 7th International Congress on Color and Coatings, Tehran,
Iran, December 19–21, 2017.
51. B. Tholt, W. G. Miranda-Júnior, R. Prioli, J. Thompson, and M. Oda,
“Surface roughness in ceramics with different finishing techniques
using atomic force microscope and profilometer,” Oper. Dent. 31,
442–449 (2006).
52. P. F. Chauvy, C. Madore, and D. Landolt, “Variable length scale analy-
sis of surface topography: characterization of titanium surfaces for
biomedical applications,” Surf. Coat. Technol. 110, 48–56 (1998).
53. S. Ghodrati, S. G. Kandi, and M. Mohseni, “Dependence of adhesion
strength of an acrylic clear coat on fractal dimension of abrasive blasted
surfaces using image processing,” in 6th International Congress on
Color and Coatings, Tehran, Iran, November 10–12, 2015.
54. M. Behzadnasab, S. M. Mirabedini, K. Kabiri, and S. Jamali,
“Corrosion performance of epoxy coatings containing silane treated
ZrO2 nanoparticles on mild steel in 3.5% NaCl solution,” Corros.
Sci. 53, 89–98 (2011).
55. T. T. X. Hang, T. A. Truc, M.-G. Olivier, C. Vandermiers, N. Guérit, and
N. Pébère, “Corrosion protection mechanisms of carbon steel by an
epoxy resin containing indole-3 butyric acid modified clay,” Prog. Org.
Coat. 69, 410–416 (2010).
56. R. M. Hosseini, A. A. Sarabi, H. E. Mohammadloo, and M. Sarayloo,
“The performance improvement of Zr conversion coating through Mn
incorporation: with and without organic coating,” Surf. Coat. Technol.
258, 437–446 (2014).
57. B. Bhushan and B. K. Gupta, Handbook of Tribology: Materials,
Coatings, and Surface Treatments (McGraw-Hill, 1991).
58. D. M. Karpinos, V. G. Zil’berberg, A. M. Vyal’Tsev, and V. S. Kud’,
“Shot-blasting as a means of preparing surfaces for plasma deposi-
tion,” Powder Metall. Met. Ceram. 17, 675–678 (1978).
59. S. Amada and T. Hirose, “Influence of grit blasting pre-treatment on
the adhesion strength of plasma sprayed coatings: fractal analysis of
roughness,” Surf. Coat. Technol. 102, 132–137 (1998).
60. D. J. Whitehouse and P. Vanherck, “Survey of reference lines in the
assessment of surface texture,” CIRP Ann. 21, 267–273 (1972).
61. S. Ghodrati, “Investigation of the image processing methods for sur-
face fractal dimension calculation and its relation with surface rough-
ness and organic coatings adhesion,” Master Dissertation (Amirkabir
University of Technology, 2016).
62. ISO, “Geometrical product specifications (GPS)–Surface texture: pro-
file method–Terms, definitions and surface texture parameters,” ISO
4287 (International Organization for Standardization, 1997).
63. ISO, “Geometric product specification (GPS)—surface texture—
profile method: rules and procedures for the assessment of surface tex-
ture,” ISO 4288 (International Organization for Standardization, 1996).
 Research Article Vol. 35, No. 6 / June 2018 / Journal of the Optical Society of America A
64. E. S. Gadelmawla, M. M. Koura, T. M. A. Maksoud, I. M. Elewa, and
H. H. Soliman, “Roughness parameters,” J. Mater. Process. Technol.
123, 133–145 (2002).
65. S. Tyan, “Median filtering: deterministic properties,” in Two-Dimensional
Digital Signal Processing II, T. S. Huang, ed. (Springer-Verlag, 1981),
pp. 197–217.
66. M. Sonka, V. Hlavac, and R. Boyle, Image Processing, Analysis, and
Machine Vision (Cengage Learning, 2014).
67. F. J. Harris, “On the use of windows for harmonic analysis with the
discrete Fourier transform,” Proc. IEEE 66, 51–83 (1978).
68. V. Vapnik, Statistical Learning Theory (Wiley, 1998).
69. W. Liu, X. Tu, Z. Jia, W. Wang, X. Ma, and X. Bi, “An improved surface
roughness measurement method for micro-heterogeneous texture in
1013
deep hole based on gray-level co-occurrence matrix and support vec-
tor machine,” Int. J. Adv. Manuf. Technol. 69, 583–593 (2013).
70. H. Yi, J. Liu, P. Ao, E. Lu, and H. Zhang, “Visual method for measuring
the roughness of a grinding piece based on color indices,” Opt.
Express 24, 17215–17233 (2016).
71. D. Zhang and W. Sui, “The application of AR Model and SVM in rolling
bearings condition monitoring,” in Advanced Research on Computer
Science and Information Engineering, G. Shen and X. Huang, eds.
(Springer, 2011), pp. 326–331.
72. J. D. Huang, L. S. Wang, G. F. Li, X. Z. Zhang, and J. Wang,
“Prediction system of surface roughness based on LS-SVM in
cylindrical longitudinal grinding,” Opt. Precis. Eng. 18, 2407–2412
(2010).