roughness

Precision Engineering 29 (2005) 95–100
Accurate estimation of surface roughness from texture features of the
surface image using an adaptive neuro-fuzzy inference system
Kuang-Chyi Lee a,∗, Shinn-Jang Ho a
, Shinn-Ying Ho b
a
Department of Automation Engineering, National Huwei University of Science and Technology, Huwei, Taiwan 632, ROC
b Department of Information Engineering, Feng China University, Taichung, Taiwan 407, ROC
Received 7 October 2003; received in revised form 15 March 2004; accepted 6 May 2004
Available online 8 July 2004
Abstract
Accurate estimation of surface roughness of workpieces in turning operations play an important role in the manufacturing industry.
This paper proposes a method using an adaptive neuro-fuzzy inference system (ANFIS) to establish the relationship between actual
surface roughness and texture features of the surface image. The accurate modeling of surface roughness can effectively estimate surface
roughness. The input parameters of a training model are spatial frequency, arithmetic mean value, and standard deviation of gray levels
from the surface image, without involving cutting parameters (cutting speed, feed rate, and depth of cut). Experiments demonstrate the
validity and effectiveness of fuzzy neural networks for modeling and estimating surface roughness. Experimental results show that the
proposed ANFIS-based method outperforms the existing polynomial-network-based method in terms of training and test accuracy of
surface roughness.
Keywords: Adaptive neuro-fuzzy inference system (ANFIS); Computer vision; Fuzzy neural network; Modeling; Polynomial network; Surface roughness
1. Introduction
Standard roughness measurement procedures depend
heavily on stylus instruments which have only limited flexi-
bility in handling different parts. Furthermore, the procedure
is a post-process approach, which is applied off-line and
not amenable for automation, and the measurement is also
relatively slow. In recent years, the modeling and prediction
problems of surface roughness of a workpiece by computer
vision in turning operations have received a great deal of
attention [1–8]. Although it has been shown that the surface
roughness is strongly characterized by the surface image,
practical surface roughness instruments based on computer
vision technology are still difficult [1]. The main problem
is how to conveniently and accurately estimate the actual
surface roughness of workpieces.
Lee and Tarng [1] established amodel of surface roughness
based on features of the surface image and its associated cut-
ting operations using a self-organizing adaptive learning tool
called polynomial network (PN) [9]. Without using cutting
parameters, Lee et al. [2] used the same PN-based method to
∗ Corresponding author. Tel.: +886 939 298948.
E-mail address: kclee@nhust.edu.tw (K.-C. Lee).
study themodeling and estimation problems for conveniently
measuring surface roughness in the turning process. As a re-
sult, the PN-based method can effectively estimate the sur-
face roughness with reasonable accuracy. Recently, Ho et al.
[3] dealt with the same problem as that in [1] using a fuzzy
neural network (FNN) approach, and obtained more accurate
results in terms of modeling and prediction accuracy.
In this paper, the FNN approach [10–12] is used to solve
the same problem as that in [2] for modeling and estima-
tion of surface roughness from texture features of the sur-
face image without using cutting parameters. It is known
that the adaptive neuro-fuzzy inference system (ANFIS) is
efficient for non-linear mapping. ANFIS is a fuzzy infer-
ence system implemented in the framework of an adaptive
FNN. By using a hybrid learning procedure, ANFIS can
construct an input–output mapping based on both human
knowledge (in the form of fuzzy if-then rules) and stipulated
input–output data pairs [10,11].In [10,11], it shows that
the ANFIS method is superior to other modeling methods,
such as auto-regressive (AR) model, cascaded-correlation
NN, back-propagation NN, sixth-order polynomial, and
linear predictive methods, in terms of non-dimensional
error index (NDEI), by a chaotic time series prediction
problem.96 K.-C. Lee et al. / Precision Engineering 29 (2005) 95–100
In this paper,we utilizeANFIS to establish the relationship
between actual surface roughness and texture features of the
surface image. An experimental setup for measuring texture
features of surface images can be found in [2]. Consequently,
once the texture features (spatial frequency, arithmetic mean
value, and standard deviation) are given, surface roughness
can be accurately estimated. Using the same database of sur-
face images as that in [2], the absolute arithmetic mean error
and rootmean square error ofmodeling surface roughness are
0.56% and 0.0473m, respectively. The encouraging exper-
imental results show that the proposed ANFIS-based method
outperforms the existing PN-based method [2] in terms of
modeling accuracy. Furthermore, the proposed method can
also improve the estimation accuracy of the PN-basedmethod
on unseen test images.
2. Computer vision system for inspecting surface
roughness
For comparison with the PN-based method [2], the same
database of surface images is used in this study. The com-
puter vision systemfor inspecting surface roughness adopted
by Lee et al. [2] is briefly described as follows. A number
of turning experiments on S55C steel were carried out on a
PC-based CNC lathe (ECOCA PC4610) using a carbide tool
(Mitsubishi TNMG160404R-2G) formachining S55C work-
pieces. The schematic diagram of the computer vision sys-
tem, including a vision system with a digital camera (Olym-
pus C-1400L) and an appropriate lighting arrangement, is
shown in Fig. 1. The camera captures surface images with
1280 × 1024 resolution, 1/30 s grabbing speed, and 8-bit
digit output. The feasible ranges of the cutting parameters are
as follows: cutting speed (53.44–199.49m/min), feed rate
(0.16–0.52mm/rev.), depth of cut (0.5–1.5mm), and actual
average surface roughness Ra (1.87–26.39m). Based on
the cutting parameter combinations, 55 turning experiments
were performed for obtaining the training database.
The actual roughness ofmachined surface ismeasured by a
profile meter (Surfcorder SE-1100, Kosaka Laboratory Ltd.)
within a sampling length of 8mm and measurement speed of
0.5mm/s. The surface roughness Ra is the arithmetic average
Fig. 1. Schematic diagram of the computer vision system.
Fig. 2. Experimental results for texture analysis, gleaned from [2].
of the absolute value of the heights of roughness irregularities
from the mean value measured
Ra = 1
n
n
i=1
|yi| (1)
where yi is the height of roughness irregularities from the
mean value and n is the number of sampling data. An impor-
tant feature of surface image used to estimate surface rough-
ness is the arithmetic average G of gray levels
G = 1
n
n
i=1
|gi| (2)
where gi is the gray level of surface image deviated from
the mean gray value [5]. The image of the workpiece sur-
face captured by the digital camera is shown in Fig. 2(a).
Fig. 2(b) shows the distribution of gray level on the center-line
of Fig. 2(a) and Fig. 2(c) shows the amplitude of gray level
in Fig. 2(b). The details for obtaining G from surface images
can be found in [2]. Then, the spatial frequency (F), the arith-
meticmean value of gray level (G), and the standard deviation
of gray level (S) can be calculated from the amplitude of the
gray level by statistical methods. The database with param-
eters (F, G, S) and actual roughness Ra is listed in Table 1,
gleaned from [2].
Table 1
Comparisons with other modeling methods [10,11]
Methods Training data NDEI
ANFIS 500 0.007
AR model 500 0.19
Cascade-correlation NN 500 0.06
Back-propagation NN 500 0.02
Sixth-order polynomial 500 0.04
Linear predictive method 2000 0.55K.-C. Lee et al. / Precision Engineering 29 (2005) 95–100 97
Fig. 3. The architecture of ANFIS.
3. The proposed ANFIS-based method
3.1. ANFIS Architecture
Let l, m, and n are the numbers of fuzzy sets for input
parameters F, G, and S, respectively. A typical rule base with
lmn fuzzy if-then rules using a first-order Takagi and Sugeno
fuzzy model can be represented as:
Rule k:If F is AFk andGis AGk and S is ASk, then fk = pkF
+ qkG+ rkS+ sk, k= 1, ... , lmn,where AFk, AGk, and ASk
are the antecedent fuzzy sets of F, G, and S, respectively. The
antecedent parameters of l + m + n fuzzy sets and lmn sets
of consequent parameters (pk, qk, rk, sk) are to be optimized
using a five-layerANFIS. The architecture of theANFIS used
in the proposed method is shown in Fig. 3. There are three
input parameters (F, G, S) and one output parameter (the
estimated surface roughness ¯ Ra). Denote the output node k in
the layer l asOl ,k.The solution usingANFIS to the antecedent
and consequent parameters can be found in [10,11], and the
setup of the five-layer network are described below.
Layer 1: Every node in this layer is an adaptive node with
a node output defined
O1,h = µFh (F), h = 1, ..., l,
O1,i+l = µGi
(G), i = 1, ...,m,
O1,j+l+m = µSj
(S), j = 1, ..., n,
(3)
where F (G, S) is the input to the node, Fh (Gi, Sj)isthe hth
(ith, jth) fuzzy set of the membership function of F (G, S).
In other words, the outputs of this layer O1,k, k = 1, ... , l
+ m + n, are the membership values of the antecedent parts.
Here, we choose the membership function to be bell-shape
with maximum equal to 1 and minimum equal to 0, such as
O1,k = µZk
(Z) = 1
1 + [(Z − cZk
)2/a2
Zk
]
bZk
,
k = 1, ..., l + m + n, (4)
where Z∈{F,G, S} is the input parameter, and {aZk
,bZk
,cZk
}
are the antecedent parameters.
Layer 2: Every node in this layer is a fixed node labeled
, which multiplies the incoming signals and outputs the
product as follows
O2,k = wk = µFh (F) × µGi
(G) × µSj
(S),
h = 1, ..., l, i = 1, ...,m, j = 1, ..., n,
k = 1, ..., lmn. (5)
Each node output represents the firing strength of a rule.
Layer 3: Every node in this layer is a fixed node labeled N.
The kth node calculates the normalized firing strength which
is a ratio of the kth rule’s firing strength to the sum of all
rules’ firing strength
O3,k =¯ wk = wk
w1 + w2 +···+ wlmn
,
k = 1, 2, ..., lmn. (6)
Layer 4: Every node k in this layer is an adaptive node with
a node function
O4,k =¯ wkfk =¯ wk(pkF + qkG + rkS + sk),
k = 1, 2, ..., lmn. (7)
Layer 5: The single node in this layer is a fixed node labeled
  that computes the overall output as the summation of all
incoming signals
O5,1 = Ra =
lmn
k=1
¯ wkfk. (8)
3.2. Comparisons with other modeling methods
ANFIS can be applied to a wide range of areas, such as
non-linear function modeling, time series prediction, on-line
parameter identification, and fuzzy controller design. The
time series used for an illustrative example is generated by
the Mackey–Glass differential delay.
˙ x = 0.2x(t − τ)
1 + x10(t − τ)
− 0.1x(t) (9)
The prediction of future values of this time series is a
benchmark problem that has been reported by many re-
searchers [10,11]. The simulation results by ANFIS method
were reported in [10,11]; more details can be found therein.
Table 1, which is from [10,11] directly, lists the result com-
parisons with other modeling methods using a chaotic time
series prediction problem. The non-dimensional error index
(NDEI) is defined as the rootmean square error divided by the
standard deviation of the target series. Table 1 shows that the
ANFISmethod is superior to othermodelingmethods, which
involve auto-regressive (AR) model, cascaded-correlation
NN (neural network), back-propagation NN, sixth-order
polynomial, and linear predictivemethods, in terms ofNDEI.98 K.-C. Lee et al. / Precision Engineering 29 (2005) 95–100
Table 2
Experimental image parameters and surface roughness for training tests
Test
number
Frequency line
mm−1 (F)
Gray level (G) Standard
deviation (S)
Stylus instrument
Ra (m)
Vision estimation
¯ Ra (m)
Error (%)
1 1.92 62.66 69.33 21.790 21.789 0.00
2 1.92 62.18 69.41 23.150 23.152 0.01
3 1.92 56.61 65.03 23.460 23.460 0.00
4 1.92 50.13 56.55 22.440 22.439 0.00
5 1.92 59.82 67.74 25.710 25.710 0.00
6 1.92 74.39 81.83 23.710 23.710 0.00
7 1.92 63.54 71.58 24.720 24.720 0.00
8 1.92 61.34 68.60 26.390 26.389 0.00
9 1.92 57.62 63.97 22.250 22.249 0.00
10 1.92 58.72 65.71 22.330 22.331 0.01
11 1.92 52.88 59.87 22.720 22.721 0.00
12 1.92 48.16 53.99 23.260 23.260 0.00
13 2.50 47.44 52.68 14.420 14.405 0.10
14 2.50 51.50 56.88 14.750 14.761 0.07
15 2.50 51.36 56.89 15.970 15.980 0.06
16 2.50 50.40 55.06 15.080 15.102 0.15
17 2.50 65.25 72.00 15.820 15.820 0.00
18 2.50 47.46 52.96 14.810 14.827 0.11
19 2.50 53.65 59.30 15.130 15.136 0.04
20 2.50 46.76 64.88 15.260 15.260 0.00
21 2.50 59.83 58.20 17.160 17.160 0.00
22 2.50 53.00 58.20 15.050 15.032 0.12
23 2.50 50.01 55.16 17.900 17.866 0.19
24 2.50 45.35 50.46 17.010 17.011 0.00
25 3.44 37.02 41.37 6.595 6.520 1.14
26 3.44 38.65 43.04 6.815 6.812 0.05
27 3.44 36.69 40.59 7.176 7.134 0.58
28 3.44 38.83 43.32 7.273 7.238 0.49
29 3.44 45.17 49.86 7.042 7.042 0.00
30 3.44 41.46 45.83 6.432 6.421 0.17
31 3.44 40.77 45.16 6.713 6.752 0.58
32 3.44 38.85 43.07 7.026 7.202 2.51
33 3.44 42.63 47.49 9.625 9.625 0.00
34 3.44 39.61 43.82 9.573 9.474 1.03
35 3.44 39.95 44.41 8.332 8.318 0.17
36 3.44 35.84 40.12 9.034 9.096 0.68
37 3.85 36.17 40.98 6.284 6.277 0.11
38 3.85 35.16 40.01 5.332 5.347 0.29
39 3.85 32.79 37.53 6.667 6.662 0.08
40 4.35 33.07 37.10 4.911 4.919 0.15
41 4.35 33.13 37.56 4.555 4.546 0.20
42 5.00 28.85 34.45 3.596 3.594 0.05
43 5.00 27.73 31.90 3.797 3.796 0.03
44 5.00 30.48 35.59 3.707 3.711 0.11
45 6.25 19.58 24.70 2.525 2.524 0.05
46 6.25 11.06 13.95 2.331 2.316 0.66
47 6.25 10.60 13.61 2.184 2.191 0.31
48 6.25 8.90 11.65 2.161 2.161 0.01
49 6.25 13.56 16.55 2.167 2.187 0.94
50 6.25 12.44 16.07 1.871 1.978 5.73
51 6.25 12.61 15.81 2.265 2.362 4.26
52 6.25 12.76 16.25 2.385 2.177 8.71
53 6.25 13.69 17.15 2.213 2.208 0.23
54 6.25 11.72 14.22 2.465 2.455 0.40
55 6.25 14.99 19.42 2.545 2.554 0.33
4. Modeling and estimation of surface roughness
The database with 55 turning experiments listed in Table 2
is used to train the ANFIS model with l = 3, m = 4, and
n= 4. The hybrid batch learning rules are used in the training
phase [10–12]. The values of antecedent and consequent pa-
rameters obtained after training are given in the Appendix A.
The trained model using ANFIS establishes the relationshipK.-C. Lee et al. / Precision Engineering 29 (2005) 95–100 99
Table 3
Experimental results using unseen test images
Test
number
Frequency line
mm−1 (F)
Gray level
(G)
Standard
deviation (S)
Stylus instrument
Ra (m)
Vision estimation
¯ Ra (mm)
Error (%)
1 3.91 32.79 37.53 6.667 6.407 3.91
2 3.93 35.16 40.01 5.332 4.951 7.14
3 4.41 33.07 37.10 4.911 4.799 2.27
4 4.52 33.13 37.56 4.555 4.411 3.17
Table 4
Performance comparisons among methods with and without using cutting parameters
Absolute arithmetic mean error Modeling (%) Test (%)
The proposed ANFIS method – pure image 0.56 4.12
The ANFIS method – using cutting parameters [3] 5.88 × 10−8 0.38
The PN method – pure image [2] Above 5 6.2
Fig. 4. Effect of ¯ Ra vs. Ra by the proposed ANFIS-based method.
between actual surface roughness and texture features of the
surface image. Once the texture parameters (F, G, S) are
given, then the estimated surface roughness ¯ Ra can be eas-
ily obtained. The absolute arithmetic mean error and root
mean square error ofmodeling surface roughness using the 55
sets of training data are 0.56% and 0.0473 m. The effect of
the estimated surface roughness ¯ Ra versus the actual surface
roughness Ra using training data by the proposed method is
shown in Fig. 4, and the errors are listed in Table 2. Fig. 5
shows the effect of ¯ Ra versus Ra using the same training
database by the PN-based method, gleaned from [2]. From
Figs. 4 and 5, experimental results show that the proposed
method is superior to the PN-based method in terms of accu-
racy. To further verify the estimation accuracy, several exper-
iments using unseen test images are performed and the results
are listed in Table 3. It is shown that the errors are smaller
than 7.14% and the mean error is only 4.12%. Performance
comparisons among methods with and without using cutting
Fig. 5. Effect of ¯ Ra vs. Ra by the PN-based method [2].
parameters are shown in Table 4. In terms of mean error for
modeling and test on unseen images, the ANFIS-based ap-
proach with using cutting parameters [3] is the best method.
When the cutting parameters are unknown, the proposedAN-
FIS method without using cutting parameters is convenient
to practical applications and can accurately estimate surface
roughness, compared with [2].
5. Conclusion
This paper proposed a method using an adaptive
neuro-fuzzy inference system to accurately establish the
relationship between actual surface roughness and texture
features of the surface image, and consequently can effec-
tively estimate surface roughness. The input parameters used
for a training model are spatial frequency, arithmetic mean
value, and standard deviation of gray levels from the surface100 K.-C. Lee et al. / Precision Engineering 29 (2005) 95–100
Appendix A
See Tables A.1 and A.2.
Table A.1
Values of antecedent parameters using ANFIS
ZF1 F2 F3 G1 G2 G3 G4 S1 S2 S3 S4
aZ 0.0806 0.3983 1.0538 11.0327 10.8124 9.2904 11.4437 11.8574 11.7167 10.8305 11.9153
bZ 2.4039 2.7653 1.7329 −0.343 4.1399 4.8649 2.0981 0.1686 2.3291 1.2904 1.7551
cZ 1.8078 4.1132 6.4814 8.968 31.4635 51.7346 73.8872 11.7078 35.0346 56.7837 81.6527
Table A.2
Values of consequent parameters using ANFIS
k 1 2345 67 89101112
pk 1.171 0.0177 3.149 1.5401 −0.3277 −0.1966 −0.6476 −0.0023 1.9631 −0.0505 4.2779 0.031
qk −16.405 3.9324 15.2496 −15.149 −7.0218 −4.7322 −13.328 0.0418 −0.2996 3.6161 −3.7459 −20.36
rk −7.3354 4.9407 1.2096 17.328 −7.6757 −5.284 −14.49 0.1082 21.103 5.1228 −8.9835 13.331
sk 0.6125 0.0064 1.6424 0.8021 −0.165 −0.1024 −0.3237 −0.0012 1.0229 −0.0268 2.2208 0.0152
i 13 14 15 16 17 18 19 20 21 22 23 24
pi 0.2252 0.0532 0.0504 2.4158 −4.5951 −11.702 −21.117 −0.9404 10.209 −0.1648 −16.266 −0.6459
qi −13.439 1.6642 9.5266 −5.095 −45.612 73.279 −145.42 −7.0828 4.2167 −22.756 −30.386 −0.6475
ri −11.457 1.7139 2.9993 14.904 −16.841 30.476 −62.414 −1.9876 83.726 −60.293 143.87 6.0572
si 0.1178 0.0278 0.0263 1.2584 0.1634 1.9113 −0.6443 −0.0474 3.0663 3.0198 2.9199 0.1181
i 25 26 27 28 29 30 31 32 33 34 35 36
pi 2.6868 −7.6767 5.7717 0.1711 −0.9523 −1.653 −1.2776 −0.0737 −4.2227 −116.73 −57.534 −3.6997
qi 47.358 −23.558 57.134 1.1699 −9.2521 −12.246 −14.177 −1.1084 0.099 49.546 −30.2 −20.441
ri 42.246 −38.18 45.75 3.9726 −9.1933 −12.297 −13.963 −1.1488 3.3225 −6.0922 46.753 35.594
si 0.8029 −2.1354 1.7437 0.0603 −0.1781 −0.2197 −0.2935 −0.0196 9.9769 13.716 4.4541 0.0731
i 37 38 39 40 41 42 43 44 45 46 47 48
pi 43.035 32.325 −95.516 −3.9468 0.4615 −1.353 20.535 2.3445 12.346 −1.9403 −0.3931 −0.3742
qi −105.39 20.582 −30.567 14.924 −123.56 73.662 30.372 8.6191 −4.4328 4.2874 7.0398 −11.49
ri 52.843 −18.63 −0.0545 24.098 −96.709 53.732 25.87 86.549 −2.9877 5.8815 8.9186 −11.915
si 21.118 48.139 −0.3875 0.32 0.0854 −0.1013 8.0256 0.9363 1.8735 0.1697 0.0099 −0.1312
image, without involving cutting parameters (cutting speed,
feed rate, and depth of cut). Experiments demonstrate the
validity and effectiveness of fuzzy neural networks for mod-
eling and estimating surface roughness. Experimental results
show that the proposed ANFIS method outperforms the ex-
isting polynomial-network method in terms of training and
test accuracy of surface roughness. The proposed method is
also useful in computing other surface texture parameters
such as spacing parameters.
References
[1] Lee BY, Tarng YS. Surface roughness inspection by computer vision
in turning operations. Int JMachine ToolsManufacture 2001;41:1251–
63.
[2] Lee BY, Juan H, Yu SF. A study of computer vision for measuring sur-
face roughness in the turning process. Int JAdvManufacturing Technol
2002;19:295–301.
[3] Ho SY, Lee KC, Chen SS, Ho SJ. Accurate modeling and prediction of
surface roughness by computer vision in turning operations using an
adaptive neuro-fuzzy inference system. Int J Machine Tools Manufac-
ture 2002;42:1441–6.
[4] Damodarasamy S, Raman S. Texture analysis using computer vision.
Comput Industry 1991;16:25–34.
[5] Hoy DEP, Yu F. Surface quality assessment using computer vision
methods. J Mater Process Technol 1991;28:265–74.
[6] Galante G, PiacentiniM, Ruisi VF. Surface roughness detection by tool
image processing. Wear 1991;148:211–20.
[7] Al-Kindi GA, Baul RM, Gill KF. An application of machine vision
in the automated inspection of engineering surfaces. Int J Product Res
1992;30(2):241–53.
[8] Kiran MB, Ramamoorthy B, Radhakrishnan V. Evaluation of sur-
face roughness by vision system. Int J Machine Tools Manufacture
1998;38(5):685–90.
[9] MontgomeryGJ,DrakeKC.Abductive reasoning network.Neurocom-
puting 1991;2:97–104.
[10] Roger Jang JS. ANFIS: adaptive-network-based fuzzy inference sys-
tem. IEEE Trans Syst, Man Cybernetics 1993;23(3):665–85.
[11] Roger Jang JS, Sun CT. Neuro-fuzzy modeling and control. Proc IEEE
1995;83(3):378–406.
[12] Shen JC. Fuzzy neural networks for tuning PID controller for plants
with underdamped responses. IEEE Trans Fuzzy Syst 2001;9(2):
333–42.