J Electr Eng Technol.

2016; 11(6): 1872-1879 ISSN(Print) 1975-0102

http://dx.doi.org/10.5370/JEET.2016.11.6.1872 ISSN(Online) 2093-7423

Identification of Plastic Wastes by Using Fuzzy Radial Basis Function

Neural Networks Classifier with Conditional Fuzzy C-Means Clustering

Seok-Beom Roh* and Sung-Kwun Oh†

Abstract – The techniques to recycle and reuse plastics attract public attention. These public
attraction and needs result in improving the recycling technique. However, the identification technique
for black plastic wastes still have big problem that the spectrum extracted from near infrared radiation
spectroscopy is not clear and is contaminated by noise. To overcome this problem, we apply Raman
spectroscopy to extract a clear spectrum of plastic material. In addition, to improve the classification
ability of fuzzy Radial Basis Function Neural Networks, we apply supervised learning based clustering
method instead of unsupervised clustering method. The conditional fuzzy C-Means clustering method,
which is a kind of supervised learning based clustering algorithms, is used to determine the location of
radial basis functions. The conditional fuzzy C-Means clustering analyzes the data distribution over
input space under the supervision of auxiliary information. The auxiliary information is defined by
using k Nearest Neighbor approach.

Keywords: Identification of plastic wastes, Conditional fuzzy C-means clustering, k nearest neighbor
approach, Fuzzy radial basis function neural networks

1. Introduction each material. NIR spectroscopy equipment is applicable

to classify and identify plastic material whose color is
These days, one key issue of mankind is the alteration of not black. However, when classifying and identifying
global weather patterns. The alteration of weather patterns black plastic which contains carbon black material, NIR
is assumed to be caused by global warming. Many spectroscopy equipment is not able to be used for a sensor.
industrialized countries are trying to slow down global Because black colored materials can absorb light of all
warming by reducing carbon dioxide emission. One way to wave length. In this paper, to overcome the drawback of
reduce carbon dioxide emission as well as the pace of NIR spectroscopy, Raman spectroscopy technique is used.
energy depletion is the waste recycling and reuse. Nowadays, Raman spectroscopy technique is used in
Among the various municipal wastes generated every very wide applications such as pharmaceuticals, forensics,
day around world, plastic is one of major material. It is archeology, art, medicine, semiconductors in micro-
natural that the techniques to recycle and to reuse plastic electronics, geology, planetary science and materials science
waste come into the spotlight. The various techniques to [3]. It is well known that Raman spectroscopy is a
recycle and reuse plastic waste have been studied. The key spectroscopic technique exploiting the inelastic scattering
issue in recycling and reusing plastic waste is the phenomenon of monochromatic light, usually from a laser
identification plastic material. When recycling and reusing source [4]. Inelastic scattering means that the frequency of
plastic waste, the purity of the recycled plastic is very photons in monochromatic light from a light source
important. To obtain homogeneous plastic material, the changes by interacting with a material.
identification technique is need. In addition, in order to improve identification perfor-
In 1995, Scott proposed an identification technique to mance of the identification system to classify some plastics
identify polyethylene terephthalate (PET) and polyvinyl into several categories, intelligent techniques such as fuzzy
chloride (PVC) resins [1]. The identification technique set theory and neural networks are used. Especially, fuzzy
proposed by Scott was based on Near Infrared Radiation Radial Basis Function Neural Networks (fRBFNNs) are
(NIR) to extract absorbance spectra of each plastic material. used as a classifier to identify black plastic wastes. In
In addition, a method to classify plastic resins into PET general, it is well known that these intelligent techniques
and PVC categories was proposed by Edward and have same advantages of dealing with the uncertain data
Sommer in 2001 [2]. The classification system invented by and having learning ability. The locations of radial basis
Edward Sommer was based on NIR transmission spectra of functions are very important to improve the classification
† Corresponding Author: Dept. of Electrical Engineering, The performance of fRBFNNs. The locations of radial basis
University of Suwon, Korea. ([email protected]) function are determined by using unsupervised clustering
* Dept. of Electrical Engineering, The University of Suwon, Korea. technique (usually fuzzy C-means clustering method).
([email protected])
Received: March 7, 2016; Accepted: June 13, 2016
However, the unsupervised clustering method does not use

1872
Copyright ⓒ The Korean Institute of Electrical Engineers
This is an Open-Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/
licenses/by-nc/3.0/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.
 Seok-Beom Roh and Sung-Kwun Oh

the auxiliary information that can be extracted the output 2

m
(al - bl ) 2
space but the data distribution information extracted over a-b =å (3)
l =1 s l2
the input space [5]. We use conditional fuzzy C-means
clustering (cFCM) algorithm to define radial basis
Here, sl2 is the variance of the l-th variable, l =1,2,… , m.
functions in supervised learning manner. The auxiliary
In fuzzy clustering procedure, the objective function is
information over the output space is defined by fuzzy k-
minimized by sequential optimization operation which
Nearest Neighbors (kNN) approach.
optimizes the partition matrix and the associated prototypes.
In this paper, we make some experiments with several
Partition matrix W: wik is the i-th row, j-th column
machine learning standard datasets to verify the proposed
element of partition matrix W
intelligent identification system. Finally, we use the
proposed identification algorithm to classify black plastics
into several material categories. 1
wik = 2/(p -1)
c
æ ||x k - v i || ö (4)
å ç ÷
l=1 è ||x k - v l || ø
2. Fuzzy C-Means Clustering and Conditional
Fuzzy C-Means Clustering Prototypes v1, v2, …, vc

The location of radial basis functions is usually defined N

p
by using unsupervised method that analyzes the data åw
k =1
ik xk
distribution on the input space. In some literatures, this step vi = N (5)
to analyze the data distribution can be considered as the åw p
ik
granulation of information [6-8]. The information granules k=1

are extracted by analyzing numerical data distribution and

other source of experimental evidence. In this study, we The partition matrix and the prototypes are updated in an
use here Fuzzy C-Means clustering algorithm to extract iterative fashion and the iterative optimization process
fuzzy set. keeps proceeding until an already determined certain
termination criterion has been satisfied.
2.1 Fuzzy C-means clustering algorithms
2.2 Conditional fuzzy c-means clustering algorithms
A brief explanation of the fuzzy clustering algorithm is
outlined below. In order to determine the center points of radial basis
Let us consider a finite set of data X = {x1 , x 2 , L , x N } , functions of fRBFNNs, unsupervised clustering algorithms
x k Î Âm , 1 £ k £ N . The FCM clustering should optimize are generally used. As mentioned before, conventional
the following objective function. unsupervised clustering methods cannot use the whole
information involved in the given data patterns. Conventional
c N
2 FCM algorithm is known as a sort of unsupervised learning
J p = åå ( wik ) p x k - v i , 1< p < ¥ (1)
i =1 k =1
algorithms. The supervised learning version of conventional
FCM is conditional fuzzy C-means clustering algorithm
The above optimization is carried out subject to the (cFCM) that was proposed by Pedrycz [9]. cFCM can
following constraints. analyze input space under supervision of other auxiliary
information that can be defined by a designer or from the
c experimental evidence.
åw
i =1
ik =1 (2-1) In this study, we apply cFCM to define the center points
of radial basis functions. When applying cFCM to analyze
N
0 < å wik < N the data distribution, the input data pattern x k Î Âm are
(2-2)
k =1 grouped by taking into account the auxiliary information
( f1 , f 2 , L , f N ). While the objective function for cFCM is
Here, x k - v i is any distance between the data x k the same as (1), the constraints are different from (2). The
and the prototype v i . The parameter “p” used in the constraint for cFCM is as follows.
objective function denotes the fuzzification coefficient.
c
“N” and “c” mean the number of data patterns and clusters,
respectively. Usually, the fuzzification coefficient is åw i =1
ik = fk (6)
equal to 2. The distance between data and prototypes is
considered as the generalized weighted version of Euclidean The optimization problem to minimize (1) subject to a
distance as (3). constraint (6) is expressed as (7).

http://www.jeet.or.kr │ 1873
Identification of Plastic Wastes by Using Fuzzy Radial Basis Function Neural Networks Classifier with Conditional Fuzzy C-Means Clustering

min J subject to
c
3. Radial Basis Function Neural Networks based
W, v åwi =1
ik = fk (7)
on Auxiliary Information

The optimal solution that minimizes the objective On the contrary to the general definition methods for
function (1) subject to (6) can be obtained by iterative the locations of RBFs which are based on the unsupervised
update formulas such as (8) and (9). clustering algorithm, we use cFCM to determine the
locations of RBFs under the supervision of the auxiliary
fk information. The auxiliary information includes the
wik = 2 / (p -1 ) information of the class label of the given data point.
C æ || x k - v i || ö (8)
å çç ÷÷ This statement means that we can use the information,
j =1 è || x k - v j || ø which the given data patterns represent over input space an
output space, as much as we can use.
N
p In some literatures [10, 11], it is said that the RBFNNs
å (u )
k =1
ik × xk
have some advantages including global optimal approxi-
vi = N (9)
p mation and classification capabilities as well as rapid
å (u )
k =1
ik
convergence of the underlying learning procedures.
The generic topology of RBF NNs is depicted in Fig. 1.
In Fig. 1, Gi , i=1, 2,…, c denote receptive fields (radial
2.3 k-Nearest neighbor approach for auxiliary basis functions), while the parameter “m” denotes the
information number of input variables. The output of the generic RBF
NN is expressed as a linear combination of the outputs
To apply the above explained cFCM, the already defined ( G(x) ) of the corresponding hidden nodes with the
auxiliary information should be needed. We use k Nearest connection weights w1 , w2 , L , wc as (12).
Neighbor approach to extract the auxiliary information
which is defined on the output space. We can say that the c

data point, whose nearest neighbors are involved in the same yˆ(x) = å wi × Gi (x) (12)
i =1
class, has homogeneous characteristic. The heterogeneous
data point is defined as the data point whose nearest
Here, G(x) means the activation level of i-th hidden
neighbors are involved in different classes.
node.
In order to build homogeneous clusters over the input
We expand the generic RBFNNs by adopting cFCM to
space, the homogeneous data points should be considered
locate RBFs with the aids of auxiliary information and
more importantly in comparison with the heterogeneous
applying polynomials to connection weights. In the
data points. Therefore, we define the auxiliary information
expanded version of RBFNNs, the auxiliary information
of a given data point as how much homogeneous the data
extracted by kNN help cFCM determine the location of
point is. In pattern recognition fields, kNN approach is
RBFs in the viewpoint of the homogeneity of the data
considered as a kind of non-parametric method to directly
points involved in the RBFs. We assume that in order to
classify the given data point. On the contrary, we use kNN
improve the identification performance, the data points
approach to extract the auxiliary information.
which are involved in a cluster defined by a RBF should be
We define the auxiliary information as (10).
homogeneous. To obtain the homogeneous clusters, we
nij
f ij = (10)
K

Here, f ij denotes the auxiliary information of the j-th

data point x j on the i-th class, K is the number of the
nearest neighbors of the given data point, and nij means
the number of the nearest neighbors of the given data
which are involved in i-th class. The auxiliary information
defined by (10) should satisfy (11).

L L nij
åf
j =1
ij =å
j =1 K
=1 (11)

Here, L means the number of classes. Fig. 1. General architecture of the generic RBF neural
networks

1874 │ J Electr Eng Technol.2016; 11(6): 1872-1879

 Seok-Beom Roh and Sung-Kwun Oh

Fig. 3. Integrated raman spectrometer produced by ocean

optics

Fig. 2. An overall topology of radial basis function neural

networks with auxiliary information

apply cFCM with the auxiliary information obtained by

Fig. 4. Spectra of plastics such as PET, PP, and PS obtained
using kNN. The auxiliary information through kNN means
from RAMAN spectrometer
how much homogeneous the data points are.
The overall topology of the Radial Basis Function
Neural Networks with the auxiliary information is shown scattered will commune with the vibrational and rotational
in Fig. 2. states of a sample molecule and emit light of a slightly
We use three types of polynomials as the consequent different wavelength.
functions of the proposed RBF neural networks as follows. The Raman spectroscopy equipment used in this experi-
ment is produced by Ocean Optics (www.oceanoptics.com).
Fj = a j (13-1) The Raman equipment is shown in Fig. 3.
The light source of the Raman spectroscopy equipment
Fj = a j 0 + a j1 x1 + L + a jm xm (13-2) is 785nm diode laser. The spectrum acquired from the
m m
above mentioned Raman spectroscopy of each plastic
Fj = a j 0 + å a jk xk + å a j , m + k xk2 (13-3)
k =1 k =1 material is shown in Fig. 4. The raw spectrum obtained
from Raman spectrometer should be preprocessed to be
The output of RBF neural networks is defined as (14). used as the input of the intelligent classifier (in this study,
fRBFNNs).
c´ L We should determine at which wave number the
åw
i =1
ik fi (x k ) characteristic peaks happens according plastic materials
yˆ(x k ) = c´ L (14) such as PET, PP, and PS. The information of characteristic
å wik peaks of each plastic material can be obtained from the
i =1 field experts. In this study, we do not try to depend on the
field experts’ knowledge but to extract the information
from the raw spectra obtained from Raman spectroscopy.
4. Extraction of Input Variables from Raman In order to distinguish a plastic material from the other
Spectrum plastic materials, we extract unique wave number where
the unique peaks of a specific plastic material happen. We
For implementing the identification system of black compare a spectrum of a specific plastic material with the
plastic, Raman spectroscopy is used as a sensor to obtain spectra of the other plastic materials. In this comparison
the specific data pattern from a black plastic material. operation, the “XOR” operator is used and is defined as
Generally, Raman spectroscopy investigates materials (15).
not through direct absorption, but by scattering of high
intensity light in the hopes that one in a million photons XOR (S, T) = (1 - S) Ä T + S Ä (1 - T) (15)

http://www.jeet.or.kr │ 1875
Identification of Plastic Wastes by Using Fuzzy Radial Basis Function Neural Networks Classifier with Conditional Fuzzy C-Means Clustering

Here, Ä means the element-wise product. The result of “ Ä ” operation of two spectra is calculated as
T
For example, let the spectrum A and B be A Ä B = éë a1 × b1 a2 × b2 L aq × bq ùû .
T T
A = éë a1 a2 L aq ùû and B = éëb1 b2 L bq ùû respectively. The results of XOR operation between two different
plastic materials are shown in Fig. 5.
The element-wise operation version of (15) is
xor (a, b) = (1 - a ) × b + a × (1 - b) .
In the case of a = 0 & b = 1 or a = 1 & b = 0 , the
output of the operation is maximized. In other words,
when the difference between two spectra is big, the
output of the XOR operation becomes big. We can
consider the 15 maximum points of the result of the
“XOR” operation between three different plastic materials
as the characteristic peaks which are shown in only one
plastic material.
In the plastic material identification experiment, we
(a) XOR operation between PET and PP obtain 100 raw spectra for each plastic material and total
300 row spectra (100 PET, 100 PP, and 100 PS). From this
“XOR” operation, we can estimate 15 characteristic peaks
as shown in Table 1.

5. Experimental Studies

We make several experiments in order to verify the

proposed design method based on k-NN and C-FCM to
implement RBF neural networks. In these experiments,
we use a series of machine learning datasets (http://www.
ics.uci.edu/~mlearn/MLRepository.html).
(b) XOR operation between PET and PS In this experiment, we use classification rate to evaluate
the proposed classifier and 10-Fold cross-validation to
evaluate the classification result.
In addition, to verify the black plastic identification
performance of the proposed intelligent classifier, we use
black plastic raw spectra from Raman spectroscopy
equipment. At first, we should predefine some design
parameters whose values are summarized in Table 2.
Table 3 summarizes the pertinent details of each dataset

Table 1. The characteristic peaks estimated by “XOR”

operation
(c) XOR operation between PP and PS
Wave Wave Wave
No. No No
Number Number Number
1 493 6 853 11 1612
2 640 7 999 12 1656
3 706 8 1282 13 1725
4 724 9 1367 14 2330
5 746 10 1553 15 2876

Table 2. The selected numerical values of the parameters

of the proposed classification technique
Parameter Value
Number of Rules in each class(C) 2, 3, 4, 5
Number of Nearest Neighbors(K) 3, 5, 7, 10, 20, 30
(d) XOR operation between three types of plastic materials Fuzzification Coefficient (p) 1.2, 1.5, 2.0, 2.5, 3.0
Fig. 5. XOR operation result between plastic Polynomial Type (T) 1, 2, 3

1876 │ J Electr Eng Technol.2016; 11(6): 1872-1879

 Seok-Beom Roh and Sung-Kwun Oh

Table 3. Machine learning datasets used in the experiments

Number Number Number
Data set
of Classes of features of patterns
Australian 2 42 690
Balance 3 4 625
Bupa 2 6 347
German 2 24 1000
Glass 6 9 214
Hayes 3 5 132
Ionosphere 2 34 351
Iris 3 4 150
Pima 2 8 768
Sonar 2 60 208
Thyroid 3 5 215 (a) PET
Vehicle 4 18 846
Wine 3 13 178
Zoo 7 16 101

Table 4. Classification performance of the proposed

classifier and generic RBF neural networks
Generic RBF NN Proposed RBF NN
Data Sets Training Data Test Data Training Data Test Data
mean STD mean STD mean STD mean STD
Australian 84.36 1.24 85.65 4.61 86.23 0.81 86.38 4.33
Balance 89.17 0.94 88.16 3.72 79.11 0.63 78.87 4.91
Bupa 59.91 2.67 57.39 9.49 67.34 1.01 62.95 7.27
German 67.07 0.80 67.00 3.40 69.71 0.62 70.00 4.57
Glass 66.82 1.88 60.76 11.70 65.99 0.98 62.62 8.22 (b) PP
Hayes 54.21 4.15 53.19 10.26 74.24 3.47 66.87 10.95
Ionosphere 91.50 1.13 89.84 4.79 90.63 1.16 90.98 4.87
Iris 95.85 1.12 96.00 7.17 96.37 0.89 97.33 3.44
Pima 69.78 1.28 70.05 4.14 74.65 0.61 73.69 3.64
Sonar 75.75 1.20 75.00 7.50 84.83 1.46 81.71 7.21
Thyroid 95.92 0.90 97.66 3.33 96.49 0.41 97.21 3.22
Vehicle 59.26 1.38 57.68 3.92 69.33 0.75 67.13 4.47
Wine 97.13 0.73 97.78 3.88 95.57 0.55 97.22 3.93
Zoo 99.23 1.05 96.00 6.99 97.47 0.90 93.00 8.23

Table 5. The identification performance of the proposed

plastic identification system
C Training Data Test Data (C) PS
m R p K
C1 C2 C3 mean STD mean STD
7 5 2.0 10 4 4 4 86.50 2.35 86.0 3.65
Fig. 6. Raman spectra of sampled plastics such (a) PET, (b)
7 7 2.0 3 4 3 4 87.75 1.20 87.0 2.17 PP, and (C) PS
9 5 1.5 7 3 3 4 86.92 1.00 85.67 3.46
9 7 1.5 7 3 4 4 86.75 2.07 84.67 2.17
spectroscopy such as PP, PS, and PET are shown.
15 5 3.0 10 4 3 3 92.92 1.21 93.33 6.56
15 7 3.0 7 2 3 4 93.25 0.62 93.33 3.53
In this black plastic identification experiment, the constant
20 5 3.0 10 2 4 2 90.08 1.68 90.67 4.80 is used as the output polynomial used in fRBFNNs classifier.
20 7 3.0 3 3 3 3 90.83 1.77 91.0 4.35 Because the higher order polynomial may result in the
over-fitting problem.
Table 5 summarizes the identification performance
such as the number of classes, number of features, and (expressed %) of the developed neural networks with
number of patterns. respect to the number of selected variables (m), the number
Table 4 summarizes the classification performance (%) of nearest neighbors (K), the range of the input variables
of the proposed neural networks and the generic RBF (R), the number of radial basis functions (C), and the
neural networks. From the experimental results shown in fuzzification coefficient (p). The best result of the proposed
Table 4, the proposed RBFNNs are generally superior to identification system is the RBFNN with 9 RBFs, 3.0
the generic RBFNNs in the majority cases of data sets in fuzzification coefficient. The prediction performance of the
terms of the prediction capability. best RBFNN is 93.33%.
In Fig. 6, the spectra of plastics obtained from Raman In order to validate the identification ability of the

http://www.jeet.or.kr │ 1877
Identification of Plastic Wastes by Using Fuzzy Radial Basis Function Neural Networks Classifier with Conditional Fuzzy C-Means Clustering

Table 6. Results of comparative analysis Acknowledgements

Parameters Training Data Test Data
Proposed R=7, m =15, This study was supported by Basic Science Research
93.25±0.62 93.33±3.53
Classifier p=3.0, K=7 Program through the National Research Foundation of Korea
Bayes
R=5, m=7 N/A 77.60 (NRF) funded by the Ministry of Science, ICT & Future
Networks
R=5, m=7 Planning [grant number NRF-2015R1A2A1A15055365] and
N/A 91.37
SVM
(Polynomial Kernel) also by the GRRC program of Gyeonggi province [grant
R=5, m=7 number GRRC Suwon 2016-B2, Centre for U-city Security
N/A 82.10
(RBF kernel)
R=5, m=7, k=1 N/A 85.50
& Surveillance Technology]
k-NN
Approach R=5, m=7, k=3 N/A 87.73
AdaBoost R=5, m=7 N/A 82.17
PART R=5, m=7 N/A 81.77 References
J48 R=5, m=7 N/A 80.83
Random
Tree
R=5, m=7 N/A 79.93 [1] Scott D. M., “A two color near-infrared sensor for
sorting recycled plastic waste,” Measurement Science
and Technology, vol. 6, pp. 156-159, 1995.
[2] Edward, J. and Sommer J. R., “Method and
proposed classifier, the identification performance of the
Apparatus for Near Infrared Sorting of Recycled
proposed classifier is compared with that of the already
Plastic Waste,” Patent in United States, Pub. No.: US
studied classifiers such as Bayesian Networks, SVM, k-
2001/0045518 A1, 2001.
Nearest Neighbors Approach, AdaBoost, PART, J48, and
[3] Sorely J. Cocjrane and Jordana Blacksberg, “A Fast
Random Tree. Weka was used to obtain identification
Classification Scheme in Raman Spectroscopy for the
performances of these already studied classifiers. Weka is
identification of Mineral Mixtures Using a Large
the well-known machine learning tool which is a collection
Database with Correlated Predictors,” IEEE Trans.
of machine learning algorithms and can be downloaded at
On Geoscience and Remote Sensing, vol. 53, No. 8,
the website www.cs.waikato.ac.nz. In this experiment, we
pp. 4259-4274, Aug. 2015.
use 10-fold cross validation to evaluate the identification
[4] E. Smith and G. Dent, “Modern Raman Spectroscopy-
performance. Table 6 shows the comparison between the
A Practical Approach,” Hoboken, NJ, USA: Wiley,
proposed classifier with the other well-known classification
2005.
algorithms which are available in the Weka machine
[5] L. Zhang, K. Li, H. He, and G. W. Irwin, “A New
learning package. From the experimental results described
Discrete-Continuous Algorithm for Radial Basis
in Table 6, the proposed identification system is said to be
Function Networks Construction,” IEEE Trans. On
superior to the other machine learning algorithms in terms
Neural Networks and Learning Systems, vol. 24, no.
of the prediction ability.
11, pp. 1785-1798, 2013.
[6] W. Pedrycz, R. Al-Hmouz, and A. S. Balamash,
“Hierarchical Granular Clustering: An Emergence of
6. Conclusion
Information Granules of Higher Type and Higher
Order,” IEEE Trans. On Fuzzy Systems, vol. 23, no. 6,
In this paper, we proposed the new identification system pp. 2270-2283, 2015.
for black plastic based on a kind of intelligent system and [7] A. Gacek and W. Pedrycz, “Clustering Granular Data
Raman spectroscopy equipment. Near infrared radiation and Their Characterization with Information
spectroscopy is usually used to extract spectrum of plastic Granules of Higher Type,” IEEE Trans. On Fuzzy
material. However, when dealing with black plastic, near Systems, vol. 23, no. 4, pp. 850-860, 2015.
infrared radiation spectroscopy cannot be used. In order [8] X. Q. Tang and P. Zhu, “Hierarchical Clustering
to extract the raw spectrum from black plastic, we use Problems and Analysis of Fuzzy Proximity Relation
Rama spectroscopy. In addition, to improve the classification on Granular Space,” IEEE Trans. On Fuzzy Systems,
ability of the fRBFNNs, we apply a supervised clustering vol. 21, no. 5, pp. 814-824, 2013.
algorithm not an unsupervised learning method to [9] W. Pedrycz, “Conditional fuzzy C-Means,” Pattern
determine the parameters of the hidden layers with aids Recognition Letters, vol.17, no,6, pp.625-632, 1996.
of the auxiliary information. This auxiliary information [10] M.J. Er, S.Q. Wu, J.W. Lu, and H.L. Toh, “Face
involves the information which can be extracted over recognition with radical basis function (RBF) neural
output space. From the experimental results, we can say networks,” IEEE Transactions on Neural Networks,
that the auxiliary information can help improve the vol. 13, no. 5, pp. 697-710, 2002.
classification ability of RBF neural networks. [11] X.Y. Jing, Y.F. Yao, D. Zhang, J.Y. Yang, and M. Li,
“Face and palm print pixel level fusion and Kernel

1878 │ J Electr Eng Technol.2016; 11(6): 1872-1879

 Seok-Beom Roh and Sung-Kwun Oh

DCV-RBF classifier for small sample biometric

recognition,” Pattern Recognition, vol.40, pp.3209 -
3224, 2007.
[12] S.-B. Roh, S.-K Oh, and W. Z. Choi, “Design of
fuzzy radial basis function neural networks classifier
based on conditional fuzzy C-means clustering
algorithm,” in conference of ICMIT 2015, pp.67 -70,
2015.

Seok-Beom Roh received the B.Sc.,

M.Sc., and Ph.D. degrees in control
and instrumentation engineering from
Wonkwang University, Korea, in 1994,
1996, and 2006 respectively. He is
currently a Research Professor in the
University of Suwon. His research
interests include fuzzy set, neural net-
works, evolutionary algorithms, computational intelligence
and statistical learning.

Sung-Kwun Oh He received the B.Sc.,

M.Sc., and Ph.D. degrees in electrical
engineering from Yonsei University,
Seoul, Korea, in 1981, 1983, and 1993,
respectively. During 1983-1989, he
was a Senior Researcher of R&D Lab.
of Lucky-Goldstar Industrial Systems
Co., Ltd. From 1996 to 1997, he was a
Postdoctoral Fellow with the Department of Electrical and
Computer Engineering, University of Manitoba, Winnipeg,
MB, Canada. He is currently a Professor with the
Department of Electrical Engineering, The University of
Suwon, Suwon, South Korea. His research interests include
fuzzy system, fuzzy-neural networks, automation systems,
advanced computational intelligence, and intelligent
control. He currently serves as an Associate Editor of the
KIEE Transactions on Systems and Control, International
Journal of Fuzzy Logic and Intelligent Systems of the KFIS,
and Information Sciences.

http://www.jeet.or.kr │ 1879