This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 1

Superpixel-based Fast Fuzzy C-Means Clustering

for Color Image Segmentation
Tao Lei, Member, IEEE, Xiaohong Jia, Yanning Zhang, Senior Member, IEEE, Shigang Liu, Hongy-
ing Meng, Senior Member, IEEE, and Asoke K. Nandi, Fellow, IEEE

Abstract—A great number of improved fuzzy c-means (FCM) I. I NTRODUCTION

clustering algorithms have been widely used for grayscale and
color image segmentation. However, most of them are time-
consuming and unable to provide desired segmentation results
for color images due to two reasons. The first one is that
I MAGE segmentation is a key step of object recognition
and classification in computer vision. Although a large
number of algorithms used for image segmentation have been
the incorporation of local spatial information often causes a
high computational complexity due to the repeated distance proposed, image segmentation remains one of the most chal-
computation between clustering centers and pixels within a local lenging research topics because none of them is able to provide
neighboring window. The other one is that a regular neighboring a unified framework for achieving fast and effective image
window usually breaks up the real local spatial structure of segmentation. The difficulty of image segmentation can be
images and thus leads to a poor segmentation. In this work,
we propose a superpixel-based fast FCM clustering algorithm attributed to two reasons. The first is that image segmentation
(SFFCM) that is significantly faster and more robust than state- is a multiple solution problem, i.e., there are multiple best seg-
of-the-art clustering algorithms for color image segmentation. To mentation results for one image. The second is that an image
obtain better local spatial neighborhoods, we firstly define a mul- is always complex because of noise, background, low signal-
tiscale morphological gradient reconstruction (MMGR) operation to-noise ratio, and intensity nonuniformity. Consequently, it
to obtain a superpixel image with accurate contour. In contrast
to traditional neighboring window of fixed size and shape, is difficult to propose a general segmentation framework to
the superpixel image provides better adaptive and irregular achieve complex image segmentation tasks.
local spatial neighborhoods that are helpful for improving color Image segmentation algorithms can be roughly grouped
image segmentation. Secondly, based on the obtained superpixel into two categories - unsupervised and supervised image
image, the original color image is simplified efficiently and its
segmentation. Unsupervised approaches, such as clustering
histogram is computed easily by counting the number of pixels
in each region of the superpixel image. Finally, we implement [1], [2], GraphCut [3], active contour model [4], watershed
FCM with histogram parameter on the superpixel image to transform (WT) [5], hidden Markov random field (HMRF)
obtain the final segmentation result. Experiments performed on [6], fuzzy entropy [7], etc. are useful and popular due to their
synthetic images and real images demonstrate that the proposed simplicity without depending on training samples and labels.
algorithm provides better segmentation results and takes less
In contrast to unsupervised image segmentation approaches,
time than state-of-the-art clustering algorithms for color image
segmentation. although some supervised approaches such as convolutional
neural network (CNN) [8] and fully convolution networks
Index Terms—Color image segmentation, fuzzy c-means (FCN) [9], are able to achieve image segmentation by using
(FCM) clustering, superpixel, morphological reconstruction.
feature learning, but they require a lot of training samples
and label images. In addition, the segmentation result has a
This work was supported in part by the National Natural Science Foundation coarse contour since CNN and FCN essentially achieve image
of China under Grant 61461025, Grant 61871259,Grant 61811530325 (NSFC-
RC), Grant 61672333, Grant 61873155, in part by China Postdoctoral Science
classification. In this paper, we mainly discuss unsupervised
Foundation under Grant 2016M602856, and in part by the National Science image segmentation.
Foundation of Shanghai under grant 16JC1401300. In unsupervised algorithms, clustering represents one kind
T. Lei is with the School of Electronical and Information Engineering,
Shaanxi University of Science and Technology, Xi’an 710021, P. R. China, of important and popular algorithms for grayscale and color
and also with the School of Computer Science, Northwestern Polytechnical image segmentation because it is suitable and useful for
University, Xi’an 710072, China. (E-mail: [email protected]) both low- and high-dimensional data. Generally, clustering
X. Jia is with the School of Electronical and Information Engineering,
Shaanxi University of Science and Technology, Xi’an 710021, China. (E-mail: algorithms can be roughly categorized into three groups -
[email protected]) minimizing an objective function [10], decomposing a density
Y. Zhang is with the School of Computer Science, Northwestern Polytech- function [11], and graph theory [12]. In this paper, we will fo-
nical University, Xi’an 710072, China. (E-mail: [email protected])
S. Liu is with with the School of Computer Science, Shaanxi Normal cus on image segmentation based on clustering by minimizing
University, Xi’an 710119, China. (E-mail: [email protected]) an objective function. It is well-known that k-means and FCM
H. Meng is with the Department of Electronic and Computer Engineering, are clustering algorithms by minimizing an objective function.
Brunel University London, Uxbridge, Middlesex, UB8 3PH, U.K. (E-mail:
[email protected]) Because k-means is a hard clustering algorithm, it is sensitive
A. K. Nandi is with the Department of Electronic and Computer Engineer- to initial clustering centers or membership. In contrast, FCM
ing, Brunel University London, Uxbridge, Middlesex, UB8 3PH, U.K., and is a soft algorithm that improves the shortcomings of k-
also the Key Laboratory of Embedded Systems and Service Computing, Col-
lege of Electronic and Information Engineering, Tongji University, Shanghai means at the cost of increasing iterations. However, both
200092, China. (E-mail: [email protected]) k-means and FCM are sensitive to noise because the local

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 2

spatial information of pixels is missed for image segmentation. a better segmentation result than state-of-the-art algorithms.
To address this shortcoming, a great number of improved Nevertheless, the FRFCM requires much execution time for
clustering algorithms that incorporate local spatial information color image segmentation because it is difficult to compute
into their objective function, have been proposed in recent the histogram of color images. To address the issue, we
years [13]-[15]. These algorithms can be grouped into two propose a superpixel-based fast FCM (SFFCM) for color
groups. The first group employs neighborhood information image segmentation. The proposed algorithm is able to achieve
of a center pixel using a window of fixed size to improve color image segmentation with a very low computational cost,
image segmentation effect, e.g., FCM algorithm with spa- yet achieve a high segmentation precision.
tial constraints (FCM S) [16], FCM S1 [17], FCM S2 [17], Two contributions are presented:
fast generalized FCM algorithm (FGFCM) [18], fuzzy lo- • We present a multiscale morphological gradient recon-
cal information c-means clustering algorithm (FLICM) [19], struction (MMGR) operation to generate superpixel im-
neighborhood weighted FCM clustering algorithm (NWFCM) age with accurate boundaries, which is helpful for in-
[20], FCM algorithm based on noise detection (NDFCM) tegrating adaptive neighboring information and reducing
[21], Memon’s algorithm [22], and the FLICM based on the number of different pixels in a color image.
kernel metric and weighted fuzzy factor (KWFLICM) [23]. • Based on a superpixel image obtained by MMGR, we
The advantage of these algorithms is that the neighborhood propose a simple color histogram computational method
information can be computed in advance, except FCM S and that can be used to achieve a fast FCM algorithm for
FLICM, to reduce the computational complexity. However, a color image segmentation.
neighborhood window of fixed size and shape is unable to The rest of this paper is organized as follows. In Section
satisfy the requirement of robust image segmentation. The II, we illustrate the motivation of our work. In Section III, we
second group employs adaptive neighborhood information propose our model and analyze its superiority. The experimen-
instead of the window of fixed size and shape, e.g., Liu’s tal results on synthetic images and real images are described
algorithm [24], Bai’s algorithm [25], and adaptive FLICM in Section IV. Finally, we present our conclusion in Section
(ADFLICM) [26]. As adaptive neighborhood information is V.
consistent with real image structuring information, the second
group of algorithms obtains a better robustness for noisy II. M OTIVATION
images and a better segmentation effect than the first group.
FCM often miss spatial information leading to a poor
Though improved FCM algorithms consider the neighbor-
result for image segmentation. Although a great number of
hood information of an image, the neighborhood information
improved algorithms address the problem by incorporating
of the corresponding membership that is helpful for improving
local spatial information into the objective function, this, in
classification effect, is ignored. HMRF [27]-[29] is a popu-
turn, increases the computational complexity of algorithms.
lar algorithm for addressing the issue. In [30], the current
Fortunately, superpixel [34] is able to address the problem.
membership called posterior probability depends on clustering
Superpixel is an image pre-processing tool that over-segments
centers and the prior probability of neighborhood. Because
an image into a number of small regions. A superpixel region
HMRF considers the previous state of current membership, it
is usually defined as perceptually uniform and homogenous
obtains better result than FCM for image segmentation [30].
regions in the image [35]. Superpixel is able to improve the
Based on the idea, Zhang et al. [31] incorporated the local
effectiveness and efficiency of image segmentation due to two
spatial information of membership into the objective function
advantages. On the one hand, superpixel is able to achieve
of FCM, which obtains better results for image segmentation
a pre-segmentation based on the local spatial information of
than the algorithm proposed in [30]. Furthermore, Liu et al.
images. The pre-segmentation provides better local spatial in-
[24] improved FCM algorithm by integrating the distance
formation than traditional neighboring windows employed by
between different regions obtained by mean-shift and the
FCM S, FLICM, FGFCM, KWFLICM, NWFCM, NDFCM,
distance of pixels into its objective function. Although these
and FRFCM. On the other hand, superpixel is able to reduce
HMRF-based clustering algorithms [24], [30], [31] effectively
the number of different pixels in an image by replacing all
improve the effect of image segmentation, they have a high
pixels in a region with the mean value of the superpixel region
computational complexity caused by the computation of neigh-
[36], [37]. In this paper, we will employ superpixel technology
borhood information provided by original image and previous
to obtain adaptive local spatial information, and then compute
state’s membership in every iteration.
the histogram of superpixel image to achieve fast color image
It is clear that the algorithms mentioned above improve
segmentation.
image segmentation effect at the cost of increasing the compu-
tational complexity. Therefore, the question arises how one can
maintain local spatial information while reducing the compu- A. Motivation for Using Superpixel
tational complexity efficiently. Lei et al. [32] proposed a fast In early improved FCM algorithms, local spatial information
and robust FCM algorithm (FRFCM) to address the problem is often insufficient in a neighboring window of fixed size
by employing morphological reconstruction [33] and mem- and shape. If the window is too small, the local spatial
bership filtering. Because the repeated distance computation information will be limited for improving segmentation effect.
between pixels within neighborhood window and clustering But if the window is too large, the computational complexity
centers is removed, the algorithm is very fast and provides of the corresponding algorithm will be very high. Recently,

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 3

some improved FCM algorithms [26] incorporate adaptive a high computational complexity when the resolution of an
local spatial information into their objective function to obtain image is high. The enhanced FCM (EnFCM) proposed by
better robustness and higher performance for image segmenta- Szilágyi et al. [38] solves the problem by performing clus-
tion. Adaptive local spatial information means that the pixels tering on gray levels instead of pixels. The idea is efficient
within a neighboring region have variable weighting factors for the reduction of the computational complexity because
depending on local characteristics of an image. For example, the repeated distance computation is removed by integrating
in Liu’s algorithm [24], the adaptive neighborhood of a pixel histogram to its objective function. The objective function of
is decided by its neighboring window and the corresponding EnFCM is defined as
region obtained by a prior mean-shift algorithm [11]. Xq X c
In [24], the objective function denoted by Jm is defined as Jm = γl um l 2
kl kf − vk k , (5)
N X c N X c   l=1 k=1
X X uki
Jm = uki Dki + $ uki log (1) where ukl represents the fuzzy membership of gray value l
i=1 i=1
πki with respect to the kth clustering center vk , m is the weighting
k=1 k=1

where uki is the membership between the ith pixel and the exponent, f is a grayscale image, f l is the gray level, 1 ≤ l ≤
kth clustering center, 1 ≤ i ≤ N , 1 ≤ k ≤ c, N is the number q, q denotes the number of the gray levels of f (it is generally
of data items, c is the number of clusters, N, c ∈ N + , $ is far smaller than N ), γl is the number of pixels whose gray
the degree of fuzziness of uki , the distance function Dki is the level equals to f l , and
combination of the pixel dissimilarity and region dissimilarity, q
X
dki + dkRi γl = N. (6)
Dki = . (2) l=1
2
In (2), dki is the dissimilarity distance between the ith pixel Clearly, the introduction of histogram is able to reduce
and the kth clustering center, dkRi is the region dissimilarity the computational complexity of FCM. Because the level of
between the region Ri obtained P by mean-shift and the kth histogram is far less than the number of pixels in an image,
clustering center, dkRi = sum(R1 it is faster to implement FCM on gray levels than pixels for
) j∈Ri dkj , Ri is the region
i
grayscale image segmentation. However, it is difficult to ex-
that contains the ith pixel, and the sum(Ri ) denotes the
tend this idea of EnFCM to FCM for color image segmentation
number of pixels in the region Ri . Furthermore,
P [39] because the number of different colors is usually close to
j∈N wj ζj ukj the number of pixels in a color image. This is also the reason
πki = Pc P i . (3)
k=1 j∈Ni wj ζj ukj
that FRFCM [32] usually requires a longer execution time
to segment a color image than the corresponding grayscale
In (3), wj is a weighting parameter of the neighborhood pixels, image.
Ni is the neighborhood of the ith pixel, and j ∈ Ni , and ζ is
To address the issue, in this work, we will compute the
the region-level iterative strength,
histogram of a color image according to the corresponding
1
−1 superpixle image since the number of regions in the superpixel
ζj = ERi ,Rj + 1 , (4)
Z image is far smaller than the number of pixels in the original
where ERi ,Rj = kmean(Ri ) − mean(Rj )k is the Euclidean color image. We will use the mean value of all pixels within an
distance between the mean values of region Ri and Rj . Z = area instead of these pixels to reduce the number of different

−1
P colors in the original color image. It is easy to compute the
j∈Ni ERi ,Rj + 1 is a normalized constant.
Clearly, a prior over-segmentation obtained by mean-shift is histogram of the superpixel image because there is only a
necessary for Liu’s algorithm. However, mean-shift is sensitive small number of different colors in the superpixel image. And
to parameters. Moreover, the fuzzy membership depends on then, the fast FCM algorithm will be achieved for color image
both the pixel’s neighboring window and the region containing segmentation, which will be presented in detail in Section
the pixel. III.B.
Based on the analysis above, although Liu’s algorithm is
able to improve image segmentation effect by incorporating
III. M ETHODOLOGY
adaptive local spatial information into the objective function,
it has a high computational complexity due to the repeated Since a superpixel image is able to provide better lo-
computation of adaptive neighboring information in every cal spatial information than a neighboring window of fixed
iteration. Although, we also employ adaptive neighboring size and shape, superpixel technologies such as mean-shift
information obtained by a superpixel algorithm to improve [11], simple linear iterative clustering (SLIC) [40], and WT
the segmentation effect, significantly different from Liu’s [41], are usually considered as pre-segmentation algorithms
algorithm is the fact that the proposed superpixel algorithm for improving segmentation results generated by clustering
has a lower computational complexity. algorithms [42], [43]. Compared to SLIC, mean-shift and
WT produce irregular superpixel areas that are better than
B. Motivation for Using Histogram of Color Images hexagonal regions obtained by SLIC. In practical applications,
Traditional FCM algorithm has to compute the distance mean-shift is more popular than WT since the latter is sensitive
between each pixel and clustering centers, which leads to to noise leading to a serious over-segmentation.

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 4

Input image Gradient image MMGR WT Color feature and histogram of superpixel Segmentation result

Superpixel based on MMGR-WT Superpixel-based Fast FCM

Fig. 1. Framework of the proposed algorithm.

TABLE I. Comparison of the number of superpixel regions for WT based on MGR and MMGR respectively.
MGR MMGR
Parameters r=1 r=3 r = 10 r1 = 2, r2 = 7 r1 = 2, r2 = 11 r1 = 2, r2 = 20
Number 1210 263 10 264 95 95

Even though mean-shift is able to provide better superpixel

results, it is sensitive to parameter values, e.g., the spatial
bandwidth denoted by hs , the range bandwidth denoted by
hr , and the minimum size of final output regions denoted
by hk . Moreover, the computational complexity of mean-shift (a) (b) (c) (d)
is higher than WT. Therefore, we need to develop a fast Fig. 2. Watershed segmentation based on MGR with different SEs.
superpixel algorithm that can provide better pre-segmentation (a) Original image “12003” (image size: 481 × 321). (b) r = 1. (c)
result and requires less time than mean-shift. Because WT r = 3. (d) r = 10.
only depends on region minima of gradient images to obtain
per-segmentation, it has a very low computational complexity.
In this work, we employ a novel WT based on MMGR
(MMGR-WT) to produce superpixel images. The MMGR-WT
is able to provide more appropriate pre-segmentation results
(a) (b) (c) (d)
using shorter execution time than mean-shift. Moreover, it is
insensitive to parameters. Fig. 3. Segmentation results using MMGR-WT with different r1 ,
Based on the superpixel image obtained by MMGR-WT, where r2 = 10. (a) r1 = 1. (b) r1 = 3. (c) r1 = 5, (d) r1 = 8.
we compute the histogram of superpixel images to achieve
fast FCM algorithm. The computation of the histogram of
where Rε and Rδ represent morphological erosion and dilation
superpixel images is easy because the number of different
reconstruction, respectively, f is the original image, i.e., the
colors from superpixel images is far smaller than that from the
mask image, g is the marker image, ε is the erosion operation
original color image. Finally, the histogram is considered as a
and δ is the dilation operation. Erosion reconstruction requires
parameter of the objective function to achieve fast color image
that g ≥ f , but dilation reconstruction requires g ≤ f ,
segmentation. The framework of our proposed algorithm is (1) (i) (1)
εf (g) = ε(g) ∨ f , εg (f ) = ε(ε(i−1) (g)) ∨ f , δf (g) =
shown in Fig. 1. (i)
δ(g)∧f , and δg (f ) = δ(δ (i−1) (g))∧f . The symbols ∨ and ∧
stand for the pointwise maximum and minimum, respectively.
Because morphological erosion and dilation are a pair of
A. Superpixel-based on MMGR-WT
dual operators, they always appear in pairs such as mor-
Watershed transform is a fast algorithm used for image seg- phological opening and closing operators. The morphological
mentation via computing local minima of a gradient image and opening and closing are more popular than erosion and dilation
searching the watershed line between adjacent local minima. because they have stronger capability for feature extraction
The algorithm easily causes an over-segmentation because it or noise removal. Consequently, the morphological opening
is sensitive to noise. To address the problem, many algorithms reconstruction denoted by RO and closing reconstructions
have been proposed by modifying the gradient image of denoted by RC , are defined as
the original image. Among these algorithms, morphological  O
R (g) = Rδ (Rε )
gradient reconstruction (MGR) [44] is a simple and efficient , (8)
RC (g) = Rε (Rδ )
algorithm for overcoming over-segmentation because it is able
to preserve the contour details of objects while removing noise where the marker image g is generally considered as g =
and useless gradient details. Firstly, the basic definition of εB (f ) in Rδ or g = δB (f ) in Rε . B is a structuring element
morphological reconstruction is presented as follows (SE).
( Both RO and RC are able to remove region minima in a
(i)
Rfε (g) = εf (g) gradient image to reduce over-segmentation. For instance, we
(i) , (7)
Rfδ (g) = δf (g) use RC to reduce over-segmentation as shown in Fig. 2.

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 5

(a)
(a) (b) (c) (d)
Fig. 4. Watershed segmentation based on MMGR-WT with different
sized SEs. (a) r1 = 2, r2 = 3. (b) r1 = 2, r2 = 7. (c) r1 = 2, (b)
r2 = 11, (d) r1 = 2, r2 = 20.

TABLE II. The values of r2 for ten images from BSDS for different
(c)
values of η.
Images η = 10−2 η = 10−3 η = 10−4 η = 10−5
“2092” 12 17 26 26
“3096” 10 10 10 10 (d)
“8023” 10 10 14 14
“8049” 14 19 22 22
“8143” 7 10 10 10
“12003” 12 18 18 18 (e)
“12074” 10 18 24 24
“12084” 14 15 15 15 Fig. 5. Superpixel images using different methods. (a) Original im-
“14037” 10 14 17 17 ages. (b) Superpixel images obtained by SLIC (sk = 500, sm = 50,
ss = 1). (c) Superpixel images obtained by mean-shift1 with hs = 7,
“15004” 14 18 18 18
hr = 7, hk = 30. (d) Superpixel images obtained by mean-shift2
with hs = 15, hr = 15, hk = 50. (e) Superpixel images obtained
by MMGR-WT (r1 = 2).
In Fig. 2, the SE is defined as a disk, where r is the radius of
the SE. Fig. 2 shows that the number of segmentation regions
decreases quickly by increasing the value of r. However, a the superpixel result has a clearly low contour precision when
small SE easily leads to over-segmentation while a large SE r1 = 8. Consequently, we choose 1 ≤ r1 ≤ 3 here. Because r2
easily leads to under-segmentation. Therefore, it is difficult to controls the size of the maximal region, the superpixel image
obtain a superpixel image with both fewer regions and accurate is better when the value of r2 is larger as shown in Fig. 4.
contour by using MGR. To balance the number of regions However, the superpixel image is unchanged when the value
in superpixel image and contour precision, a suitable SE is of r2 is larger than a threshold; for example, the threshold is
required, but it is difficult to choose a suitable SE for different 11 in Fig. 4. Clearly, the superpixel image is convergent via
images. increasing the value of r2 . Moreover, the convergent result
To solve the problem, we try to use different SEs to is perfect because it includes fewer regions and yet provides
reconstruct a gradient image, and then fuse these reconstructed accurate contour. Therefore, the MMGR is insensitive to the
gradient images to remove the dependency of segmentation change of r2 when r2 is larger than a threshold. Table I shows
result on SEs. Thus, we propose a MMGR operation denoted the comparison of the number of superpixel regions for WT-
by RM C that is defined as follows MGR and WT-MMGR, respectively.
MC C C C
Rf (g, r1 , r2 ) = ∨{Rf (g)Br1 , Rf (g)Br1 +1 , · · · , Rf (g)Br2 }. As can be seen from Table I that r2 can be variant. But
(9) it is difficult to set different values of r2 for each image. In
where r1 and r2 represent minimal and maximal r, respec- practical applications, r2 is adaptive and it is not required for
+
tively, r1 ≤ r ≤ r2 , r1 , r2 ∈ N , g ≤ f . MMGR as long as we set a minimal error threshold denoted
We can see that R MC
employs multiscale SEs to reconstruct by η instead of r2 , i.e.,
a gradient image to obtain multiple reconstructed images. max Rf (g, r1 , r2 ) − RfM C (g, r1 , r2 + 1) ≤ η. (10)
 MC
By computing the pointwise maximum of these reconstructed
gradient images, an excellent gradient image that removes In (10), r2 can be replaced by η because r2 is supposed to
most of useless local minima while preserving important edge have different values for each image in a dataset, but a fixed
details is obtained. value of η can be used for all images in the dataset. Note that
The proposed MMGR includes two parameters, r1 and r2 , if η is too large, r2 will be small but the error will be large.
where r1 controls the size of the minimal region and r2 On the contrary, if η is too small, the error will be small but
controls the size of the maximal region. If r1 is too small, r2 will be large leading to a high computational burden for
there will be many small regions in segmentation results, but MMGR. Therefore, it is important to choose an appropriate
if r1 is too large, the boundary precision will be low. An η for a dataset. We perform MMGR on ten images from the
example is shown in Fig. 3. It can be seen that the superpixel Berkeley segmentation dataset and benchmark (BSDS), we can
result has a high contour precision but includes some small obtain different values of r2 according to a fixed value of η
regions when r1 = 1, the superpixel result has a high contour as shown in Table II.
precision and excludes small regions when r1 = 2 or r1 = 3, Table II shows that the values of r2 will be larger when

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 6

TABLE III. Comparison of execution time (in seconds) of different

methods used to generate superpixel images. The best values are in
bold.
Algorithms “100075” “124084” “100007” “145086” Average
SLIC 3.86 4.07 3.89 3.88 3.93
mean-shift1 1.02 1.22 0.94 1.20 1.10 (a) (b) (c)
mean-shift2 2.66 1.22 2.67 2.86 2.79
MMGR-WT 0.32 0.32 0.31 0.36 0.33 Fig. 7. The color distribution of different color images. (a) The color
distribution of Fig. 6(a). (b) The color distribution of Fig. 6(b). (c)
The color distribution of Fig. 6(d).

(a) (b) (c) (d) (e)

Fig. 6. The quantization of a color image and the corresponding
histogram. (a) The original image. (b) Color quantization using the
algorithm proposed in [39] (c = 10). (c) The histogram of Fig.
1(b). (d) The superpixel image using MMGR-WT (r1 = 2). (e) The
histogram of Fig. 1(c). (a) (b)
Fig. 8. Segmentation reuslts on Fig. 6(a). (a) The segmentation result
decreasing η. However, r2 will be unchanged when η is using FCM for quantized image. (b) The segmentation result using
the proposed SFFCM.
smaller or equal to 10−4 . Therefore, we set η = 10−4 in
this paper.
To demonstrate the effectiveness of the MMGR, Fig. 5
shows superpixel images obtained by SLIC, mean-shift, and levels which is usually far smaller than the number of pixels
MMGR-WT, respectively, where sk is the number of desired in an image. But the number of different colors in a color
superpixels, sm is the weighting factor between color and image is far larger than 256. The quantization technology is
spatial differences, and ss is the threshold used for region usually used to reduce the number of colors in an image.
merging. These parameters are selected depending on [40], The basic idea of quantization technology is that a clustering
[42]. It can be seen from Fig. 5 that the superpixel images algorithm is performed on each channel of a color image to
generated by SLIC include lots of areas with similar shape and obtain an image with fewer color levels than before. However,
size, but the superpixel images generated by the mean-shift and the traditional color quantization only reduces the number of
MMGR-WT include lots of areas with different shapes sizes. different colors, but the color distribution of the quantized
It is clear that later two algorithms provide better visual effect image is still similar to that of the original image because
for the requirement of real images. the local spatial information is ignored. Because a superpixel
Although SLIC and mean-shift are able to generate super- image carries the spatial information of the image and reduces
pixel images according to task requirements by changing pa- the number of different colors, the superpixel image is superior
rameters, they have a longer execution time than the proposed to images quantized by clustering algorithms. We applied
MMGR as shown in Table III, where SLIC corresponds to the clustering algorithm proposed in [39] and the proposed
Fig. 5(b), mean-shift1 corresponds to Fig. 5(c), mean-shift2 MMGR-WT to quantize a color image, and then computed
corresponds to Fig. 5(d), and MMGR-WT corresponds to Fig. the histogram of the quantized image as shown in Fig. 6,
5(e). Because our purpose is to propose a fast FCM algorithm where the number of different colors is 57214 in the original
for color image segmentation, MMGR is more appropriate image. Furthermore, Fig. 7, the color distribution of Fig. 6,
than SLIC and mean-shift for our task requirement. shows that the proposed MMGR-WT is more appropriate than
clustering algorithm proposed in [39] for subsequent image
B. Superpixel-based Fast FCM segmentation.
In Section III.A, we proposed the MMGR-WT to obtain It is clear that the histogram of Fig. 6(b) and Fig. 6(d) are
better local spatial information used for fuzzy clustering. simpler with only a small number of different colors appearing
Because MMGR-WT depends on the local feature of an image in the quantized images. According to Fig. 6(c) and Fig. 6(e),
while FCM depends on the global feature, the combination of we can extend EnFCM to color image segmentation easily.
MMGR-WT and FCM is able to improve image segmentation Compared to Fig. 6(c), Fig. 6(e) has even fewer color levels.
result. In this section, we propose a superpixel-based fast FCM In addition, it is clear that the color distributions of Fig. 7(c)
algorithm by incorporating adaptive local spatial information is different from Fig. 7(a) and Fig. 7(b), and the former is
into the objective function of FCM. helpful for subsequent pixel classification.
EnFCM is popular and efficient for achieving fast image Based on the superpixel image obtained by MMGR-WT, we
segmentation because a gray image only includes 256 gray proposed the objective function of SFFCM for color image

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 7

0 0
segmentation as follows Step 1: Set values for c, m, r1 , η, η , where η is the
q X
c convergence condition used for SFFCM.
X 1 X
Jm = Sl um
kl k( xp ) − vk k2 , (11) Step 2: Compute a superpixel image using (9-10), and then
Sl compute its histogram.
l=1 k=1 p∈Rl
(1) Compute the gradient image using Sobel operators.
where l is the color level, 1 ≤ l ≤ q, q is the number of regions
(2) Implement MMGR using (9-10) and η.
of the superpixel image, l, q ∈ N + , Sl is the number of pixels
(3) Implement WT to obtain the superpixel image.
in the lth region Rl , and xp is the color pixel within the lth
Step 3: Initialize randomly the membership partition matrix
region of the superpixel image obtained by MMGR-WT. The
U(O) according to the superpixel image.
new objective function only introduces histogram information
Step 4: Set the loop counter b = 0.
compared with the old one in FCM. Because each color pixel
Step 5: Update the clustering centers using (15).
in the original image is replaced by the mean value of color
Step 6: Update the membership partition matrix U(t) using
pixels within the corresponding region of the superpixel image,
(16).
the number of color level is equivalent to the number of regions 0
Step 7: If max{U(b) − U(b+1) } < η then stop, otherwise,
in the superpixel image. Thus, the computational complexity set b = b + 1 and go to Step 5.
is efficiently reduced due to l  N . We applied the proposed SFFCM to Fig. 6(a) following the
Utilizing the Lagrange multiplier technique, the aforemen- previous steps. Then, the segmentation result is shown in Fig.
tioned optimization problem can be converted to an uncon- 8. We can see that the proposed SFFCM is able to obtain better
strained optimization problem that minimizes the following segmentation result than the traditional algorithm. Based on
objective function, i.e., the analysis mentioned above, we conclude that the proposed
q X
c c SFFCM has following advantages:
X 1 X X
Jem = Sl um
kl k( xp ) − vk k2 − λ( ukl − 1), • SFFCM is very fast for color image segmentation because
Sl
l=1 k=1 p∈Rl k=1
the number of different colors is reduced efficiently due
(12)
to superpixel and color histogram.
where λ is a Lagrange multiplier. We compute the partial
• SFFCM is insensitive to the change of parameters be-
differential equation of Jem with respect to ukl and vk , re-
cause the superpixel image obtained by MMGR-WT is
spectively,
convergent.
q Xc 1
∂Sl um 2
P
∂ Jem kl k( Sl p∈Rl xp ) − vk k • SFFCM obtains an excellent result for color image seg-
X
= −λ mentation because both adaptive local spatial information
∂ukl ∂ukl
l=1 k=1
q X
c
and global color feature are incorporated into the objec-
X 1 X , (13) tive function.
= mSl um−1
kl k( xp ) − vk k2 − λ
Sl
l=1 k=1 p∈Rl
IV. E XPERIMENTS
=0
We conduct experiments on two synthetic color images
q X c 1
∂Sl um 2
P
∂ Jem kl k( Sl p∈Rl xp ) − vk k of size 256 × 256 and real color images from the Berkeley
X
=
∂vk ∂vk Segmentation Dataset and Benchmark (BSDS) [45] and the
l=1 k=1
q X c Microsoft Research Cambridge (MSRC) [46]. The first syn-
∂k( S1l p∈Rl xp ) − vk k2
P
thetic image includes four different colors while the second
X
= Sl um
kl
∂vk includes five different colors. The experiments are conducted
l=1 k=1
q 1
∂k( Sl p∈Rl xp ) − vk k2
P on a DELL desktop with Intel(R) Core (TM) CPU, i7-6700,
X . (14)
= Sl um
kl 3.4GHz, 16GB RAM.
∂vk
l=1
q
X 1 X A. Comparative Algorithms
= −2 Sl um
kl k( xp ) − vk k
Sl To assess the effectiveness and efficiency of the proposed
l=1 p∈Rl
SFFCM, nine comparative algorithms based on clustering used
=0
for color image segmentation are presented, i.e., FCM [10],
Combing (13)-(14) together, the corresponding solutions for FGFCM [18], HMRF-FCM [30], FLICM [19], NWFCM [20],
ukl and vk are obtained: KWFLICM [23], NDFCM [21], Liu’s algorithm [24], and
Pq m
P FRFCM [32]. Since these algorithms employ different local
l=1 ukl p∈Rl xp
vk = Pq m , (15) spatial neighborhoods to improve segmentation results, they
l=1 Sl ukl have different advantages and disadvantages.
−2/(m−1)
1 P
( Sl p∈Rl xp ) − vk

B. Parameters Setting
ukl = −2/(m−1) . (16)
Pc 1 P Since both comparative algorithms and the proposed SF-
x ) − v

j=1 ( Sl p∈Rl p j
FCM belong to clustering algorithms based on objective func-
Based on (9)-(16), the proposed SFFCM algorithm can be tion optimization, three indispensable parameters: the weight-
summarized as follows: ing exponent, the convergence condition, and the maximal

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 8

(a) (b) (c) (d) (a) (b) (c) (d)

sy5
Gaussian sy_4
Original noise_4
(h) (e) (f) (h)
(e) (f) (g) 1 4~10 12 13 (g)

(i) (j) (k) (l) (i) (j) (k) (l)

Fig. 9. Comparison of segmentation results on the first symmetric Fig. 10. Comparison of segmentation results on the second symmetric
image. (a) The first synthetic image. (b) Noisy image (Gaussian image. (a) The second synthetic image. (b) Noisy image (Salt &
noise, the noise level is 10%). (c) FCM. (d) FGFCM. (e) HMRF- Pepper, the noise level is 40%). (c) FCM. (d) FGFCM. (e) HMRF-
FCM. (f) FLICM. (g) NWFCM. (h) KWFLICM. (i) NDFCM. (j) FCM. (f) FLICM. (g) NWFCM. (h) KWFLICM. (i) NDFCM. (j)
Liu’s algorithm. (k) FRFCM. (l) SFFCM. Liu’s algorithm. (k) FRFCM. (l) SFFCM.

number of iteration must be set before iterations. In our exper- FCM, HMRF-FCM, FLICM, and NWFCM provide poor
iments, the three parameters are 2, 10−5 , and 50, respectively. results as shown in Figs. 9-10 (c, e, f, and g), which show
In addition, the value of the minimal error threshold used for that they are sensitive to both Gaussian and Salt & Pepper
MMGR is 10−4 . In the comparative algorithms, a window of noise. HMRF-FCM, FLICM, and NWFCM cannot improve
size 3 × 3 is employed by those algorithms required a neigh- the FCM algorithm for color images. FGFCM, NDFCM and
boring window of fixed size for fair comparison. Moreover, the FRFCM obtain good segmentation results as shown in Fig. 9
computational complexity is also an important reason for the (d, i, and k) for the image corrupted by Gaussian noise, but
choice of the window of size 3×3. In addition, a neighborhood poor segmentation results as shown in Fig. 10 (d, i, and k) for
window is unnecessary for FCM. According to the criterion of the image corrupted by Salt & Pepper noise. It is clear that the
parameters setting mentioned in those comparative algorithms, three algorithms are insensitive to Gaussian noise but they are
the spatial scale factor and the gray-level scale factor in sensitive to Salt & Pepper noise of high density. KWFLICM,
FGFCM and NDFCM, are λs = 3 and λg = 5, respectively. Liu’s algorithm, and the proposed SFFCM provide better
The third parameter of the NDFCM, a new scale factor is results as shown in Figs. 9-10 (h, j, and l), which demonstrates
λa = 3. The NWFCM only refers to the gray-level scale factor, that they are robust against both Gaussian noise and Salt &
λg = 5. Because Liu’s algorithm requires a pre-segmentation Pepper noise as adaptive neighboring information is employed
obtained by mean-shift, three parameters hs = 10, hr = 10, by the three algorithms.
and hk = 100 follow the original paper. Except three indis- To assess the performance of different algorithms on noisy
pensable parameters mentioned above and the number of the image segmentation, two performance indices, the quantitative
cluster prototypes, HMRF-FCM, FLICM and KWFLICM do score (S) that is the degree of equality between pixel sets
not require any other parameters. In FRFCM, the structuring Ak and the ground truth Ck and the optimal segmentation
element used for multivariate morphological reconstruction is accuracy (SA) that is the sum of the correctly classified pixels
a square of size 3 × 3, and the filtering window used for divided by the sum of the total number of the pixels [23], are
membership filtering is also a square of size 3 × 3. As the adopted. S and SA are defined as
proposed SFFCM needs a minimal structuring element for c T
X Ak Ck
MMGR, we set r1 = 2 for MMGR. S= S , (17)
Ak Ck
k=1
c T
A Ck
Pkc
X
C. Results on Synthetic Images SA = . (18)
k=1 j=1 Cj
First, we test these comparative algorithms and the proposed
SFFCM on two synthetic color images to show their robustness where Ak is the set of pixels belonging to the kth class found
to noise. In this experiment, three kinds of different noise by the algorithm while Ck is the set of pixels belonging to
Gaussian, Salt & Pepper, and Uniform noise are added to the class in the ground truth. We implemented each of these
these synthetic images. All algorithms mentioned above are algorithms on two synthetic images, and computed the mean
implemented and segmentation results are shown in Figs. 9- value and the root mean square error (RMSE) of S and SA as
10. shown in Tables IV-V.

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 9

TABLE IV. Comparison scores (S%) of the ten algorithms on the first synthetic image corrupted by noise of different levels (c = 4). The
best values are in bold.
Noise FCM FGFCM HMRF-FCM FLICM NWFCM KWFLICM NDFCM Liu’s algorithm FRFCM SFFCM
Gaussian 5% 95.53 99.85 95.60 95.60 95.61 99.78 99.87 99.41 99.82 99.20
Gaussian 10% 85.69 99.75 85.70 85.69 85.70 99.68 99.61 98.46 99.81 99.17
Gaussian 15% 66.35 98.20 66.24 66.26 66.25 99.24 99.50 97.54 99.62 99.15
Gaussian 20% 55.22 94.05 55.29 55.31 55.29 98.45 99.53 96.76 99.48 98.98
Salt & Pepper 10% 80.21 95.56 80.21 80.21 80.21 99.65 81.91 99.67 99.68 99.18
Salt & Pepper 20% 65.54 87.55 65.54 65.54 65.54 98.66 71.28 99.23 98.58 98.91
Salt & Pepper 30% 51.51 77.21 54.07 54.07 54.07 95.21 69.39 98.55 97.11 98.68
Salt & Pepper 40% 42.19 65.94 44.51 42.19 44.51 87.69 66.60 96.76 78.64 92.83
Uniform 10% 84.41 99.49 84.44 84.41 84.45 99.81 82.31 99.77 99.81 99.22
Uniform 20% 71.08 97.95 71.02 71.08 71.02 99.57 77.59 99.28 99.68 99.19
Uniform 30% 59.83 94.31 60.00 60.08 60.00 98.91 89.23 98.72 99.39 98.99
Uniform 40% 50.05 87.46 49.94 50.05 49.94 96.50 99.46 97.23 98.44 98.83

Mean value 67.30 91.44 67.71 67.54 67.72 97.76 86.36 98.45 97.51 98.53
RMSE 16.49 10.48 15.98 16.28 15.98 3.48 13.15 1.11 6.00 1.80

TABLE V. Segmentation accuracy (SA%) of ten algorithms on the second synthetic image corrupted by noise of different levels (c = 5).
The best values are in bold.
Noise FCM FGFCM HMRF-FCM FLICM NWFCM KWFLICM NDFCM Liu’s algorithm FRFCM SFFCM
Gaussian 3% 93.10 99.43 93.09 93.10 93.10 99.47 99.87 99.96 99.64 99.59
Gaussian 5% 87.03 98.44 86.85 87.01 86.86 99.04 99.70 99.96 99.55 99.52
Gaussian 10% 74.71 95.74 74.24 74.64 74.33 97.52 99.52 98.58 97.09 99.38
Gaussian 15% 66.02 92.30 65.70 66.06 65.70 95.69 98.84 97.09 94.87 99.41
Salt & Pepper 10% 86.42 97.33 86.46 86.42 73.86 99.44 91.26 99.84 99.57 99.55
Salt & Pepper 20% 74.40 92.41 74.54 74.40 74.54 98.55 89.52 99.68 97.08 99.48
Salt & Pepper 30% 63.52 84.59 58.15 63.52 58.15 95.59 86.49 99.03 92.60 99.00
Salt & Pepper 40% 49.86 75.76 48.89 49.86 48.89 76.97 62.54 97.56 85.48 98.99
Uniform 10% 90.41 99.26 90.38 90.42 90.38 99.47 94.13 99.91 99.62 99.59
Uniform 20% 81.09 97.77 80.97 81.09 80.97 99.16 90.17 99.81 98.95 99.53
Uniform 30% 72.34 94.96 62.67 72.34 72.02 98.13 99.59 99.64 96.71 99.41
Uniform 40% 63.42 89.73 62.80 63.42 54.73 82.16 99.82 82.75 93.71 99.20

Mean value 75.19 93.14 73.73 75.19 72.79 95.10 96.62 97.81 96.24 99.39
RMSE 13.01 7.02 14.16 13.01 14.00 7.46 10.67 4.84 4.17 0.21

In Tables IV-V, FCM, HMRF-FCM, FLICM, and NWFCM filtering to improve segmentation results.
obtain similar S values as well as SA values, which further As KWFLICM, Liu’s algorithm, and the proposed SFFCM
demonstrates that HMRF-FCM, FLICM and NWFCM are employ adaptive local spatial information to improve seg-
inefficient for color image segmentation. FCM misses the mentation results, they obtain larger S and SA than those
local spatial information leading to poor segmentation result. comparative algorithms that employ local spatial information
HMRF-FCM, FLICM, and NWFCM only employ a small in a window of fixed size. Liu’s algorithm obtains higher
neighboring window to incorporate local spatial information values of S and SA because of the combination of mean-
into their objective function, which is helpful for segmenting shift, FCM and HMRF. In some cases, the proposed SFFCM
images corrupted by low-density noise but not useful for provides smaller S and SA than Liu’s algorithm but higher
segmenting images corrupted by high-density noise. FGFCM values than those comparative algorithms because contour
and NDFCM obtain higher values of S and SA than FCM, details are smoothed in segmentation results obtained by the
HMRF-FCM, FLICM and NWFCM because the tested images SFFCM. However, SFFCM provides the best mean value of S
are synthetic and the added noise is known. Because FGFCM for two synthetic images and the best root of mean square error
and NDFCM employ a filter to suppress noise before iterations (RMSE) of SA for the second synthetic image, which shows
in clustering, they obtain larger S and SA than FCM, HMRF- that SFFCM is able to obtain good segmentation results for
FCM, FLICM and NWFCM for synthetic images corrupted images corrupted by different noises.
by known noise. FRFCM obtains high S and SA when noisy
density is low, but small S and SA when noisy density D. Results on Real Images
is high because FRFCM employ multivariate morphological To demonstrate that the proposed SFFCM is useful for real
reconstruction to simplify image and use the membership image segmentation, we further conducted experiments on the

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 10

Images FCM FGFCM HMRF-FCM FLICM NWFCM KWFLICM NDFCM Liu’s algorithm FRFCM SFFCM

“12003” GT-1 GT-2 GT-3 GT-4 GT-5

c=2

c=3

“113009” GT-1 GT-2 GT-3 GT-4 GT-5

Fig. 11. Ground truth (GT) segmentations of images “12003” and c=3

“113009” from BSDS.

c=3
c=2 c=3 c=3 c=4 c=4

c=4
Images

c=4
FCM

Fig. 13. Comparison of segmentation results on color images from

BSDS using different models.
FGFCM

dataset contains 23 object classes and comprises of 591 natural

HMRF-FCM
images. For each image in MSRC, there is only one ground
truth segmentation that is pixel-wise labelled.
Figs. 12-14 show segmentation results of images from the
FLICM
BSDS and MSRC. The parameters setting is the same as that
in Section IV.B. Since the size of images in BSDS and MSRC
NWFCM
is different from the size of synthetic images, the structuring
element used for multivariate morphological reconstruction in
FRFCM is a disk of size 5 × 5 for the BSDS and MSRC.
KWFLICM The value of r1 is set to 3 for real images in this Section. In
addition, the CIE-Lab color space is used for all algorithms
for fair comparison.
NDFCM As can be seen from Figs. 12-13 that segmentation results
obtained by FCM, FGFCM, HMRF-FCM, FLICM, NWFCM,
KWFLICM, and NDFCM include a great number of small
Liu’s algorithm regions because only a small local neighboring window is
employed (a large neighboring window will cause a very
high computational complexity). FRFCM obtains better results
FRFCM
than algorithms mentioned above due to the introduction
of multivariate morphological reconstruction and membership
filtering. However, Liu’s algorithm and the proposed SFFCM
SFFCM
obtain better results than FRFCM due to the use of adaptive
Fig. 12. Comparison of segmentation results on color images from local spatial information provided by pre-segmentation. Al-
BSDS using different models.
Groud Truth
though Liu’s algorithm provides a better segmentation result
than the proposed SFFCM on the left image, the later provides
better results than the former on four other images in Fig.
BSDS and MSRC. The BSDS is a popular benchmark that 12. In practical applications, since it is difficult to propose
has been widely used by researchers for the task of image an algorithm to achieve the best segmentation result for every
segmentation [45]. The early BSDS is named as BSDS300 image in a dataset, researchers usually use the average result
that is composed of 300 images and the current BSDS is on all images in the dataset, e.g., BSDS and MSRC, to estimate
an extended version that is composed of 500 images, called the algorithm performance.
BSDS500. For each image in BSDS, there are 4-9 ground In Fig. 14, all algorithms are efficient for images in which
truth segmentations. These ground truth segmentations are the foreground is clearly different from the background as
delineated by different human subjects. For instance, there shown in the first row. A large number of small regions appear
are 5 ground truth segmentations delineated by 5 subjects on in segmented images except images obtained by SFFCM as
image “12003” and “113009” as shown in Fig. 11. The MSRC shown in the second to fifth rows. Liu’s algorithm, FRFCM,

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 11

Images FCM FGFCM HMRF-FCM FLICM NWFCM KWFLICM NDFCM Liu’s algorithm FRFCM SFFCM

Fig. 14. Comparison of segmentation results on color images from MSRC using different models (c = 2).

TABLE VI. Average performance of ten algorithms on the BSDS300 TABLE VII. Average performance of ten algorithms on the BSDS500
that includes 300 images. The best values are in bold. that includes 500 images. The best values are in bold.
Algorithms PRI↑ CV↑ VI↓ GCE↓ BDE↓ Algorithms PRI↑ CV↑ VI↓ GCE↓ BDE↓
FCM 0.74 0.43 2.87 0.41 13.78 FCM 0.74 0.43 2.88 0.40 13.48
FGFCM 0.74 0.43 2.80 0.40 13.63 FGFCM 0.75 0.44 2.81 0.39 13.28
FLICM 0.74 0.43 2.82 0.40 13.69 FLICM 0.74 0.43 2.83 0.40 13.38
NWFCM 0.74 0.43 2.78 0.41 13.78 NWFCM 0.74 0.43 2.88 0.40 13.47
KWFLICM 0.74 0.43 2.82 0.40 13.70 KWFLICM 0.74 0.44 2.83 0.40 13.40
NDFCM 0.74 0.44 2.87 0.39 13.52 NDFCM 0.75 0.44 2.78 0.39 13.13
HMRF-FCM 0.74 0.43 2.77 0.40 13.71 HMRF-FCM 0.75 0.43 2.78 0.40 13.22
Liu’s algorithm 0.77 0.48 2.53 0.35 12.57 Liu’s algorithm 0.76 0.47 2.58 0.36 12.31
FRFCM 0.75 0.46 2.62 0.36 12.87 FRFCM 0.76 0.45 2.67 0.37 12.35
SFFCM 0.78 0.55 2.02 0.26 12.90 SFFCM 0.78 0.54 2.06 0.26 12.80

and SFFCM obtain better results than other algorithms as distance between two segmentations in terms of their average
shown in the sixth row. All algorithms fail to segment images conditional entropy. The GCE computes the degree to which
except SFFCM, as shown in the last three rows. two segmentations are mutually consistent. The BDE is an
To evaluate segmentation results obtained by different al- error measure that is used to measure the average displacement
gorithms, five performance measures [45], [47], namely, the error of boundary pixels between two segmentations. If the
probabilistic rand index (PRI), the covering (CV), the variation segmentation result is more similar to the ground truth, PRI
of information (VI), the global consistency error (GCE), and and CV will be larger but VI, GCE and BDE will be smaller.
the boundary displacement error (BDE), are computed in this In the BSDS, each image corresponds to multiple ground truth
experiment. The PRI is a similarity measure that counts the segmentations, which leads to a result that a segmentation
fraction of pairs of pixels whose labels are consistent between result corresponds to multiple groups of performance index.
the computed segmentation and the corresponding ground Therefore, the average value of multiple groups of perfor-
truth segmentation. The CV is an overlap measure that can mance index is usually considered as the final performance
be also used to evaluate the segmentation effect. The VI index of the segmentation result.
is a similarity measure that is always used to measure the All these algorithms are evaluated on the BSDS and MSRC

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 12

TABLE VIII. Average performance of ten algorithms on the MSRC TABLE IX. Average performance of SFFCM on BSDS300.
that includes 591 images. The best values are in bold. Exponent PRI↑ CV↑ VI↓ GCE↓ BDE↓
Algorithms PRI↑ CV↑ VI↓ GCE↓ BDE↓ m=2 0.78 0.55 2.02 0.26 12.90
FCM 0.70 0.55 1.93 0.32 12.67 m=5 0.78 0.55 2.00 0.26 12.91
FGFCM 0.70 0.56 1.85 0.31 12.39 m = 10 0.78 0.55 2.01 0.25 12.87
FLICM 0.72 0.59 1.73 0.28 12.29 m = 30 0.78 0.55 2.02 0.26 13.03
NWFCM 0.69 0.55 1.90 0.32 12.61 m = 100 0.78 0.55 2.01 0.25 12.89
KWFLICM 0.69 0.55 1.93 0.32 12.67
NDFCM 0.69 0.55 1.90 0.32 12.54 TABLE X. Average performance of SFFCM on BSDS500.
HMRF-FCM 0.70 0.56 1.84 0.31 12.38 Exponent PRI↑ CV↑ VI↓ GCE↓ BDE↓
Liu’s algorithm 0.71 0.54 1.77 0.34 12.43 m=2 0.78 0.54 2.06 0.26 12.80
FRFCM 0.71 0.58 1.79 0.30 12.23 m=5 0.78 0.54 2.04 0.26 12.72
SFFCM 0.73 0.62 1.58 0.25 12.49 m = 10 0.78 0.54 2.06 0.26 12.72
m = 30 0.78 0.54 2.06 0.26 13.80
m = 100 0.78 0.54 2.05 0.26 12.82
datasets. The value of c is set from 2 to 6 for each image in
the BSDS, while its value is set from 2 to 4 for each image
in the MSRC. We choose the best value of c corresponding images used in Section IV.C-IV.D. We computed the average
to the highest PRI. Because the BSDS and MSRC includes execution time of algorithms on images from the BSDS and
lots of images, the average values of PRI, CV, VI, GCE and MSRC, respectively.
BDE corresponding to segmentation results of all images in the It can be seen from Table XI that FCM is faster than other
BSDS or MSRC are presented in Tables VI-VIII. In Tables VI- algorithms, except SFFCM, because no additional computa-
VII, we can see that FCM, FGFCM, FLICM, NWFCM, and tion is implemented. FGFCM and NDFCM are faster than
KWFLICM have similar values of the PRI, CV, VI, GCE and FLICM, NWFCM and KWFLICM because the neighboring
BDE. NDFCM has the similar performance with HMRF-FCM. information is computed in advance. FLICM, NWFCM and
FRFCM clearly outperforms other algorithms on PRI and KWFLICM repeatedly compute the neighboring information
BDE due to the introduction of multivariate morphological re- in each iteration leading to a high computational complexity.
construction. Liu’s algorithm obtains better performance than Both HMRF-FCM and Liu’s algorithm require a long execu-
FRFCM because it computes the distance between pixels and tion time because a prior probability used for HMRF model
clustering centers according to the combination of superpixel must be computed in each iteration. FRFCM is fast because
image and the original image. Similarly, Table VIII shows that multivariate morphological reconstruction and membership
FLICM, NWFCM, and KWFLICM have similar values of the filtering are implemented only once. The proposed SFFCM
PRI, CV, VI, GCE and BDE. The performance of FGFCM is is very fast even faster than FCM for some images because
similar to NDFCM. Different from Tables VI-VII, HMRF- the number of different colors in superpixel image obtained by
FCM obtains better performance than Liu’s algorithm and MMGR-WT is decreased efficiently and the color histogram
FRFCM as shown in Table VIII. Clearly, the proposed SFFCM is integrated into SFFCM.
is the most excellent because it obtains the best values of PRI,
V. C ONCLUSION
CV, VI, and GCE, as well as within the 0.04 of the best value
obtained of BDE as shown in Tables VI-VIII and the best In this paper, a superpixel-based fast FCM clustering al-
segmentation results as shown in Figs. 12-14. gorithm (SFFCM) for color image segmentation has been
To demonstrate that the proposed SFFCM is insensitive to proposed. Two main contributions are presented. The first
parameters, we further discussed the relationship between the contribution is that we presented the MMGR operation to
weighting exponent m and the SFFCM. We have known that obtain a good superpixel image. The second contribution is
FCM algorithm is insensitive to m when the FCM is used that we incorporated color histogram into objective function
for image segmentation. The proposed SFFCM has the same to achieve fast image segmentation. The proposed SFFCM is
objective function with FCM. The difference between them tested on synthetic and real images. The experimental results
is that the proposed SFFCM employ color histogram created demonstrate that the proposed SFFCM is superior to state-
by MMGR-WT to speed up the FCM algorithm. Therefore, of-the-art clustering algorithms because it provides the best
theoretically, the performance of the proposed SFFCM is segmentation results and requires the shortest running time.
also insensitive to the value of m. Tables IX-X show the Our algorithm is very fast for color image segmentation,
performance of SFFCM for different values of m. Fig. 15 but similar to other k-means clustering algorithms, it has lim-
shows the plot of Tables IX-X. It is clear that the performance itations in practical applications since the number of clusters
of SFFCM is changed slightly via changing the value of m. must be set prior. In the future work, we will explore fast
clustering algorithms that automatically estimate the number
E. Execution Time of clusters [48], [49].
Execution time is an important index used to measure R EFERENCES
the performance of an algorithm. Table XI shows execution [1] B. Abu-Jamous, R. Fa, A. K. Nandi, Integrative Cluster Analysis in
time of different algorithms on two synthetic images and real Bioinformatics. Hoboken, NJ, USA: Wiley, 2015.

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 13

TABLE XI. Comparison of execution times (in seconds) of ten algorithms. The best values are in bold.
FCM FGFCM HMRF-FCM FLICM NWFCM KWFLICM NDFCM Liu’s algorithm FRFCM SFFCM
The first image 0.48 1.65 43.31 57.88 48.45 64.46 4.62 23.49 0.69 0.19
The second image 0.87 2.06 47.43 121.05 80.39 133.9 4.82 23.69 0.91 0.21
BSDS500 1.15 4.67 73.43 195.53 136.20 206.70 11.26 79.88 1.87 0.74
MSRC 0.22 1.59 19.35 37.76 30.07 39.83 4.60 17.75 0.53 0.29

[17] S. Chen and D. Zhang, “Robust image segmentation using FCM with
spatial constraints based on new kernel-induced distance measure,” IEEE
Trans. Syst., Man, Cybern., B, Cybern., vol. 34, no. 4, pp. 1907-1916,
Aug. 2004.
[18] W. Cai, S. Chen, and D. Zhang, “Fast and robust fuzzy c-means cluster-
ing algorithms incorporating local information for image segmentation,”
Pattern Recognit., vol. 40, no. 3, pp. 825-838, Mar. 2007.
[19] S. Krinidis and V. Chatzis, “A robust fuzzy local information c-means
clustering algorithm,” IEEE Trans. Image Process., vol. 19, no. 5, pp.
1328-1337, May 2010.
(a) (b) [20] Z. Zhao, L. Cheng and G. Cheng, “Neighbourhood weighted fuzzy c-
means clustering algorithm for image segmentation,” IET Image Pro-
Fig. 15. Plot corresponding to Tables IX-X. (a) BSDS300. (b) cess., vol. 8, no. 3, pp. 150-161, Mar. 2014.
BSDS500. [21] F. Guo, X. Wang and J. Shen, “Adaptive fuzzy c-means algorithm based
on local noise detecting for image segmentation,” IET Image Process.,
vol. 10, no. 4, pp. 272-279, Apr. 2016.
[2] S. Zeng, X. Wang, H. Cui, C. Zheng and D. Feng, “A unified collab- [22] K. H. Memon and D. H. Lee, “Generalised fuzzy c-means clustering
orative multi-kernel fuzzy clustering for multiview data,” IEEE Trans. algorithm with local information,” IET Image Process., vol. 11, no. 1,
Fuzzy Syst., vol. 26, no. 3, pp. 1671-1687, Jun. 2018. pp. 1-12, Jan. 2017.
[3] J. Ma, S. Li, H. Qin and A. Hao, “Unsupervised multi-class co- [23] M. Gong, Y. Liang, J. Shi, W. Ma and J. Ma, “Fuzzy c-means clustering
segmentation via joint-cut over L1 -manifold hyper-graph of discrimi- with local information and kernel metric for image segmentation,” IEEE
native image regions,” IEEE Trans. Image Process., vol. 26, no. 3, pp. Trans. Image Process., vol. 22, no. 2, pp. 573-584, Feb. 2013.
1216-1230, Mar. 2017. [24] G. Liu, Y. Zhang and A. Wang, “Incorporating adaptive local information
[4] M. Gong, H. Li, X. Zhang, Q. Zhao and B. Wang, “Nonparametric into fuzzy clustering for image segmentation,” IEEE Trans. Image
statistical active contour based on inclusion degree of fuzzy sets,” IEEE Process., vol. 24, no. 11, pp. 3990-4000, Nov. 2015.
Trans. Fuzzy Syst., vol. 24, no. 5, pp. 1176-1192, Oct. 2016. [25] X. Bai, Z. Chen, Y. Zhang, Z. Liu and Y. Lu, “Infrared ship target
[5] M. Bai and R. Urtasun, “Deep watershed transform for instance seg- segmentation based on spatial information improved FCM,” IEEE Trans.
mentation,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Cybern., vol. 46, no. 12, pp. 3259-3271, Dec. 2016.
Honolulu, HI, 2017, pp. 2858-2866. [26] H. Zhang, Q. Wang, W. Shi and M. Hao, “A novel adaptive fuzzy local
[6] M. Pereyra and S. McLaughlin, “Fast unsupervised bayesian image information c-means clustering algorithm for remotely sensed imagery
segmentation with adaptive spatial regularisation,” IEEE Trans. Image classification,” IEEE Trans. Geosci. Remote Sens., vol. 55, no. 9, pp.
Process., vol. 26, no. 6, pp. 2577-2587, Jun. 2017. 5057-5068, Sept. 2017.
[7] S. Yin, Y. Qian and M. Gong, “Unsupervised hierarchical image [27] P. Ghamisi, J. A. Benediktsson and M. O. Ulfarsson, “Spectral-spatial
segmentation through fuzzy entropy maximization,” Pattern Recognit., classification of hyperspectral images based on hidden Markov random
vol. 68, pp. 245-259, Aug. 2017. fields,” IEEE Trans. Geosci. Remote Sens., vol. 52, no. 5, pp. 2565-2574,
[8] A. Krizhevsky, I. Sutskever and G. E. Hinton, “Imagenet classification May 2014.
with deep convolutional neural networks,” in Proc. Adv. Neural Inf. Proc. [28] Q. Zhao, X. Li, Y. Li and X. Zhao, “A fuzzy clustering image
Syst. (NIPS), New York, NY, 2012, pp. 1097-1105. segmentation algorithm based on hidden markov random field models
[9] E. Shelhamer, J. Long and T. Darrell, “Fully convolutional networks for and voronoi tessellation,” Pattern Recognit. Lett., vol. 85, no. 1, pp.
semantic segmentation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 49-55, Jan. 2017.
39, no. 4, pp. 640-651, Apr. 2017. [29] G. Liu, Z. Zhao and Y. Zhang, “Image fuzzy clustering based on the
[10] N. R. Pal and J. C. Bezdek, “On cluster validity for the fuzzy c-means region-level Markov random field model,” IEEE Geosci. Remote Sens.
model,” IEEE Trans. Fuzzy Syst., vol. 3, no. 3, pp. 370-379, Aug. 1995. Lett., vol. 12, no. 8, pp. 1770-1774, Aug. 2015.
[11] D. Comaniciu and P. Meer, “Mean shift: A robust approach toward [30] S. P. Chatzis and T. A. Varvarigou, “A fuzzy clustering approach toward
feature space analysis,” IEEE Trans. Pattern Anal. Mach. Intell., vol. hidden markov random field models for enhanced spatially constrained
24, no. 5, pp. 603-619, May 2002. image segmentation,” IEEE Trans. Fuzzy Syst., vol. 16, no. 5, pp. 1351-
[12] A. Y. Ng, M. I. Jordan and Y. Weiss, “On spectral clustering: Anal- 1361, Oct. 2008.
ysis and an algorithm,” in Proc. Adv. Neural Inf. Proc. Syst. (NIPS), [31] H. Zhang, Q. M. J. Wu, Y. Zheng, T. M. Nguyen and D. Wang, “Effective
Vancouver, British Columbia, 2001, pp. 849-856. fuzzy clustering algorithm with bayesian model and mean template for
[13] Y. Zhang, X. Bai, R. Fan and Z. Wang, “Deviation-sparse fuzzy c-means image segmentation,” IET Image Process., vol. 8, no. 10, pp. 571-581,
with neighbor information constraint,” IEEE Trans. Fuzzy Syst., 2018. Oct. 2014.
Doi: 10.1109/TFUZZ.2018.2883033. [32] T. Lei, X. Jia, Y. Zhang, L. He, H. Meng and A. K. Nandi, “Significantly
[14] J. Fan and J. Wang, “A two-phase fuzzy clustering algorithm based fast and robust fuzzy c-means clustering algorithm based on morpholog-
on neurodynamic optimization with its application for PolSAR image ical reconstruction and membership filtering,” IEEE Trans. Fuzzy Syst.,
segmentation,” IEEE Trans. Fuzzy Syst., vol. 26, no. 1, pp. 72-83, Feb. vol. 26, no. 5, pp. 3027-3041, Oct. 2018.
2018. [33] T. Lei, Y. Zhang, Y. Wang, S. Liu and Z. Guo, “A conditionally invariant
[15] R. Shang, P. Tian, L. Jiao, R. Stolkin, B. Hou and X. Zhang, “A spatial mathematical morphological framework for color images,” Inf. Sci., vol.
fuzzy clustering algorithm with kernel metric based on immune clone 387, pp. 34-52, May 2017.
for SAR image segmentation,” IEEE J. Sel. Topics Appl. Earth Observ., [34] M. Liu, O. Tuzel, S. Ramalingam and R. Chellappa, “Entropy rate
vol. 9, no. 4, pp. 1640-1652, Apr. 2016. superpixel segmentation,” in Proc. IEEE Conf. Comput. Vis. Pattern
[16] M. N. Ahmed, S. M. Yamany, N. Mohamed, A. A. Farag and T. Moriarty, Recognit. (CVPR), Colorado Springs, CO, 2011, pp. 2097-2104.
“A modified fuzzy c-means algorithm for bias field estimation and [35] S. Wang, H. Lu, F. Yang and M. Yang, “Superpixel tracking,” in Proc.
segmentation of MRI data,” IEEE Trans. Med. Imag., vol. 21, no. 3, IEEE Int. Conf. Comput. Vis. (ICCV), Barcelona, Spain, 2011, pp. 1323-
pp. 193-199, Mar. 2002. 1330.

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 14

[36] S. Kim, C. D. Yoo, S. Nowozin and P. Kohli, “Image segmentation using Yanning Zhang (M’08-SM’12) received the B.S.
higher-order correlation clustering,” IEEE Trans. Pattern Anal. Mach. degree in Electronic Engineering from Dalian Uni-
Intell., vol. 36, no. 9, pp. 1761-1774, Sept. 2014. versity of Technology, Dalian, China, in 1988, and
[37] J. Chen, Z. Li and B. Huang, “Linear spectral clustering superpixel,” the Ph.D. degree in Signal and Information Pro-
IEEE Trans. Image Process., vol. 26, no. 7, pp. 3317-3330, July 2017. cessing from Northwestern Polytechnical University,
[38] L. Szilágyi, Z. Benyó, S. Szilágyi and H. Adam, “MR brain image Xi’an, China, in 1996.
segmentation using an enhanced fuzzy c-means algorithm,” in Proc. 25th She is currently a Professor with the School
Annu. Int. Conf. IEEE Eng. Med. Biol. SOC., Cancun, 2003, pp. 724- of Computer Science, Northwestern Polytechnical
726. University. Her current research interests include
[39] D. Özdemir and L. Akarun, “A fuzzy algorithm for color quantization computer vision and pattern recognition, image and
of images,” Pattern Recognit., vol. 35, no. 8, pp. 1785-1791, Aug. 2002. video processing, and intelligent information pro-
[40] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua and S. Susstrunk, cessing. She was the Organization Chair of the Asian Conference on Computer
“SLIC superpixels compared to state-of-the-art superpixel methods,” Vision 2009. She served as the Program Committee Chair of several interna-
IEEE Trans. Pattern Anal. Mach. Intell., vol. 34, no. 11, pp. 2274-2282, tional conferences.
Nov. 2012.
[41] Z. Hu, Q. Zou and Q. Li, “Watershed superpixel,” in Proc. IEEE Int.
Conf. Image Process. (ICIP), Quebec City, QC, 2015, pp. 349-353.
[42] T. H. Kim, K. M. Lee and S. U. Lee, “Learning full pairwise affinities
for spectral segmentation,” IEEE Trans. Pattern Anal. Mach. Intell., vol.
35, no. 7, pp. 1690-1703, Jul. 2013. Shigang Liu received the B.S. and M.S. degrees
[43] B. Wang and Z. Tu, “Affinity learning via self-diffusion for image from Harbin Engineering University, Harbin, China,
segmentation and clustering,” in Proc. IEEE Conf. Comput. Vis. Pattern in 1997 and 2001, respectively. He received his
Recognit. (CVPR), Providence, RI, 2012, pp. 2312-2319. Ph.D. degrees from Xidian University of China, in
[44] L. Vincent, “Morphological Grayscale reconstruction in image analysis: 2005.
applications and efficient algorithms,” IEEE Trans. Image Process., vol. From 2009 to 2010, he was a Visiting Scholar with
2, no. 2, pp. 176-201, Apr. 1993. the Department of Computing at The Hong Kong
[45] P. Arbelaez, M. Maire, C. Fowlkes and J. Malik, “Contour detection Polytechnic University. From 2015 to 2016, he was
and hierarchical image segmentation,” IEEE Trans. Pattern Anal. Mach. a Visiting Scholar with the Quantum Computation
Intell., vol. 33, no. 5, pp. 898-916, May 2011. and Intelligent System group at University of Tech-
[46] J. Shotton, J. Winn, C. Rother and A. Criminisi, “Textonboost: joint nology Sydney, Sydney, Australia. He is currently a
appearance, shape and context modeling for multi-class object recogni- Professor with the School of Computer Science, Shaanxi Normal University.
tion and segmentation,” in Proc. Eur. Conf. Comput. Vis. (ECCV), Graz, His research interests include pattern recognition and image processing.
Austria, 2006, pp. 1-15.
[47] X. Wang, Y. Tang, S. Masnou and L. Chen, “A global/local affinity
graph for image segmentation,” IEEE Trans. Image Process., vol. 24,
no. 4, pp. 1399-1411, Apr. 2015.
[48] A. Rodriguez and A. Laio, “Clustering by fast search and find of density Hongying Meng (M’10-SM’17) received his Ph.D.
peaks,” Science, vol. 344, no. 6191, pp. 1492-1496, Jun. 2014. degree in Communication and Electronic Systems
[49] J. Hou and W. X. Liu, “A parameter independent clustering framework,” from Xi’an Jiaotong University, Xi’an China, in
IEEE Trans. Ind. Informat., vol. 13, no. 4, pp. 1825-1832, Aug. 2017. 1998. He is currently a senior lecturer at the De-
partment of Electronic and Computer Engineering,
Brunel University London, U.K. He is also a mem-
ber of Institute of Environment, Health and Soci-
Tao Lei (M’17) received the Ph.D degree eties, Human Centred Design Institute (HCDI), and
in Information and Communication Engineering Wireless Networks and Communications Research
from Northwestern Polytechnical University, Xi’an, Center at Brunel. He is a Fellow of The Higher
China, in 2011. Education Academy (HEA) and a member of En-
From 2012 to 2014, he was a Postdoctoral Re- gineering Professors Council in UK.
search Fellow with the School of Electronics and Dr Meng’s current research interests include digital signal processing,
Information, Northwestern Polytechnical University, affective computing, machine learning, human computer interaction, computer
Xi’an, China. From 2015 to 2016, he was a Visiting vision, and embedded systems with over 90 publications in these areas.
Scholar with the Quantum Computation and Intel- Especially, his audio based and video based emotion recognition systems
ligent Systems group at University of Technology have won the International Audio/Visual Emotion Challenges AVEC2011 and
Sydney, Sydney, Australia. He is currently a Pro- AVEC2013 prizes respectively. He is a senior member of the IEEE.
fessor with the School of Electrical and Information Engineering, Shaanxi
University of Science and Technology. His current research interests include
image processing, pattern recognition, and machine learning.

Xiaohong Jia received the M.S. degree in Signal

and Information Processing in Lanzhou Jiaotong
University, Lanzhou, China, in 2017. He is currently
pursuing the Ph.D. degree at the School of Electroni-
cal and Information Engineering, Shaanxi University
of Science and Technology, Xi’an 710021, China.
His current research interests include image pro-
cessing and pattern recognition.

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TFUZZ.2018.2889018, IEEE
Transactions on Fuzzy Systems
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. XXX, NO. XXX, XXX 2018 15

Asoke K. Nandi (F’11) received the degree of Ph.D.

in Physics from the University of Cambridge (Trin-
ity College), Cambridge (UK). He held academic
positions in several universities, including Oxford
(UK), Imperial College London (UK), Strathclyde
(UK), and Liverpool (UK) as well as Finland Dis-
tinguished Professorship in Jyvaskyla (Finland). In
2013 he moved to Brunel University London (UK),
to become the Chair and Head of Electronic and
Computer Engineering. Professor Nandi is a Dis-
tinguished Visiting Professor at Tongji University
(China) and an Adjunct Professor at University of Calgary (Canada).
In 1983 Professor Nandi co-discovered the three fundamental particles
known as W + , W − and Z 0 , providing the evidence for the unification
of the electromagnetic and weak forces, for which the Nobel Committee
for Physics in 1984 awarded the prize to his two team leaders for their
decisive contributions. His current research interests lie in the areas of signal
processing and machine learning, with applications to communications, gene
expression data, functional magnetic resonance data, and biomedical data. He
has made many fundamental theoretical and algorithmic contributions to many
aspects of signal processing and machine learning. He has much expertise in
“Big Data”, dealing with heterogeneous data, and extracting information from
multiple datasets obtained in different laboratories and different times.
He has authored over 570 technical publications, including 230 journal
papers as well as four books, entitled Automatic Modulation Classification:
Principles, Algorithms and Applications (Wiley, 2015), Integrative Cluster
Analysis in Bioinformatics (Wiley, 2015), Blind Estimation Using Higher-
Order Statistics (Springer, 1999), and Automatic Modulation Recognition of
Communications Signals (Springer, 1996). Recently he published in Blood,
International Journal of Neural Systems, BMC Bioinformatics, IEEE TWC,
NeuroImage, PLOS ONE, Royal Society Interface, Mechanical systems and
Signal Processing, and Molecular Cancer. The h-index of his publications is
70 (Google Scholar) and ERDOS number is 2.
Professor Nandi is a Fellow of the Royal Academy of Engineering (UK)
and also a Fellow of seven other institutions, including the IEEE and the
IET. Among the many awards he received are the Institute of Electrical and
Electronics Engineers (USA) Heinrich Hertz Award in 2012, the Glory of
Bengal Award for his outstanding achievements in scientific research in 2010,
the Water Arbitration Prize of the Institution of Mechanical Engineers (UK) in
1999, and the Mountbatten Premium, Division Award of the Electronics and
Communications Division, of the Institution of Electrical Engineers (UK) in
1998. Professor Nandi is an IEEE EMBS Distinguished Lecturer (2018-19).

1063-6706 (c) 2018 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.