information

Article
Edge Detection Method Based on General Type-2
Fuzzy Logic Applied to Color Images
Claudia I. Gonzalez, Patricia Melin * ID
and Oscar Castillo ID

Tijuana Institute of Technology, Tijuana 22379, Mexico; cgonzalez@tectijuana.mx (C.I.G.);

ocastillo@tectijuana.mx (O.C.)
* Correspondence: pmelin@tectijuana.mx; Tel.: +52-664-6236318

Received: 1 July 2017; Accepted: 25 August 2017; Published: 28 August 2017

Abstract: This paper presents a new general type-2 fuzzy logic method for edge detection applied
to color format images. The proposed algorithm combines the methodology based on the image
gradients and general type-2 fuzzy logic theory to provide a powerful edge detection method. General
type-2 fuzzy inference systems are approximated using the α-planes approach. The edge detection
method is tested on a database of color images with the idea of illustrating the advantage of applying
the fuzzy edge detection approach on color images against grayscale format images, and also when
the images are corrupted by noise. This paper compares the proposed method based on general
type-2 fuzzy logic with other edge detection algorithms, such as ones based on type-1 and interval
type-2 fuzzy systems. Simulation results show that edge detection based on a general type-2 fuzzy
system outperforms the other methods because of its ability to handle the intrinsic uncertainty in
this problem.

Keywords: α-planes; general type-2 fuzzy sets; general type-2 fuzzy sets; interval type-2 fuzzy sets;
type-1 fuzzy sets; image processing; color edge detection; computer vision

1. Introduction
Edge detection is a fundamental technique in many computer vision and image processing
applications, such as pattern recognition, medical image processing, object recognition, and motion
analysis. In the past few decades most research has concentrated on designing edge detector algorithms
for grayscale images; however, color image processing applications have recently been receiving
increasing attention because color images provide more information about the objects in a scene than
grayscale images and this information can be used to improve the performance of image processing
systems [1]. Edge detection in color images is computationally more complex than in grayscale images,
but there are many advantages to using color images. For example, the increase in the amount of
information can lead to more accurate object location and the possibility of processing images that are
more complex [2].
Various edge detection techniques have been proposed over the years, but the common approach
is to apply the first or second derivative. Based on this, the edge detection can be classified as gradient
edge detectors (first derivative), Laplacian method (second derivative), or Gaussian edge detectors [3].
In the fuzzy system area, there are edge detection methods that have been developed for grayscale
images [4–8]. In Gonzalez et al. [9], an improved method for processing edge detection based on
general type-2 fuzzy sets (GT2 FSs) applied to grayscale images is proposed. According with the
results presented in [9], the author demonstrated that the edge detection approach based on GT2 FSs
outperformed the results obtained by the methods based on interval type-2 fuzzy sets (IT2 FSs), type-1
fuzzy sets (T1 FSs), and of course the traditional edge detection methods.
The contribution of this paper is focused on exploiting the benefits of using digital images in
color format and combining this process with GT2 FSs, which has shown satisfactory results in image

Information 2017, 8, 104; doi:10.3390/info8030104 www.mdpi.com/journal/information

Information 2017, 8, 104 2 of 15

processing applications; especially, when the image is corrupted by noise, this kind of technique is
helpful to handle the uncertainty, so we are proposing a novel method of edge detection based on
GT2 FSs to be applied to color images. The process is designed using Lab, HSV, and RGB to show
the advantages of using one kind of color format with respect to the others. The inputs and outputs
of the fuzzy system are built with Gaussian membership functions (MFs), which have demonstrated
good results in these kinds of applications. Moreover, we present a comparative analysis of the results
achieved with this edge detection approach versus the results achieved by using TI and IT2 FSs.
The paper is organized as follows. In Section 2 important definitions about GT2 FSs are presented.
Section 3 describes the color image edge detector using the gradient approach. Section 4 shows the
color image edge detector based on GT2 FSs. In Section 5 the simulation results are presented and
explained. Finally, Section 6 offers some conclusions about the results.

2. Fuzzy Logic Systems

Fuzzy sets theory was introduced by Lofti A. Zadeh in 1965 [10]. Fuzzy sets provide a way to
model the uncertainty associated with vagueness, imprecision, and lack of information regarding a
problem. The uncertainty arises from ignorance due to a lack of knowledge, from randomness, or from
vagueness, like the fuzziness existing in human language [10,11]. In recent years we have witnessed a
rapid growth in a number of applications of fuzzy logic in various areas, such as pattern recognition,
neural networks, expert systems, artificial intelligence, theory control, automata theory, computer
science, decision making, medical diagnosis, and robotics, to name a few. In digital image processing,
which is the area of interest of this paper, there are some applications, like in edge detection [5,6,8],
noise filtering [12,13], feature extraction [14], classification and clustering [2], and it also has been
applied in video processing such as in background/foreground separation, and in 3D computer
vision [15–17].
Advances in fuzzy sets theory have made it possible to research and apply more complex forms of
fuzzy systems. In addition, great progress has been made in transitioning from traditional type-1 FS to
interval type-2 FS [18–21], and recently there has been a growing interest in general type-2 FS [22–25].
In what follows, we introduce the basic concepts and properties about general type-2 FS and the
α-planes representation, which are important because the main contribution of this paper is based on
the use of these concepts.

2.1. Definition of General Type-2 Fuzzy Sets

A general type-2 fuzzy set (GT2 FS) A,
e can be represented by Equation (1) [22,24]:
 
e = ( x, u), µ e ( x, u) ∀ x ∈ X, ∀u ∈ [0, 1] ,
A (1)
A

where X is the universe for the primary variable of A, e x. The 3D membership function is usually
denoted by µ Ae ( x, u), where x e X and u e U ⊆ [0, 1] and 0 ≤ µ Ae ( x, u) ≤ 1. If A
e is continuous it
can be expressed as Z 
Z
A=
e µ Ae ( x, u)/( x, u) , (2)
x∈X u∈[0,1]
RR
where denotes the union over all admissible x and u. For discrete universes of discourse, A
e can be
denoted by Equation (3):  
 
e = ∑ ∑ µ e ( x, u)/( x, u) .
A (3)
 x∈X u∈[0,1] A 

In Equations (1)–(3) u is called the secondary variable, and has the domain U = [0, 1] at each
x e X. When µ Ae ( x, u) = 1 for ∀ x ∈ X and ∀u ∈ U then A e is called an interval type-2 fuzzy set
(IT2 FS).
Information 2017, 8, 104 3 of 15

At each point of x, say x = x 0 , the 2D plane whose axes are u and µ Ae ( x, u) is called a vertical
slice of µ Ae ( x, u). A secondary membership function is a vertical slice of µ Ae ( x, u), i.e.,

Z
µ Ae x = x 0 , u ≡ µ Ae x 0 =


f x (u)/u. (4)
u∈ J u0 ⊆[0,1]
x

In Equation (4) Jxu0 is the subset of U that is the support of µ Ae ( x 0 ), and is called the primary
e The amplitude of the secondary membership function f x (u), is called the secondary
membership of A.
grade or secondary set.
The two-dimensional support of µ Ae ( x, u) is called the footprint of uncertainty (FOU) of A, e i.e.,
 
e) = ( x, u) ∈ X × [0, 1]µ e ( x, u) > 0 .
FOU ( A (5)
A

FOU ( A
e) can also be expressed as the union of all primary memberships, i.e.,

e) = ∪ x∈X Jxu .
FOU ( A (6)

2.2. General Type-2 Fuzzy Sets Representation

In order to limit the complexity of general type-2 fuzzy logic, several efforts have been proposed;
for example, Wagner and Hagras [1,26] have introduced the zSlices representation, and Mendel and
Liu [23,27,28] have put forward a representation based on the α-plane concept, both of which enable
the representation of, and computation with, general type-2 fuzzy sets.
An α-plane for the GT2 FS A,e denoted by A eα , is the union of all primary memberships functions
of A whose secondary grades are greater than or equal to α (0 ≤ α ≤ 1) [24]. The equation of the
e
α-plane is represented by Equation (7):
  Z Z
e α = ( x, u), µ e ( x, u) ≥ α∀ x ∈ X, ∀u ∈ ∀u ⊆ [0, 1] =
A {( x, u)| f x (u) ≥ α}. (7)
A
∀ x ∈ X ∀u∈[0, 1]

3. Edge Detection Process Based on the Gradient Approach

The edge detection process in a digital image consists in identifying the pixels, where the
brightness changes dramatically or has discontinuities with respect with their neighbor pixels.
There exist some approaches to perform the edge detection process; the majority are based on image
gradients, which are calculated with the first derivative of an image. The most important and most used
methods based on the image gradients are from Kirsch [29], Sobel [30], Prewitt [31], and Canny [32].
In these techniques the Euclidean distance is used to calculate the image gradients in conjunction with
a kernel [33].
The edge detection method presented in this paper consists of calculating the image gradients
with the Euclidean distance, but without using any kernel. This process consists of obtaining four
image gradients based on a 3 × 3 matrix to indicate the edge direction ( Di , for i = 1 . . . 4) (Figure 1).
Each matrix position (Pi ) of Figure 1 is represented in Figure 2, where f represents the image,
x-axis the columns and y-axis the rows. According with these positions, the Euclidean distance
is applied to calculate the gradients Di using the expressions illustrated in Figure 3a–d and the
gradient magnitude or the output gradient is obtained with the aggregation of the four gradients
( FE = D1 + D2 + D3 + D4).
Each matrix position (P ) of Figure 1 is represented in Figure 2, where represents the image,
x-axis the columns and y-axis the rows. According with these positions, the Euclidean distance is
applied to calculate the gradients D using the expressions illustrated in Figure 3a–d and the
gradient magnitude or the output gradient is obtained with the aggregation of the four gradients
( = 12017,
Information + 8,2 104
+ 3 + 4). 4 of 15

Information 2017, 8, 104 Figure 1. 3 × 3 matrix to indicate direction of the four gradients Di .
Information 2017, 8, 104 Figure 1. 3 × 3 matrix to indicate direction
4 of 15
4 of 15

Figure 2.
Figure 2. Matrix
Matrix position.
position.

Figure3.3.Positions
Figure Positionsto
to calculate the gradients
calculate the D1,D2,
gradientsD1, D2,D3,
D3,and D4.
andD4.

Figure 3. Positions to calculate the gradients D1, D2, D3, and D4.
Gradient Approach
Gradient Edge
Approach EdgeDetection
DetectionApplied
AppliedononColor
Color Format Images
Format Images
In In
Gradient the literature
the literature
Approach Edgewewe commonly
commonly
Detection find
findedge
Applied edge
on detection
Color methods
Formatmethods appliedon
Images applied ongrayscale
grayscale format
format images
images
(two-dimensionalimages);
(two-dimensional images);this thisisisbecause
because thethe computational
computational complexity
complexityisisreduced,
reduced, but thethe
but image
image
In the literature we commonly find edge detection methods applied on grayscale format images
could
could loselose information
information whenwhenit is itcaptured
is captured in color
in color formatformat
and itand it is changed
is changed to grayscale
to grayscale to be
to be processed.
(two-dimensional
processed. The images);
aim of this isisbecause the thecomputational complexity is reduced, but the image
The aim of this paper is this paper
to perform to perform
the edge detection edgeprocess
detection process
using imagesusingin images
color. in color.
could lose
There information
exist several when
models it is captured in color format and it is changed to grayscale to be
There exist several models to to represent
represent color
color images;
images; of of
thethe most
most usedusedwewe could
could mention
mention thetheHSV
processed.
HSV the Thethe
model, aimLab
of this paper
model, isRGB
themodel, to perform
model, the CMY
the edge color
detection process
model, using
the model,
YUV images
model, thein color.
HIS color
model, Lab model, the RGB the CMY color model, the YUV the HIS color model,
Theretheexist several models to represent color images; of thedetection
most used we could mention the
themodel,
CMYK CMYK
color color[33–35],
model model [33–35], etc. In
etc. In this this paper
paper the detection
the edge edge methodmethod is applied
is applied on the
on the HSV,
HSVHSV,model, the Lab
Lab, and RGBmodel, the RGB
color models. HSVmodel,
standstheforCMYhue, color model,
saturation, andthevalue
YUV(brightness);
model, the itHIS color
is the
model,
most the CMYK
common colorin
model model [33–35],
computer etc.applications.
vision In this paperLab the color
edge space
detectionis a method
three-axisis color
applied on the
system
HSV,
withLab, and RGB color
the dimensions L formodels.
lightnessHSV and astands
and b forfor color;
hue, saturation,
working with andthevalue (brightness);
Lab color it is the
space includes
most common
all of the colors model
in thein computer
spectrum, as vision
well as applications.
colors outside Lab color perception.
of human space is a three-axis
In the RGBcolor model,system
an
image
with the consists
dimensions of three
L forindependent
lightness and image
a andplanes,
b for one in each
color; of the
working primary
with the Labcolors:
colorred, green
space and
includes
Information 2017, 8, 104 5 of 15

Lab, and RGB color models. HSV stands for hue, saturation, and value (brightness); it is the most
common model in computer vision applications. Lab color space is a three-axis color system with the
dimensions L for lightness and a and b for color; working with the Lab color space includes all of the
colors in the spectrum, as well as colors outside of human perception. In the RGB model, an image
consists of three independent image planes, one in each of the primary colors: red, green and blue;
so, this model can be considered the fundamental color representation in computers, digital cameras,
and scanners.
To apply the gradient approach on a color format image (Lab, HSV, and RGB), we need to calculate
the four gradients (Steps 5 to 8), and after that the gradient magnitude (Step 11) for each channel and
finally the output edge (Step 16). Considering that the Lab, HSV, and RGB color formats, the gradients
are calculated for three channels. The step to process this is described in the following pseudocode:
Information 2017, 8, 104 5 of 15
Pseudocode to calculate the gradients for each channel of a color format image ( f ).


Input. Theformats,
color format image are
the gradients f r,c,dcalculated
where (rfor
) isthree
the rows size ofThe
channels. thestep
image, (c) the this
to process columns size andin(dthe
is described ) the
channel number;
followingd =pseudocode:
1, 2, 3.
Output. The gradients ( D1d , D2d , D3d , and D4d ), the gradient magnitude ( Gd ) for each channel of the image
(f), and the output calculate the gradients for each channel of a color format image ( ).
edgeto(Edge).
Pseudocode
1: Input.the
Calculate TheD1color
d , format
D2d , image
D3d , and , , D4 d for (each
where ) is(the
d) rows size to
channel of obtain ( Gd() ) the columns size and
the image,
2: ( ) the
for d = 1 to 3channel number; d = 1, 2, 3.
Output. The gradients ( 1 , 2 , 3 , 4 ), the gradient magnitude ( ) for each channel of the
3: for x = 2 to r − 1
image (f), and the output edge (Edge).
4: for y1:= 2qto c − 1
Calculate the 1 , 2 , 3 , 4 for each ( ) channel to obtain ( )

5: y)3− f ( x, y − 1))2 + ( f ( x, y) − f ( x, y + 1))2
D1d2:= for( df (=x,1 to
3: for x = 2 to − 1
q
6: D2d =
4: q
( f ( x, y) − f ( x − 1, y))2 + ( f ( x, y) − f ( x + 1, y))2
for y = 2 to − 1
7: D3d5:= 1( f = ( x, y() (−, f)(−
x −(1,, y − 1))2 ++
−1)) ( ((f (, x,) −
y) −( , f (+
x+1))1, y + 1))
2

2 = ( ( , ) − ( − 1, )) 2 + ( ( , ) − ( + 1, ))

q
6:
8: D4d = ( f ( x, y) − f ( x + 1, y − 1)) + ( f ( x, y) − f ( x − 1, y + 1))2
7: 3 = ( ( , ) − ( − 1, − 1)) + ( ( , ) − ( + 1, + 1))
9: end
8: 4 = ( ( , ) − ( + 1, − 1)) + ( ( , ) − ( − 1, + 1))
10: end 9: end
11: Gd =10:D1 d + D2d + D3d + D4d
end
12: Normalize
11: = Gradient
the 1 + 2 G+ d in3values
+ 4 between {0, 1}
12:G Normalize
− the Gradient in values
( Gbetween {0, 1}
Gd = d min ( min ( G d )) /max (max d )) − min ( min ( Gd )), where min and max represent the
13: 13: = minimum
− ( ) ( ) − ( ) , where min and max represent the
maximum and pixel value of Gd , respectively
maximum and minimum pixel value of , respectively
14: end
14: end
15: Calculate the output
15: Calculate the edge
output(Edge)
edge (Edge)
16: Edge16:= ∑dd= 3
=1=G d∑

An example of the of
An example output ( Edge
the output ( ) achieved
) achievedby theprevious
by the previous pseudocode
pseudocode is illustrated
is illustrated in Figure 4.in Figure 4.

(a) (b)
Figure 4. Output color edge detection based on the gradient approach. (a) Input image; (b) edge
Outputoutput.
Figure 4. detection color edge detection based on the gradient approach. (a) Input image; (b) edge
detection output.
4. Edge Detection Process Based on the Gradient Approach and GT2 FS
This section presents an edge detection method based on the gradient approach combined with
GT2 FSs to be applied on color format images (Lab, HSV, and RGB). The motivation of using GT FSs
in the proposed edge detection method is to provide the ability to handle higher degrees of
uncertainty in the edge detection process for color format images.
Information 2017, 8, 104 6 of 15

4. Edge Detection Process Based on the Gradient Approach and GT2 FS

This section presents an edge detection method based on the gradient approach combined with
GT2 FSs to be applied on color format images (Lab, HSV, and RGB). The motivation of using GT FSs in
the proposed edge detection method is to provide the ability to handle higher degrees of uncertainty
in the edge detection process for color format images.
The methodology to develop the proposed method consists of three steps. At first, the
input color image ( f ) is separated into (d, for d = 1, . . . , 3) channels. Secondly, the gradient
approach is applied in each single channel to obtain the four gradients ( D1d + D2d + D3d + D4d );
these gradients are calculated with the Steps 5 to 8 of the pseudocode presented in Section 3.
Finally, the 12 (four for each channel) gradients previously obtained are identified as
( D11 , D21 , D31 , D41 , D12 , D22 , D32 , D42 , D13 , D23 , D33 , D43 ) and these are used as the
inputs to the GT2 FSs.
The GT2 fuzzy system is a singleton Mamdani type, which was approximated using α-planes [24,27],
where each α-plane centroid was computed by the Karnik–Mendel algorithm [19]. In more detail, the GT2
FSs was designed with 12 inputs (D11 , D21 , D31 , D41 , D12 , D22 , D32 , D42 , D13 , D23 , D33 , D43 )
and three outputs (G1 , G2 and G3 ) to generate the gradient magnitude for each image channel. According
to the pseudocode presented in Section 3, the ( G1 , G2 and G3 ) are calculated with the aggregation (Step
11) of the four gradients obtained in each channel. In the proposed method, the idea is to eliminate
this step, and the G1 , G2 , and G3 values are inferred directly with a GT2 fuzzy system.
The inputs and outputs are fuzzified using GT2 Gaussian membership functions with uncertain
mean (gausmgausstype2), each input has three linguistic values (low, middle, and high) to determine
the grade that the evaluated gradient correspond to be the output edge. Each output is granulated in
two linguistic values (background and edge) to produce the gradient magnitude for each color image
channel and the combination of these generates the output color edge detection.
The parameters required for the GT2 MFs (gausmgausstype2) are expressed in Equation (8), and
these are calculated depending of the image gradient values, i.e., considering the image of Figure 4a,
the parameters are obtained with Equations (9) to (14).

e( x, u) = gausmgausstype2( x, u, [σx , m1 , m2 ])
µ (8)

low = min(Di ) (9)

high = max(Di ) (10)

medium = low + (high − low)/2 (11)

σ = high/4 (12)

m1 = high (13)

m2 = m1 + (m1 *FOU), where FOU is in (0, 1) (14)

In Figure 5 an example of the membership functions for the D11 , D21 , D31 , and D41 inputs
are illustrated, where the parameters were calculated with the previous equations. These inputs
correspond to the first channel of the color image. The output for the first channel ( G1 ) is shown in
Figure 6.
Another important part of a fuzzy inference system is the fuzzy rules; for this proposed
method the fuzzy rules were designed considering different combinations of the gradients
( D11 , D21 , D31 , D41 , D12 , D22 , D32 , D42 , D13 , D23 , D33 , D43 ) obtained from the different
channels of the color images to infer the gradient magnitude (outputs) for each image channel
( G1 , G2 , G3 ). In our approach, the fuzzy rules were separately combined according to the color
image channel, based on experimental results and human experts in this kind of applications, with the
idea of reaching the best accuracy edge detection with a minimum number of rules. After several tests,
Information 2017, 8, 104 7 of 15

the best performance

Information 2017, 8, 104 was obtained by the set of nine fuzzy rules presented in Table 1; these fuzzy 7rules
of 15
Informationthe
represent 2017, 8, 104
knowledge about the edge detection process. 7 of 15

Figure
Figure 5.5.5.
Figure GT2
GT2 input
input
GT2 membership
inputmembership function
membershipfunction forthe
function for
for thevariables 1 11,, , D2
variablesD1
variables 2 2,1 ,, 3D3
3, and
, and .4 1. .
1 , and4 D4

Figure 6. GT2 output membership function for the variable .

Figure 6.
Figure GT2 output
6. GT2 output membership
membership function
function for
for the variable G1 ..
the variable

Table 1. Knowledge base of nine fuzzy rules for edge detection process.
Table 1. Knowledge base of nine fuzzy rules for edge detection process.
process.
Fuzzy Rules
1. If (RD1 is HIGH) or (RD2 is HIGH) or (RD3 Fuzzy isFuzzy
HIGH) Rules
or (RD4 is HIGH) then (GR is EDGE)
Rules
1. If (RD1 is HIGH) or (RD2 is HIGH) or (RD3 is HIGH) or (RD4 is (RD4 HIGH) then (GR is EDGE)
1. If (RD1 is HIGH) or (RD2 is HIGH) or (RD3 is HIGH) or (RD4 isorHIGH)
2. If (RD1 is MIDDLE) or (RD2 is MIDDLE) or (RD3 is MIDDLE) is MIDDLE)
then (GR is EDGE)then (GR is EDGE)
2. 2.
If (RD1
3. If
If (RD1
(RD1isis MIDDLE)
is MIDDLE) or (RD2 is MIDDLE) or (RD3 is MIDDLE) or (RD4 is MIDDLE) then (GR(GR
LOW) andor (RD2
(RD2 isis MIDDLE)
LOW) and or
(RD3 (RD3is is
LOW) MIDDLE)
and (RD4 oris (RD4
LOW) is MIDDLE)
then (GR is then
BACKGROUND) is EDGE)
is EDGE)
3. 3.
If (RD1
4. If
If (RD1
(GD1isisis
LOW) and
HIGH) or (RD2
(GD2 is
is LOW)
HIGH) and
or (RD3
(GD3 is is LOW)
HIGH) and
or (GD4(RD4is is LOW)
HIGH)
LOW) and (RD2 is LOW) and (RD3 is LOW) and (RD4 is LOW) then (GR is BACKGROUND) thenthen(GG (GR is is BACKGROUND)
EDGE)
5. If
4. 4.
If If (GD1
(GD1
(GD1 isisis MIDDLE)
HIGH)
HIGH)or or (GD2
or(GD2
(GD2 is is MIDDLE)
is HIGH)
HIGH) (GD3
or (GD3 or is(GD3
isHIGH)
HIGH)is MIDDLE)(GD4isor
oror(GD4 (GD4
isHIGH)
HIGH) is
thenMIDDLE)
then(GG (GG then (GG is EDGE)
is EDGE)
is EDGE)
5. 5.
6. If
If If (GD1
(GD1
(GD1 isisis MIDDLE)
LOW) and
MIDDLE) or(GD2
or (GD2isis
(GD2 isLOW)
MIDDLE) or
or(GD3
and (GD3
MIDDLE) (GD3 isisMIDDLE)
is LOW) and (GD4
MIDDLE) oror (GD4
is(GD4
LOW) is is
MIDDLE)
then
MIDDLE)(GGthen (GG(GG
is EDGE)
is BACKGROUND)
then is EDGE)
6. 6.
7. If
If If (GD1
(BD1isisLOW)
(GD1 LOW)
HIGH)and
and (GD2isis
or (BD2
(GD2 isHIGH)
LOW) or
LOW) and
and (GD3
(BD3
(GD3 isisLOW)
is HIGH) LOW) orand
(BD4
and (GD4 is is
is HIGH)
(GD4 LOW)
LOW)then then(GB
then (GG
is(GG is BACKGROUND)
EDGE)is BACKGROUND)
7.
8. If
If (BD1
(BD1 is HIGH) or (BD2 is HIGH) or (BD3 orisis HIGH) or (BD4 isorHIGH) then (GB is EDGE)
7. If (BD1 is is MIDDLE)
HIGH) or (BD2
or (BD2 is MIDDLE)
is HIGH) or (BD3 (BD3 HIGH) is MIDDLE)
or (BD4 is (BD4 HIGH) is MIDDLE)
then (GB is then (GB is EDGE)
EDGE)
8.
9. If
If (BD1
(BD1 is
is MIDDLE)
LOW) and or (BD2
(BD2 is is MIDDLE)
LOW) and or
(BD3(BD3 is is
LOW) MIDDLE)
and (BD4or (BD4
is LOW)is MIDDLE)
then (GB then
is (GB is EDGE)
BACKGROUND)
8. 9.
If If
(BD1
(BD1isisMIDDLE)
LOW) andor(BD2 (BD2isisLOW) MIDDLE)
and (BD3or (BD3is LOW) is MIDDLE)
and (BD4or (BD4 isthen
is LOW) MIDDLE) then (GB is EDGE)
(GB is BACKGROUND)
9. If (BD1 is LOW) and (BD2 is LOW) and (BD3 is LOW) and (BD4 is LOW) then (GB is BACKGROUND)
5. Experimental Results
5. Experimental Results
5. Experimental
This sectionResults
presents the edge detection accuracy and noise robustness achieved by the
This section
proposed method presents thethe
based on edge detection
gradient accuracy
technique andand
GT2noise robustnessinachieved
FSs (explained Section 4)bywhen
the proposed
this is
This section presents the edge detection accuracy and noise robustness achieved by the
method
appliedbased on the
on color gradient
images technique
in the Lab, HSV,and GT2
and FSsformats.
RGB (explained
Theinexperiments
Section 4) when
were this is applied
applied on the on
proposed method based on the gradient technique and GT2 FSs (explained in Section 4) when this is
color images
image in the
database Lab, HSV,
presented in and
TableRGB formats.
2 and The experiments
the performance were applied
of the proposed on isthe
method image database
compared with
applied on color images in the Lab, HSV, and RGB formats. The experiments were applied on the
two other fuzzy edge detection approaches (T1 and IT2 FSs), and with the two
presented in Table 2 and the performance of the proposed method is compared with two other fuzzyclassic methods
image database presented in Table 2 and the performance of the proposed method is compared with
(Canny
edge and Sobel).
detection approaches (T1 and IT2 FSs), and with the two classic methods (Canny and Sobel).
two other fuzzy edge detection approaches (T1 and IT2 FSs), and with the two classic methods
(Canny and Sobel).
Information 2017, 8, 104 8 of 15

Information
Information 2017,
2017, 8, 104
8, 104 8 of815
of 15
Information
Information 2017,
2017, 8, 104
8, 104
Table 2. Synthetic color image database. 8 of815
of 15
Information
Information 2017,
Information 2017,
8, 1048, 1048, 104
2017, 8,2017,
Information 104 8 of 15 8 of 15 15
8 of 8
15of
Table
Table 2. Synthetic
2. Synthetic color
color image
image database.
database.
Information
Information
Image 2017,
2017,
Number 8, 104
8, 104 Synthetic Images
Table
Table 2. Synthetic
2. Synthetic color Reference
color image
image Images
database.
database. 8 of815
Resolution of 15
ImageImage Number
Number Table
Table
Table
Synthetic
Synthetic 2. 2. Synthetic
2.Images
Synthetic
Synthetic
Table
Images colorcolor
color
2. Syntheticimage
image
colorimage
image
Reference
Reference database.
database.
database.
database.
Images
Images Resolution
Resolution
ImageImage NumberNumber Synthetic
Synthetic Images
Table
Table Images
2. 2. Synthetic
Synthetic color colorReference
Reference
imageimage ImagesImages
database.
database. Type Resolution
Resolution
of file: PNG
ImageImage
Image Number
Image
Number Number NumberSynthetic Synthetic
Synthetic Images
Synthetic
Images Images
Images Reference
Reference Reference
Reference Images
Images Images
Images Type of
Resolution file:
Resolution PNG
Resolution
Resolution filefile
Type Type
Type of offile:
of
file: file:
PNG PNG
PNG file file
file
ImageImage NumberNumber Synthetic
Synthetic ImagesImages Reference
Reference Images Type
Images Type Type
TypeWidth:
Width:
of file:
of ofPNG 420
file:
of file:420
Resolution
file:
Resolution PNG pixels
pixels
file
PNG PNG filefile
file
Width: Width:
Width:420
Width:
Width: 420
pixels
420 420
420 pixels
pixels
pixelspixels
1 1 1 Type Width:
Type
Width: of
420
of file: 420
file:
PNG pixels
PNG
pixels filefile
1 11 1 Width:
Width:
Height:
Height: 420
182 420 pixels
pixels
182 pixels
pixels
1 1 Height: 182 pixels
Height: Height:
Height:
182 pixels
Height: 182182
182 pixels
pixels
pixels
1 1 Height:Height: 182 182 pixelspixels
Type Height:
Height:
Type of182
of file: 182
file:
PNG pixels
pixels
PNG filefile
TypeType of
Typefile:
Type ofofPNGfile:
of
file: file
PNG
file:
PNG PNG filefile
Type TypeWidth:
Width:
Type
of of 143
of 143
file: pixels
pixels
file: PNG
PNGfile filefile
Width: 143file:
Width:
Width:
Width: 143
PNG
pixels
143 143pixels
pixels
pixels
Width:
Type Width:
Type 143
of file:
Width: 143
pixels
of file:
PNG
143 pixels
PNG filefile
pixels
2 2 22 2
2 2 Width:
Width:
Height:
Height: 143
241 143 pixels
pixels
241 pixels
pixels
2 2 Height: 241 pixels
Height:
Height: 241241 pixelspixels
Height:
Height: 241241 pixels
pixels
2 2 Height:Height: 241 241 pixelspixels
TypeType of
Type
Height:
Height:
Type of
ofofPNG
file: 241
file:
file:
241
file:
PNG
file
PNG
pixels
pixels
PNG filefile
file
Type
Width: Type of
219
Width:
Width: of
file:file:
pixels
219 PNG
219 PNG filefile
pixels
pixels
Type Width:
Type
Typeof file: 219
ofoffile:
PNG pixels
file:PNGPNG
file filefile
Width:
Width: 219219 pixels
pixels
3 33 3 Width:
Type Width:
Type of 219 219
pixels
of file:
file: PNG
Width: 219 pixels
pixels
PNG file file
3 3 3 Height: 219 pixels
Height: 219 pixels
3 3 Width:
Width:
Height:
Height: 219
219 219
219 pixels
pixels
pixels
pixels
Height:
Height:
Height: 219219 pixels
pixels
219 pixels
3 3 Height:Height: 219 219 pixelspixels
Type of file: PNG file
Type of file: PNG file
Type Height:
Height:
Type of219
of file: 219
file:
PNG pixels
pixels
PNG filefile
Width:
Type 307
Width:
Type pixels
of
of file: 307
file:
PNGpixels
PNG filefile
Width:
Width: 307 307
Type of file: PNG file file
Type of file: pixels
pixels
PNG
4 4 Width:
Width:
Type 307
307file:
of pixels
pixels
PNG file
4 4 Width:
Type
Height: Width:
Type 307
of
of file:
183
Height: pixels
183307
pixels
file:
PNG pixels
PNG
pixels filefile
4 4 Height:
Height: 183 183 pixels
pixels
44 4 Width:
Width:
Width: 307 pixels 307
307 pixels
pixels
Height:
Height: 183183 pixels
pixels
4 4 Height:
PNG Height: 183 183
file
PNG
Height: 183pixels
pixels
file pixels
Width: PNG
310
Width:
Height:
Height: PNG
pixels
310
183 file
183 file
pixels
pixels
pixels
PNG
Width: PNG file
310 file
pixels
5 5 Width:PNG 310
PNG file pixels
file
Width:
Width: 310310 pixels
pixels
5 5 Height:
Width: 207
Width:
Height:
PNG pixels
207
PNG
310 310
file pixels
pixels
file
pixels
5 5 PNG file
5 5 Width:
Width:
Height:
Height: 207310
310 207 pixels
pixels
pixels
pixels
Height:
Height:
Width: 207207 pixels
pixels
310 pixels
5 5 5 Height:Height: 207 207 pixelspixels
The performance
The performance is measuredis measuredby using the figure
by using the figureof merit of Pratt
of merit of (Figure
Pratt (Figure of Merit; FOM)FOM)
of Merit; [36]; [36];
Height:
Height:
Height: 207207 pixels
pixels
207 pixels
wherewhere the FOM the FOM is a useful is a usefulmetricmetricfor assessing
for assessing the qualitythe quality of edge detectors,
of edge whichwhich
detectors, consists of
consists of
The The performance
performance is is measured
measured by by
usingusing the the
figure figure of of merit
merit of of Pratt
Pratt (Figure(Figure of of Merit;
Merit; FOM) FOM) [36];
[36];
evaluatingevaluating
The
The performance the distance
performance the is between
distance is measured
measured two
between by images—the
two
by
using images—the
using the
the figurefirst represents
figure first represents
of
of quality merit
meritofofof the output
of
Prattthe
Pratt edges
output
(Figure(Figure obtained
edges of
of which Merit;
Merit; by
obtainedFOM)the by
FOM) the[36];
[36];
where
where Thethe
proposed The the
FOM
performance
proposed FOM
performance
method methodaisis
is (GT2 ameasured
useful useful
is measured
fuzzy
(GT2 metric
metric
edge
fuzzy by for
edge byfor
using
detector) assessing
assessing
using the the
and
detector) figure
the
and the the
figure
second
the quality
of of
is amerit
merit
second of edge
of
is aPratt
reference edge
Pratt
(Figure
image
reference detectors,
detectors,
(Figure
or of
image of
theMerit; which
Merit;
ground
or the FOM) consists
consists
FOM)
truth
ground [36];
truth of of
[36];
where
where the
evaluating
evaluating the
FOM
the FOM
the is ais
distance
distance a
useful useful
between
between metric
metric two for
two for assessing
assessing
images—the
images—the the
firstthe
first quality
quality of
represents
represents of
edge
the edge
the
output detectors,
detectors,
output edges which
edges which
obtained consists
consists
obtained by by of
the of
the
wherewhere
(inThe
Table
theThe the
FOM2). FOM
The
(inperformance
Table is
performance isuseful
2).output
aThe aisoutput
useful
ismeasure metric
metric
measured
measured varies
measureforby
by for
in
variesassessing
the
assessing
usingusing inrange
the thethe
the
figure the
[0,
range 1],quality
quality
figure of ofwhere
[0,
merit 1],
ofof
merit of
where
edge
ofisedge
1Pratt the detectors,
isoptimal
1 (Figure
Pratt the optimal
detectors,
(Figure ofvalue,
which whichi.e.,
value,
ofobtained
Merit;
Merit; consists
the
consists
FOM) i.e.,[36];
FOM) the of
of[36];
The evaluating
evaluating
performance
proposed the the
method distance
distance
is measured
(GT2 between
between
fuzzy two two
byrespect
edge using images—the
images—the the first
figure first of represents
represents
merit the
of the
output
Pratt output edges
(Figure edges ofobtained
Merit; by by
the
FOM) the [36];
proposed
output
evaluating
evaluating
where method
edges
output
the
the theareis
edges
distance
FOM (GT2
similar
are
distance
ais fuzzy
with
similar
abetweenbetween
useful edge
respect
with
two
metrictwotodetector)
detector)
the
images—the
for to
images—theandand
reference
the
assessing the
reference
firstthe
second
image.
first
the second
image.is aof
represents
represents
quality isreference
athe
reference
the
output image
output image
edges edges orwhich
the
or obtained
the ground
ground
obtained bytruthbytruth
the ofthe
where
(in proposed
proposed
where proposed
the (inFOM
Table
the
Table
FOM
2).ismethod
method
2).
The a The (GT2
useful
output
useful
(GT2
output fuzzy fuzzy
metric
metric
edge
measure
measure edge
for
for
detector)
varies
assessing
detector)
assessing
varies in and
in
the and
the the
range
the
thethe
range
quality
secondsecond
quality
[0, [0,
1], where aof
edge
aisreference
is1], of
where
edge
reference
edge
1 is1 is
the
detectors,
detectors,
image image
detectors,
the orwhich
or value,
the
optimal
optimal the
ground
which
value,
consists
consists
ground truth
i.e., truthof
consists
i.e.,
the the of
proposed
evaluating method method
the (GT2 (GT2
distance fuzzy fuzzyedge
between edge detector)
detector)
two andand
images—the thefirstthe
second second
first is a is a the
reference
reference
represents image image or orobtained
the the
ground ground truth truth
evaluating
(in (in
Table
Fuzzy
output
output Table 2).
Edge
Fuzzy
edges
the
edges2).
The distance
The
Detection
Edge
are output
are
similar output
Method
Detection
similar
between
with measure
measure
Applied
Method
with
two
respectApplied
respect
images—the
onvaries
varies thethe
to inthe
on
to inSynthetic
the
Synthetic
the the
range range
Color
reference
reference
represents
[0, [0,
Images
Color
image. 1],
image. 1],
where
Images where 1the output
output
is1 the
is the edges
optimaledges
optimal obtained
value,value, by
i.e., by
the
i.e.,
the the
the
evaluating
(in (inthe
Table Table
proposed
proposed distance
2). 2).
The The
method
method between
output
(GT2 output
(GT2 measure
fuzzy fuzzy two
measure
edge images—the
varies
edge varies
detector)in the
detector) in
and the
range
andthe first
range
the represents
[0, is
[0,second
second 1], 1],aiswhere
where the
1 is1the
a reference
reference output
isimage
the
optimal
image edges
optimalorvalue,
or the the obtained
value, i.e.,
i.e.,truth
ground
ground the the
truth by the
output
output edges edges areare similarsimilar with with respect
respect to to the
the reference
reference image.image.
proposed(inoutput
outputmethod
(in
Table In the
edges
Table edges
2).Infirst
arethe
(GT2
2).
The are
Thetest,
first
similar
fuzzy
output the
similar
test,
output fuzzy
with with
the
edge edge
fuzzy
respect
measure
measure detection
respect
edge to
tovaries
detector)
varies the inthe methods
detection reference
reference
and
in
the the methods
the
range based
image.
range image.
second
[0, [0,on1],
based
1], T1, onaIT2,
iswhere
where 1and
T1, IT2,
reference
is isGT2
1 the and
the FSs
GT2
image were
FSs
optimal
optimal applied
orweretheapplied
value,
value, ground
i.e., thethe truth
i.e.,
Fuzzy
Fuzzy
on Edge
the onEdge
image
the Detection
Detection
database
image Method
databaseMethod
presented Applied
Applied onin
in Table
presented on 2.the
theTable Synthetic
Synthetic
These images
2.reference
These Color Color
were
images Images
Images converted
were converted to theto Lab,
theHSV,
Lab, HSV,and RGB and RGB
(in Table output
output
2).
Fuzzy
Fuzzy The edges
Edge edges
Edge are
output are
Detection
Detection similar
similar
measure
Method with
Method with respect
respect
varies
Applied
Applied onintotheto
the
onthethethe
reference
range
Synthetic
Synthetic [0,image.
Color 1],
Colorimage.
where
Images 1 is the optimal value, i.e., the output
Images
Fuzzycolor
Fuzzy
Edge
In formats
color
InEdge
the Detection
the
first in
formats
first order
Detection
test, in
Method
test, to
theorder
the analyze
Method
fuzzy Applied
fuzzy edgethe
to analyze
Applied advantage
the
on
ondetection
edge the advantage
the
Synthetic
detection of methods
Synthetic
methodsusing
Color one
of Color
using format
Images
based one
based color
onformat
Images on
T1,T1, over
color
IT2, IT2,andthe
over
and others.
GT2 the
GT2 others.
FSs FSs
were were applied
applied
edges are similar
IntheIn
the Table
InEdge
thewith
first 3test,
Infirst
Tabletherespect
FOM
test,
3the
the the FOM
fuzzy to
values
fuzzythe
values
edge reference
achieved
edge by
detection
achieved
detection the image.
byT1the
methods FSs
methods T1edgeFSs
based for
based
edge the Lab,
on
for T1,
the
onconverted
T1, HSV,
IT2, IT2,
Lab, andandand
HSV, RGB
GT2 GT2
and FSsformats
FSs
RGB were are
were
formats applied
applied are
on Fuzzy
Fuzzy
on
the Edge
image
In image
the Detection
Detectiondatabase
database MethodMethod
presented Applied
Applied
presented in on
in
Table on
the
Table the
2. Synthetic
Synthetic
2.
These These Color
images Color
images Images
Images
were were converted to to
the the
Lab, Lab,HSV, HSV, and andRGB RGB
In the
presented. firstfirst
presented. test,test,the the fuzzy fuzzyedge edge detection
detection methods methods based basedon T1, on T1, IT2,IT2, andand GT2GT2 FSs FSs were were applied
applied
on on
the
color
color the
image image
formats
formats in database
database
in
orderorder to presented
presented
to analyze
analyze in
the in
Table
the Table 2.
advantage
advantage 2.
These These
of images
of
using images
using onewere
onewere
format converted
converted
format colorcolor to
over to
the
over the
theLab,
theLab,HSV, HSV,
others.
others. and andRGB RGB
Fuzzy Edge onIn
on the the
imageIn
Detection
theimage
the
first database
database
first
Method
test, test,the thepresented
presented
fuzzy fuzzy
Applied inedge
edge on in
Table Table
the 2.
detection
detection 2. methods
These
Synthetic These images
methods images
Color were
based were
based
Images on on converted
converted
T1, T1,
IT2, to
IT2,and toGT2
the
and the
Lab,
GT2 Lab,
FSsHSV,
FSs HSV,
were and
were and
RGB RGB
applied
applied
color
color formats
formats
In Table in in
order
3intheorder to
FOM to analyze
analyze
values the the advantage
advantage of of
using using one oneformatformat color colorover overthe the others.
others.
oncolor
coloron Inimage
Inthe
the Table
formats formats
Inimage
Table Table
3 the
in order
3 the
FOM
order
database
database
3 theFOM
to
FOM
values
to
analyze analyze
presented
presented
valuesvalues inachieved
achieved
the the
in Table
Table
achieved
achieved
by
advantage
advantage
by
by
2.
2. These
by
the
the
theThese
of T1 ofT1
FSs
using
T1images
the T1
FSs
FSs
using edge
one
images
FSsedge
edge
one
format
were
edgewere
for
for
thethe
forformat Lab,
color
theconverted
converted
for the
Lab,
Lab,
color
over
Lab, toHSV,
HSV,
HSV,
over
the
HSV,
and
tothethe and
the
Lab,
and
RGB
others.
Lab,
and
RGB
RGB
others.
HSV,
RGB HSV, formats
formats
and and
formats
formats RGBare
are
are
RGB
are
thepresented.
In presented.
color first
Informats
color Table test,
Informats
Table3 inthe
the3order
inthe
FOMfuzzy
order FOM
tovalues edge
to values detection
achieved
achieved
analyze
analyze the the by the
advantage
advantage methods
by the T1
of FSs T1using
of
using based
FSs
edgeoneedge
for
one onfor
the
format T1,
formatthe
Lab, IT2,
Lab,
color HSV,
color and
HSV,
over and
over GT2
theandRGB
the FSs
RGB were
formats
formats
others.
others. applied
are are on
presented.
presented.
the image database
presented.
presented. In Table
In Table 3presented
the3 theFOM FOM invalues
values Table 2. These
achieved
achieved by by images
the the
T1 T1 FSsFSs were
edge edge converted
for for
thethe Lab, Lab, to HSV,
HSV, theand Lab,
andRGB HSV,
RGB andare
formats
formats RGBare color
inpresented.
formatspresented.
order to analyze the advantage of using one format color over the others.
In Table 3 the FOM values achieved by the T1 FSs edge for the Lab, HSV, and RGB formats
are presented.
Information 2017, 8, 104 9 of 15

Table 3. T1 FSs edge detection applied on Lab, HSV, and RGB images.

T1 FSs Edge Detection

Image Number FOM
Lab HSV RGB
1 0.9430 0.9380 0.9393
2 0.9504 0.9372 0.9478
3 0.9485 0.9403 0.9464
4 0.9493 0.9453 0.9491
5 0.9428 0.9353 0.9415
Mean 0.9468 0.9392 0.9448

The results achieved by the IT2 fuzzy edge detection for the Lab, HSV, and RGB color format
images are shown in Table 4.

Table 4. IT2 FSs edge detection applied on Lab, HSV, and RGB images.

IT2 FSs Edge Detection

Image Number FOM
Lab HSV RGB
1 0.9437 0.9401 0.9401
2 0.9510 0.9478 0.9479
3 0.9491 0.9464 0.9470
4 0.9496 0.9491 0.9493
5 0.9430 0.9415 0.9417
Mean 0.9473 0.9450 0.9452

The FOM values achieved by the proposed method based on GT2 FSs are presented in Table 5.

Table 5. GT2 FSs edge detection applied on Lab, HSV, and RGB color format images.

GT2 FSs Edge Detection

Image Number FOM
Lab HSV RGB
1 0.9457 0.9416 0.9416
2 0.9523 0.9481 0.9481
3 0.9497 0.9472 0.9473
4 0.9503 0.9497 0.9497
5 0.9434 0.9419 0.9420
Mean 0.9483 0.9457 0.9457

In Table 6, a summary of Tables 3–5 is presented and we also include the results achieved by the
Canny and Sobel edge detection methods. According to these results, we can conclude that the edge
detection method based on IT2 FSs achieved better FOM values than the T1 FSs for the three different
color formats. Nevertheless, the proposed method based on GT2 FSs improved most of the results in
comparison with the IT2 FSs, T1 FSs, Canny, and Sobel methodologies; additionally, we can notice that
the FOM values obtained by the Sobel operator are the lowest, followed by those of the Canny method.
An interesting finding in Table 6 is that, according to the FOM values, four of the five methods
achieved better results when the Lab color format was used; this means that the Lab format provides
more information about the image, which is a good point to consider in image processing applications,
especially those that work with the edge detection process. In Figure 7, we present a plot of the FOM
values achieved by the three different fuzzy edge detection methods when these are applied over
Information 2017, 8, 104 10 of 15

Information 2017, 8, 104 10 of 15

the Lab, HSV, and RGB formats; we can see the advantage of using the Lab color format in this kind
applied
of over the Lab, HSV, and RGB formats; we can see the advantage of using the Lab color format
application.
in this kind of application.
Table 6. Results achieved by T1FSs, IT2FSs, and GT2FSs edge detectors applied on Lab, HSV, and RGB
Table 6.
format Results
color achieved by T1FSs, IT2FSs, and GT2FSs edge detectors applied on Lab, HSV, and
images.
RGB format color images.
FOM
Fuzzy Edge Detector FOM
Fuzzy Edge Detector Lab HSV
Lab HSV RGBRGB
T1 FSs
T1 FSs 0.9468
0.9468 0.9392
0.9392 0.9448
0.9448
IT2 FSs
IT2 FSs 0.9473
0.9473 0.9450
0.9450 0.9452
0.9452
GT2 FSs 0.9483 0.9457 0.9457
Canny
GT2 FSs 0.9483
0.8113
0.9457
0.9264
0.9457
0.8021
SobelCanny 0.8113
0.5343 0.9264
0.4762 0.8021
0.4396
Sobel 0.5343 0.4762 0.4396

7. FOM
Figure 7. FOMvalues
valuesachieved byby
achieved thethe
fuzzy edgeedge
fuzzy detectors in Lab,
detectors in HSV,
Lab, and
HSV,RGB
andcolor
RGBformat
color images.
format
images.
In order to evaluate the ability of the fuzzy techniques to handle the uncertainty (noise) of the
In of
images order to2,evaluate
Table they werethecorrupted
ability ofusing
the fuzzy techniques
Gaussian to handle
noise levels (20, 30,the
40,uncertainty
and 50 dBi).(noise)
In Tablesof 7–9
the
images
we of Table
present 2, they
the FOM were corrupted
average using by
value achieved Gaussian
the T1,noise levels
IT2, and GT2(20,FSs
30,edge
40, and 50 dBi).after
detection In Tables
these
7–9 we present the FOM average value
were applied on Lab, HSV, and RGB images. achieved by the T1, IT2, and GT2 FSs edge detection after
theseFrom
wereTable
applied on Lab, HSV, and RGB images.
7 (T1 FSs) we see that for images with Gaussian noise of 20 dBi, the FOM value (0.9367)
From Table 7 (T1color
was better in the HSV FSs)format.
we seeForthat30for
andimages
40 dBi with Gaussian
the FOM valuesnoise of 0.9485)
(0.9489, 20 dBi, were
the FOM
bettervalue
with
the Lab color format; for 50 dBi the FOM result (0.9457) was higher using the RGB color format. were
(0.9367) was better in the HSV color format. For 30 and 40 dBi the FOM values (0.9489, 0.9485)
better with the Lab color format; for 50 dBi the FOM result (0.9457) was higher using the RGB color
format. Table 7. FOM values achieved by T1 FSs in images with Gaussian noise.

Table 7. FOM values achieved byT1T1FSs

FSs in images with Gaussian noise.

T1 FSs FOM
Color Format
20 dBi 30 dBi FOM40 dBi 50 dBi
Color Format
Lab 0.849420 dBi 0.9489
30 dBi 400.9485
dBi 50 dBi 0.8972
HSV Lab 0.93670.8494 0.9393 0.9392
0.9489 0.9485 0.8972 0.9391
RGB 0.8844 0.9461 0.9458 0.9457
HSV 0.9367 0.9393 0.9392 0.9391
RGB 0.8844 0.9461 0.9458 0.9457
From Table 8 (IT2 FSs) we observe that for the images with 20 dBi of Gaussian noise, the HSV
colorFrom
format has 8a (IT2
Table better FOM
FSs) wevalue (0.9385);
observe however,
that for for 30,
the images with40,20and
dBi50ofdBi the FOM
Gaussian values
noise, the were
HSV
higher in the has
color format Lab aformat.
better FOM value (0.9385); however, for 30, 40, and 50 dBi the FOM values were
higher in the Lab format.
Information 2017, 8, 104 11 of 15

Table 8. FOM values achieved by IT2 FSs in images with Gaussian noise.

IT2 FSs
FOM
Color Format
20 dBi 30 dBi 40 dBi 50 dBi
Lab 0.8597 0.9492 0.9485 0.9483
HSV 0.9385 0.9394 0.9393 0.9393
RGB 0.9081 0.9466 0.9458 0.9456

In Table 9, the simulation results after applying the GT2 FS edge detection on images with
Gaussian noise are presented. In these results we notice that the FOM value (0.9392) was better when
the HSV color format was used under a noise level of 20 dBi; for 40 and 50 dBi the FOM values (0.9491,
0.9489) were better using the Lab color format; in this test the RGB format has higher FOM for a noise
level of 30 dBi (0.9563).

Table 9. FOM values achieved by GT2 FSs in images with Gaussian noise.

GT2 FSs
FOM
Color Format
20 dBi 30 dBi 40 dBi 50 dBi
Lab 0.8600 0.9496 0.9491 0.9489
HSV 0.9392 0.9395 0.9393 0.9392
RGB 0.9158 0.9563 0.9470 0.9460

Finally, we present a comparative analysis in which the results are classified by color format.
In Tables 10–12 we present the results achieved by the three fuzzy edge detection methods and two
traditional methods (Canny and Sobel) when these are applied on Lab, HSV, and RGB color format
images, respectively. According to the results of Table 10, the FOM values were better when the GT2
FS was applied for the different noise levels (20, 30, 40, and 50 dBi). Moreover, in Table 11 (HSV)
and Table 12 (RGB) we notice that the GT2 FSs also improved the results, obtaining the best FOM
accuracy. Additionally, we notice that the Sobel method achieved the lowest results for all the cases,
followed by Canny; this means that the traditional methods do not have additional parameters to
handle uncertainty, which is the advantage of using fuzzy techniques in this kind of application.
In Figure 8, we present a plot to illustrate the FOM values by color format obtained by the five
edge detection methods when these are applied in images with noise; in this figure we cannot observe
clearly the difference, and for this reason we calculate the mean for each format color. For Lab color
we obtained a FOM mean of 0.8126, for HSV a FOM mean of 0.8385, and for RGB a mean value of
0.8333. Therefore, for these tests we conclude that when the images are corrupted by noise, the HSV
color format represents a good option, followed by the RGB color format.

Table 10. T1, IT2, and GT2 FS edge detection applied on Lab color format images with Gaussian noise.

Lab
FOM
Fuzzy Edge Detection
20 dBi 30 dBi 40 dBi 50 dBi
T1 FSs 0.8494 0.9489 0.9485 0.8972
IT2 FSs 0.8597 0.9492 0.9485 0.9483
GT2 FSs 0.8600 0.9496 0.9491 0.9489
Canny 0.8306 0.8126 0.8113 0.8110
Sobel 0.4824 0.4824 0.4824 0.4824
Information 2017, 8, 104 12 of 15

Table 11. T1, IT2, and GT2 FS edge detection applied on HSV color format images with Gaussian noise.

HSV
FOM
Fuzzy Edge Detection
20 dBi 30 dBi 40 dBi 50 dBi
T1 FSs 0.9367 0.9393 0.9392 0.9391
IT2 FSs 0.9385 0.9394 0.9393 0.9393
GT2 FSs 0.9392 0.9395 0.9393 0.9392
Canny 0.8175 0.9261 0.9264 0.9264
Information 2017, 8, 104 Sobel 0.4762 0.4762 0.4762 0.4762 12 of 15

Table 11. T1,

Table 12. T1,IT2,
IT2,and
andGT2
GT2FSFS edge
edge detection
detection applied
applied on HSV
on RGB colorcolor format
format images
images with Gaussian
with Gaussian noise.
noise.
RGB
HSV
FOM
FOM
Fuzzy
FuzzyEdge Detection
Edge Detection
20 20
dBidBi 30
30dBi
dBi 40
40 dBi
dBi 5050dBi
dBi
T1T1
FSsFSs 0.9367
0.8844 0.9393
0.9461 0.9392
0.9458 0.9391
0.9457
IT2IT2
FSsFSs 0.9081
0.9385 0.9466
0.9394 0.9458
0.9393 0.9456
0.9393
GT2
GT2FSsFSs 0.9158
0.9392 0.9563
0.9395 0.9470
0.9393 0.9460
0.9392
Canny 0.8302 0.8230 0.8220 0.8212
Canny
Sobel
0.8175
0.5343
0.9261
0.5343
0.9264
0.5343
0.9264
0.5343
Sobel 0.4762 0.4762 0.4762 0.4762

Figure 8.
Figure Comparison of
8. Comparison of results
results of
of the
the FOM
FOM values
values for
for the
the different
different color
color formats.
formats.

Table 12. T1,edge

The visual IT2, and GT2 FS edge
detections detection
for T1, applied
IT2, and GT2onFSs
RGB color
are format images
presented with13.
in Table Gaussian
Of course,
noise. we cannot clearly appreciate the difference in the detected edges and for this reason
graphically
the FOM measure was applied, but tests supportRGB
the conclusion that GT2 FSs can outperform IT2 and
T1 in edge detection process. FOM
Fuzzy Edge Detection
20 dBi 30 dBi 40 dBi 50 dBi
T1 FSs 0.8844 0.9461 0.9458 0.9457
IT2 FSs 0.9081 0.9466 0.9458 0.9456
GT2 FSs 0.9158 0.9563 0.9470 0.9460
Canny 0.8302 0.8230 0.8220 0.8212
Sobel 0.5343 0.5343 0.5343 0.5343

The visual edge detections for T1, IT2, and GT2 FSs are presented in Table 13. Of course,
graphically we cannot clearly appreciate the difference in the detected edges and for this reason the
FOM measure was applied, but tests support the conclusion that GT2 FSs can outperform IT2 and T1
in edge detection process.
 Information 2017, 8, 104 13 of 15
Information
Information
Information 2017,
2017,
2017,
Information
Information 2017,
Information
Information 8,
2017,
2017, 8,
8,2017,
104
8,
104
8, 104
8, 104
8, 104
104
104 13 13
13 of
of
13 of
15
of
1315
13 of
of 15
1315
of 15
15
15

Table 13. Edge detection output using T1, IT2, and GT2 FSs.
Table
Table 13.13.
Table
13.
Table Edge
Table
Edge
13. Edge
13. detection
Edge
detection
Edge output
detection
detection using
output
detection
output
output
output T1,T1,
using
output using
using
using T1,
using IT2,
T1,
IT2,
T1, and
IT2,
T1,
IT2,
and
IT2, GT2
and
IT2,
and
GT2
and FSs.
GT2
and
GT2 FSs.
GT2
FSs.
GT2 FSs.FSs.
FSs.
T1
T1 FSs
T1T1
T1
FSs
FSsFSs
T1
FSsFSs IT2IT2
IT2 FSs
IT2
IT2
FSsFSs
IT2
FSsFSs GT2
GT2 FSs
GT2 FSs
GT2
GT2 FSsFSs
GT2
FSs

6. Conclusions
6. Conclusions
6.
6. Conclusions
6. 6. Conclusions
Conclusions
Conclusions
In In
In this
this
In
In this
In
this
this paper
paper
thispaper
paper
paper we
wewewe
paper we presented
we presented
presented
presentedpresented
presented an anedge
an anedge
edge
an edge
anedge detection
edge
detectiondetection
detection detection
detection approach
approach approach
approach approach
approachbased based
basedbased
on ongradient
on
based
based
the on
onthe
the onthe
thegradient
thegradient
gradient
gradient
technique technique
technique
technique
gradient technique
technique
combined
combined
combined
combined
combined
combined
with three with
with with
with
fuzzythree
with
threethree
threefuzzy
three
fuzzyfuzzy
fuzzy
approaches approaches
fuzzy approaches
approaches
approaches
approaches
(T1, IT2, (T1,(T1,
and IT2,
(T1,
(T1,
IT2,
GT2 IT2,
(T1,
IT2,and
and
FSs),and
IT2,
and GT2
and
GT2 GT2
GT2
where FSs),
GT2
FSs), FSs),
FSs),
the where
where
FSs),
wherewhere
main thethe
where
the
aim main
the
main
was main
the
main aim
tomain
aim aim
aim
show was
was was
aimwasto
to
the wasshow
to
to show
show show
advantage thethe
to show
the ofthe
the
advantage
advantage
advantage
advantage
advantage of
of using
of
of
using using
of
using GT2
using
GT2 GT2
GT2 FSs
GT2
FSs FSs
FSsin
in FSs image
in
in
image image
in
image processing
image processing
processingprocessing
processing systems
systems
systems
using GT2 FSs in image processing systems and analyze their performance with respect to the higher systems
systems and
and and
and analyze
andanalyze
analyze analyze
analyze their
their their
their performance
their performance
performance
performance
performance with
with with
withwith
respect
respect
respect to the
to
respect
respect
uncertainty the
to
to theto
thehigher
higher
higher
the
levels. uncertainty
Inuncertainty
uncertainty
higher
higher uncertainty
uncertainty
general, levels.
welevels.
levels. Inconclude
In
levels.
levels.
can general,
In
In general,
general,
In general,
general, we
we
that can
we
can
we can
we
thecanconclude
conclude
conclude
can conclude
conclude
edge that
that
detection the
that
the
thatthat edge
the
edge
the
based edge
the
edge detection
detection
detection
edge
on based
Fsbased
detection
detection
GT2 has based
based on
on
abased
goodon
on on
GT2
GT2GT2
GT2 Fs
GT2
Fs
control hason
Fs
Fs
has Fs aa has
has
has good
aa good
good
the good
imagecontrol
a good control
control
control onon
control
on
gradients theon
on
the theimage
the
in the
image image
image
images gradients
image gradients
gradients
gradients
gradients
with noise in images
in images
in
in images
and in
images
withoutwith
images
with with
with noise
with
noise
noisenoise
noiseand
noise
and
for and
and
the without
andwithout
without
without
different noise
without
noisenoise
noise
color for
noise
for the
for
for
the the
for
formats.thethe
different
different
different color
color
different
different color
color formats.
formats.
formats.
color formats.
formats.
On the other hand, different color format images were analyzed (Lab, HSV, and RGB) and,
OnOn
On On
accordingthe
theOn
Onthe
theother
tothe
other
theother
other hand,
theother
hand,
other
FOM hand,
hand,hand,
hand, different
different
differentdifferent
different
metric, different
based color
color color
oncolor
color format
color
the format
format
formatformat
images
format
simulation images
images
images were
images
were
images were
were
results were analyzed
analyzed
analyzed analyzed
inanalyzed
were (Lab,
(Lab,
analyzed
images (Lab,
(Lab,
without HSV,
(Lab,
HSV,
(Lab, HSV,
HSV, and
HSV,
and
HSV,
noise and
and RGB)
and
RGB)
the and RGB)
RGB)
fuzzy and,
RGB)
and,
RGB) and,
and,
edgeand,
and,
according
according
according
according to
according
to the
to
to
the FOM
the
to
the
FOM FOM
the
FOM metric,
FOM metric,
metric,
metric, based
metric,
based based
based on
based
on the
on
on
the simulation
the
on
the simulation
the simulation
simulation
simulation
detection approaches were better for images in Lab color format, followed by the RGB format; theresults
results
results in
results
results in images
in
in images
imagesin without
images
images without
without noise
without
without noisenoise
noise the
noise
the fuzzy
the
the fuzzy
the
fuzzyfuzzy edge
fuzzy
edge edge
edge
edge
detection
detection
detection approaches
detection
detection
lowest approaches
approaches
performanceapproaches
approaches were
were were
were
was better
were better
better
better
from forfor
for
better
the images
images
for
HSV images
for in Lab
in
images
imagescolor Lab
in Lab
in
informat.
Lab color
color
Labcolor
color format,
color
In format,
format, followed
format,
format,
images followed
followed followed
followed
with byby
by
noise, the
the
by the
by
the
the RGB
RGB RGB
the
RGB format;
RGB
performanceformat;
format; the
the
format;
format; the
thethe
was
lowest
lowest
lowest
lowest
better performance
lowest performance
performance
performance
performance
for the HSVwas was was
was
format from
was
from from
from
and theworst
from
the the
theHSV
HSV HSV
theHSV color
HSV
color
for color
color format.
thecolor format.
format.
Lab format. In In
format.
In
format. images
In
imagesimages
In
images
This with
images
with with
with
representsnoise,
with noise,
noise,
noise, an the
noise,
the performance
the
the performance
the performance
performance
performance
important was
was
and helpful was
waswas
better
better
better
better for
better
for the
for
for
the the
for
theHSV
HSV HSV
the
HSV format
HSV format
formatformat and
format
and and
and worst
and
worst worst
worst for
worst
for the
for
for
the the
for Lab
the
Lab
characteristic to be considered when we are working with this kind of application. Lab
the
Lab format.
Lab format.
format. This
format.
format. This This
Thisrepresents
represents
This
representsrepresents
represents an
an important
an
an important
an important
important
important and
and and
and helpful
andhelpful
helpful
helpful
helpful
characteristic
characteristic
characteristic
characteristic
characteristic
In future to be
to be
to
to
workconsidered
be
to considered
considered
be be
theconsidered
considered when
idea when
is towhen
when wewe
we
when
improve arewe
are
we areworking
are
the working
working
are working
working
fuzzy with
with with
with
system this
this
with kind
this
kind
this
using kind
this of
aof
kind kindapplication.
of
of application.
application.
of application.
application.
metaheuristic to optimize the main
In
In future
In
In future
In
future
future work
future
work work
work the
work
the theidea
the
idea idea
the is
ideaidea
is to
is
is
to improve
to
is
to improve
to
improve improve
improve the
the thefuzzy
the
components of the GT2 FSs, such as the type of MF, the parameters of the MF, and of fuzzy
the
fuzzy fuzzy system
fuzzy system
system using
system
system usingusing
using a
using
a metaheuristic
a
a metaheuristic
a metaheuristic
metaheuristic
metaheuristic to course
to optimize
to
to optimize
to optimize
optimize
optimize thethethe the
thethe
fuzzy
main
main
main
main components
main components
components
components
components of
of the
of
of
the ofGT2
the
the
GT2 GT2
the
GT2 FSs,
GT2
FSs, FSs,
FSs,such
FSs,
such such
such as
such
as the
as
as
the type
the
as
the
type type
thetypeof
type
of MF,
of
of
MF,
rules; in addition, the next goal will be improving the GT2 FSs method to reduce the computing time MF,
of
MF, the
MF,
the parameters
the
the parameters
the
parametersparameters
parameters of
of the
of
of
the the
ofMF,
the
MF, MF,
the
MF,and
MF,
and and
andof
ofandcourse
of
of course
courseof course
course
thethe
thethe
for fuzzy
fuzzy
the
fuzzyfuzzy
the fuzzy
rules;
fuzzy
achieving rules;
rules;
rules; in in
rules;
in
rules;
fasterin addition,
addition,
in
addition,
addition,
in the
addition,
the
addition,
results, whichthenext
next
the the
thenext goal
next
is goal
next nextgoal
goal
required will
goal
will
goal willbe
be
will
in be
will
will improving
improving
be
improving
be
this improving
improving
be
kind improving
of areathe the
thethe GT2
the
GT2
fortheGT2
GT2 FSs
GT2
FSs
GT2 FSs
real-world method
FSs
method
FSs method
method
FSs to to
method
to
method
application. to reduce
reduce
to
reduce
reduce
to the
reduce thethe
thethe
reduce the
computing
computing
computing
computing
computing time
time time
time for
time
for achieving
for
for achieving
for achieving
achieving
achieving faster
fasterfaster
faster results,
faster results,
results,
results, which
results,
whichwhich
which is required
which
is required
is
is required
is required
required in in
in this
in
this in kind
this
this kind
this
kind kind of of
kind
of area
of
area area
of for
area for
area
for real-world
for forreal-world
real-world
real-world
real-world
application.
application.
application.
application.
application.
Information 2017, 8, 104 14 of 15

Author Contributions: Claudia I. Gonzalez proposed the idea of color images and implemented the software,
as well as made the experiments. Patricia Melin proposed the general idea of type 2 edge detection and verified
the correctness of implementation and results. Oscar Castillo verified the correctness of type-2 fuzzy logic
implementation and the writing of the paper.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Torre, V.; Poggio, T.A. On Edge Detection. IEEE Trans. Pattern Anal. Mach. Intell. 1986, 8, 147–163. [CrossRef]
[PubMed]
2. Koschan, A.; Abidi, M. Detection and classification of edges in color images. IEEE Signal Process. Mag. 2005,
22, 64–73. [CrossRef]
3. Bhardwaja, S.; Mittalb, A. A Survey on Various Edge Detector Techniques. Procedia Technol. 2012, 4, 220–226.
[CrossRef]
4. Biswas, R.; Sil, J. An Improved Canny Edge Detection Algorithm Based on Type-2 Fuzzy Sets. Procedia Technol.
2012, 4, 820–824. [CrossRef]
5. Bustince, H.; Barrenechea, E.; Pagola, M.; Fernandez, J. Interval-valued fuzzy sets constructed from matrices:
Application to edge detection. Fuzzy Sets Syst. 2009, 160, 1819–1840. [CrossRef]
6. Melin, P.; Gonzalez, C.I.; Castro, J.R.; Mendoza, O.; Castillo, O. Edge-Detection Method for Image Processing
Based on Generalized Type-2 Fuzzy Logic. IEEE Trans. Fuzzy Syst. 2014, 22, 1515–1525. [CrossRef]
7. Melin, P.; Mendoza, O.; Castillo, O. An improved method for edge detection based on interval type-2 fuzzy
logic. Expert Syst. Appl. 2010, 37, 8527–8535. [CrossRef]
8. Tao, C.; Thompson, W.; Taur, J. A fuzzy if-then approach to edge detection. In Proceedings of the Second IEEE
International Conference on Fuzzy Systems, San Francisco, CA, USA, 28 March–1 April 1993; pp. 1356–1360.
9. Gonzalez, C.I.; Melin, P.; Castro, J.R.; Mendoza, O.; Castillo, O. An improved sobel edge detection method
based on generalized type-2 fuzzy logic. Soft Comput. 2016, 20, 773–784. [CrossRef]
10. Zadeh, L.A. Fuzzy Sets; Academic Press Inc.: Amsterdam, The Netherlands, 1965; Volume 8.
11. Zadeh, L.A. The concept of a linguistic variable and its application to approximate reasoning—I. Inf. Sci.
1975, 8, 199–249. [CrossRef]
12. Schulte, S.; De Witte, V.; Kerre, E.E. A Fuzzy Noise Reduction Method for Color Images. IEEE Trans.
Image Process. 2007, 16, 1425–1436. [CrossRef] [PubMed]
13. Yuksel, M.E.; Basturk, A. Application of Type-2 Fuzzy Logic Filtering to Reduce Noise in Color Images.
IEEE Comput. Intell. Mag. 2012, 7, 25–35. [CrossRef]
14. Teruhisa, S.; Futoki, S.; Hiroshi, K.; Toshiaki, O. Application of an edge detection method to satellite images
for distinguishing sea surface temperature fronts near the Japanese coast. Remote Sens. Environ. 2005, 98,
21–34.
15. El Baf, F.; Bouwmans, T.; Vachon, B. Type-2 Fuzzy Mixture of Gaussians Model: Application to Background
Modeling. Int. Sympos. Vis. Comput. 2008, 2008, 772–781.
16. Guo, Y.; Ji, Y.; Zhang, J.; Gong, S.; Liu, C. Robust Dynamic Background Model with Adaptive Region Based
on T2FS and GMM. In International Conference on Knowledge Science, Engineering and Management; Springer:
Cham, Cambodia, 2015; Volume 9403, pp. 764–770.
17. Zhao, Z.; Bouwmans, T.; Zhang, X.; Fang, Y. A Fuzzy Background Modeling Approach for Motion Detection
in Dynamic Backgrounds. Multimed. Signal Process. Commun. Comput. Inf. Sci. 2012, 346, 177–185.
18. Karnik, N.N.; Mendel, J.M.; Liang, Q. Type-2 fuzzy logic systems. IEEE Trans. Fuzzy Syst. 1999, 7, 643–658.
[CrossRef]
19. Liang, Q.; Mendel, J. Interval type-2 fuzzy logic systems: Theory and design. IEEE Trans. Fuzzy Syst. 2000, 8,
535–550. [CrossRef]
20. Mendel, J. Uncertain Rule-Based Fuzzy Logic Systems: Introduction and New Directions; Prentice-Hall:
Upper-Saddle, NJ, USA, 2001.
21. Mendel, J.M. Advances in type-2 fuzzy sets and systems. Inf. Sci. 2007, 177, 84–110. [CrossRef]
22. Hao, M.; Mendel, J.M. Similarity measures for general type-2 fuzzy sets based on the α-plane representation.
Inf. Sci. 2014, 277, 197–215. [CrossRef]
Information 2017, 8, 104 15 of 15

23. Liu, F. An efficient centroid type-reduction strategy for general type-2 fuzzy logic system. Inf. Sci. 2008, 178,
2224–2236. [CrossRef]
24. Mendel, J.M. General Type-2 Fuzzy Logic Systems Made Simple: A Tutorial. IEEE Trans. Fuzzy Syst. 2014, 22,
1162–1182. [CrossRef]
25. Wagner, C.; Hagras, H. Toward general type-2 fuzzy logic systems based on zSlices. IEEE Trans. Fuzzy Syst.
2010, 18, 637–660. [CrossRef]
26. Wagner, C.; Hagras, H. Employing zSlices based general type-2 fuzzy sets to model multi level agreement.
In Proceedings of the IEEE Symposium on Advances in Type-2 Fuzzy Logic Systems (T2FUZZ), Paris, France,
11–15 April 2011; pp. 50–57.
27. Mendel, J.M.; Liu, F.; Zhai, D. α-Plane representation for type-2 fuzzy sets: Theory and applications.
IEEE Trans. Fuzzy Syst. 2009, 17, 1189–1207. [CrossRef]
28. Mendel, J.M. Comments on alpha-plane representation for type-2 fuzzy sets: Theory and applications.
IEEE Trans. Fuzzy Syst. 2010, 18, 229–230. [CrossRef]
29. Kirsch, R. Computer determination of the constituent structure of biological images. Comput. Biomed. Res.
1971, 4, 315–328. [CrossRef]
30. Sobel, I. Camera Models and Perception. Ph.D. Thesis, Stanford University, Stanford, CA, USA, 1970.
31. Prewitt, J.M.S. Object Enhancement and Extraction; Lipkin, B.S., Rosenfeld, A., Eds.; Picture Analysis and
Psychopictorics; Academic Press: New York, NY, USA, 1970; pp. 75–149.
32. Canny, J. A computational approach to edge detection. IEEE Trans. Pattern Anal. Mach. Intell. 1986, 8,
679–698. [CrossRef] [PubMed]
33. Gonzalez, R.C.; Woods, R.E.; Eddins, S.L. Digital Image Processing Using Matlab; Prentice Hall: Upper-Saddle,
NJ, USA, 2004.
34. Kulkarni, A.D. Computer Vision and Fuzzy Neural Systems; Prentice Hall: Upper-Saddle, NJ, USA, 2001.
35. Pratt, W.K. Digital Image Processing, 2nd ed.; John Wiley & Sons, Inc.: New York, NY, USA, 1991.
36. Abdou, I.A.; Pratt, W. Quantitative design and evaluation of enhancement/thresholding edge detectors.
Proc. IEEE 1979, 67, 753–766. [CrossRef]

© 2017 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).