This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Unc
orre
cted
Aut
hor P
roof
Journal of Intelligent & Fuzzy Systems xx (20xx) x–xxDOI:10.3233/IFS-131110IOS Press
1
Multichannel generalization of theUpper-Lower Edge Detector using orderedweighted averaging operators
1
2
3
C. Guerraa,∗, A. Jurioa, H. Bustincea and C. Lopez-Molinaa,b4
aDepartment Automatica y Computacion, Universidad Publica de Navarra, Spain5
bDepartment of Mathematical Modelling, Statistics and Bioinformatics, Ghent University, Gent, Belgium6
Abstract. A large number of methods in the edge detection literature are only prepared to deal with monochannel images, whichrepresent the value at each pixel by means of scalar values. This fact hinders their applicability to many fields in which multichannelare common, including remote sensing or medical imagery. Very often, multichannel images have to be turned into grayscaleimages on which edge detection can be performed, but this is coupled to a loss of information that can be unbearable in certainscenarios. In this work we propose a technique for multichannel edge feature fusion technique that can be combined with any edgedetection method using scalar edge features. In this way, we can extend edge detection methods by considering an initial phaseof monochannel feature extraction followed by a subsequent phase of multichannel feature fusion. For the information fusion wemake use of Ordered Weighted Averaging (OWA) operators, which are able to vary the relevance of each of the features to beaggregated depending upon their value. As an example, our proposal is tested with the Upper-Lower Edge Detector, despite it canbe further combined with a wide range of edge detectors.
2 C. Guerra et al. / Multichannel generalization of the ULED using OWA operators
detection is very often based on the characterization of41
gradients in the image. The concept of gradient as the42
combination of orthogonal partial differences at each43
pixel of the image is of vast utility on monochannel44
images, but has a complicated fit when pixel infor-45
mation becomes vectorial. The underlying problem is,46
as stated by Toivanen et al., that it is not possible to47
define uniquely the ordering of multivariate data [39].48
Since the most employed edge detection methods are49
grounded in signal processing and convolution opera-50
tors (and make use of gradients), we can safely state that51
a large part of the edge detection methods in the litera-52
ture cannot be applied to color images. Very often, this53
impels the images to be collapsed to a monochannel rep-54
resentation before detecting edges. This comes coupled55
to a loss of information, which often hinders the success56
of edge detection methods. Even if non-linear aggrega-57
tion of channels is used, the dimensionality reduction58
might lead to indescernibility of vectorial tones in their59
scalar representation.60
There has been a significant effort in developing61
color-specific operators for gradient characterization.62
Examples of such efforts can be found in the works by63
Karakos and Trahanias [21] or Evans and Liu [12]. As64
stated by Zhu et al. [45], the two options to generate65
such operators are (a) the extension of monochrome66
techniques [33–35] and (b) the generation of dedicated67
color operators, often based on the analysis of vectorial68
spaces [11, 12, 38, 40]. Although it is not common,69
some authors rely on the applicability of segmenta-70
tion algorithms as starting point for edge detection.71
This is the case of the proposal by Huntsberger and72
Descalzi [19], which roughly consists of postprocessing73
the results obtained using Fuzzy c-Means [6]. How-74
ever, we believe that channelwise feature extraction,75
followed by a feature fusion phase, is a feasible option,76
mainly because it enables the reuse of the knowledge77
on edge detection gathered for grayscale images.78
In this work we propose a technique for multichan-79
nel gradient fusion. In this way, edge features can be80
characterized at each of the channels independently,81
then fused to produce a single edge interpretation. The82
fusion is performed using Ordered Weighted Aver-83
aging (OWA) operators, which have been throughly84
studied in the past years, and vary the relevance of85
each of the channels depending upon the value of the86
features.87
The remainder of this paper is as follows. In Section88
2 we introduce some preliminary concepts with util-89
ity in the subsequent sections. Section 3 is devoted to90
explain in detail our proposal, while Section 4 covers91
an experimental study of its utility. Our conclusions are 92
exposed in Section 5. 93
2. Preliminary definitions 94
This section recalls some basic definitions of the 95
concepts used hereafter. 96
2.1. Triangular norms and conorms 97
Definition 1. [4, 22] A t-norm T : [0, 1]2 → [0, 1] is 98
an associative, commutative, increasing function, such 99
that, T (1, x) = x for all x ∈ [0, 1]. A t-norm T is called 100
idempotent if T (x, x) = x for all x ∈ [0, 1]. 101
Definition 2. [4, 22] A t-conorm S : [0, 1]2 → [0, 1] is 102
an associative, commutative, increasing function such 103
that S(0, x) = x for all x ∈ [0, 1]. A t-conorm S is called 104
idempotent if S(x, x) = x for all x ∈ [0, 1]. 105
In-depth studies of t-norms and t-conorms together 106
with some other aggregation functions, can be found in 107
[1, 4, 22]. The application of these operators to image 108
processing has been tackled by several authors [7, 24]. 109
2.2. Ordered weighted averaging aggregation 110
operators 111
Definition 3. [44] A function w : [0, 1]n → [0, 1] iscalled an OWA operator of dimension n if thereexists a weighing vector h = (h1, . . . , hn) ∈ [0, 1]n
with∑
i hi = 1, and such that
w(a1, . . . , an) =n∑
j=1
hja(j) (1)
with a(j) the j-th greatest element of the ai, for any 112
(a1, . . . , an) ∈ [0, 1]n. 113
Any OWA operator is completely defined by its 114
weighing vector. In his original definition, Yager con- 115
sidered functions w defined on the whole Euclidean 116
space Rn and taking values in R, but for our interest 117
it is more appropriate to reduce this to [0, 1]n. In this 118
work we use the 3-place OWA operators in Table 1. 119
In the remainder of this work we consider images 120
to have dimensions of X and Y pixels. For the sake of 121
brevity, we consider P = {1, . . . , X} × {1, . . . , Y} to 122
be the set of their positions. We denote by IQ the set of 123
all images whose pixels take value in Q (i.e. I ∈ IQ if 124
and only if for all (x, y) ∈ P , I(x, y) ∈ Q). When the 125
dimensionality of the universe Q is greater than 1, an 126
Unc
orre
cted
Aut
hor P
roof
C. Guerra et al. / Multichannel generalization of the ULED using OWA operators 3
Table 1OWA operators used in this work
Name Weighing vector
w1 (1, 0, 0)
w2 ( 23 , 1
3 , 0)
w3 ( 12 , 1
2 , 0)
w4 ( 36 , 2
6 , 16 )
w5 ( 13 , 1
3 , 13 )
image I ∈ IQ is said to be multichannel. In those cases,127
I(i) refers to the i-th channel of the image I.128
3. Multichannel Upper-Lower Edge Detector129
3.1. Multichannel edge detection130
Multichannel images have become mainstream for131
industrial applications, and edge detection methods132
should be adapted to such reality. Since grayscale edge133
detection has been deeply studied, we consider that134
grayscale edge detection methods are a good starting135
point to tackled color edge detection. We identify three136
strategies to tackle multichannel edge detection using137
grayscale edge detectors:138
a) Channel fusion. This method consists of converting139
multichannel information to a monochannel repre-140
sentation, so that classical edge detection methods141
can be applied.142
b) Edge combination. In this case, edges are detected143
at each of the channels in completely independent144
processes, leading to as many edge images are chan-145
nels. Once all the edge images are produced, their146
result is combined in a sort of multiexpert voting147
phase.148
c) Feature fusion. This option is based on gathering149
edge features or indicators at each of the channels150
of the image, and subsequently fusing them in a151
unique representation that can be used to take a final152
decision on the presence of edges.153
The first option has been explored in the literature,154
together with the use of color spaces able to appro-155
priately represent the perceived differences between156
tones [42]. Indeed, very early studies on this option157
are dedicated to edge detection, such as that by Robin-158
son [32]. The main advantage of this strategy is that it159
enables the use of most of the existing edge detection160
methods, and that the computational overhead is lim-161
ited to the initial fusion of the channels [5, 28]. Indeed,162
Wesolkowski et al. state that perhaps it is more impor- 163
tant to find the appropriate color space representation 164
rather than design a new edge detection algorithm to 165
obtain superior performance for a color edge detection 166
algorithm [42]. However, it is a fact that in any form 167
of information fusion some data must be lost. By using 168
channel fusion some information about the tones must 169
be lost, hindering the good discrimination of edges in 170
the subsequent process of edge detection. Moreover, if 171
no a priori information about the images is available, 172
the safest fusion technique is the averaging of the chan- 173
nels, which can lead to very misleading results. For 174
example, the values (1, 0, 0) and (0, 0, 1) in the RGB 175
space become indistinguishable in their monochannel 176
(scalar) representation. 177
The second option allows different edge detection 178
methods to adapt to the the specific conditions of each 179
channel. However, it induces practical complications 180
due to the fact that the final result of an edge detection 181
method (binary images with thin edges) dramatically 182
burdens the fusion. That is, the information provided 183
to the ultimate edge combination phase is very limited. 184
More specifically, a lot of information is lost in thinning 185
and binarizing the edges, such the confidence of the 186
edge detection method about the existence of an edge at 187
each pixel (edginess). Consequently, the fusion reduces 188
to some sort of multi-expert binary voting. Note also 189
that slightly displaced edges at different channels must 190
be understood to be the same edge, although represented 191
at different positions of the image, what imposes the use 192
of correspondence methods to match the edge pixels at 193
different images. If ignoring the thinness constraints, 194
as in [13], edge combination becomes easier, but the 195
result is not according to the usual conventions. For 196
these reasons, we have discarded this strategy, despite 197
it allows for the use of the existing edge detectors. 198
In the third option the preservation of the tonal 199
information is extended to the maximum. Moreover, 200
the fusion is performed at the point were most of 201
the information (i.e. the image and the edge features) 202
is available, so that the fusion can be supported by 203
as much knowledge as possible. The main drawback 204
associated with this option is that the methods used 205
for edge detection have to be explicitly modified. In 206
our case, the modification consists of adding an extra 207
phase of feature fusion after extracting edge features at 208
each of the channels. Our proposal is based on OWA 209
operators, and consequently is only valid for edge detec- 210
tion methods producing scalar features. In order to 211
illustrate the proposal, we combine it with the Upper- 212
Lower Edge Detector (ULED). Section 3.2 depicts the 213
Unc
orre
cted
Aut
hor P
roof
4 C. Guerra et al. / Multichannel generalization of the ULED using OWA operators
ULED, while Section 3.3 explains the details of the214
feature fusion phase. The resulting method takes up215
Section 3.4.216
3.2. The Upper-Lower Edge Detector217
As proposed by Bustince et al. [8], fuzzy edge images218
can be constructed from the lengths of intervals of219
Interval-Valued (IV) images. An exhaustive study of220
this fact was performed in [3], where the authors posed221
three different objectives:222
– Defining two new concepts, denoted as lower223
constructor and upper constructor, to construct224
a new interval-valued image (I ∈ IL([0,1]), where225
L([0, 1]) represents the set of all possible closed226
intervals in [0, 1]) from a given grayscale image227
(I ∈ I[0,1]);228
– Generating edge images from IV images;229
– Applying these theoretical developments to real230
images.231
The upper and lower constructors are operators cre-232
ated to generate versions of an image that are brighter233
and darker that the original, respectively. In an early234
effort, they were designed to create IV representations235
of an image, since the results provided by them can be236
used as the bounds of the interval to be assigned with237
each position in the image. The IV image generated238
with these constructors has several properties, among239
which the authors highlight the fact that pixels around240
large intensity variations are assigned intervals whose241
length is greater than that of those pixels in homoge-242
neous regions. In this way, the length of the interval243
associated with each position of the image can be taken244
as an indicator of the edginess of the position, more245
precisely of its membership degree to the edges. Con-246
sequently, a fuzzy edge image can be constructed with247
the lengths of the intervals at each position.248
Definition 4. Let I ∈ I[0,1] be a grayscale image ofdimensions X and Y . Consider two t-norms T1 andT2 and two values n, m ∈ N so that n ≤ X−1
2 andm ≤ Y−1
2 . A lower constructor associated with T1, T2,n and m is given by:
Ln,mT1,T2
: I[0,1] → I[0,1] given by
Ln,mT1,T2
[I](i, j) =mn
T1u=−nv=−m
(T2(I(i − u, j − v), I(i, j))
)
for all (i, j) ∈ P . The values of n and m indicate that the 249
considered window is a matrix of dimension (2n + 1) × 250
(2m + 1) centered at (i, j). For the sake of simplicity, 251
if n = m then we denote Ln,mT1,T2
as LnT1,T2
. 252
Definition 5. Let I ∈ I[0,1]. Consider two t-conorms S1and S2 and two values n, m ∈ N such that n ≤ X−1
2 andm ≤ Y−1
2 . The upper constructor associated with S1,S2, n and m is defined as:
Un,mS1,S2
: I[0,1] → I[0,1], given by
Un,mS1,S2
[I](i, j) =mn
S1u=−nv=−m
(S2(I(i − u, j − v), I(i, j))
)
for all (i, j) ∈ P . The values of n and m indicate 253
that the considered window is a matrix of dimension 254
(2n + 1) × (2m + 1) centered at (i, j). For the sake of 255
clarity, if n = m then we denote Un,mS1,S2
as UnS1,S2
. 256
Let I ∈ I[0,1] and consider a lower constructor Ln,mT1,T2
and an upper constructor Un,mS1,S2
. Then
Ln,mT1,T2
[I](i, j) ≤ I(i, j) ≤ Un,mS1,S2
[I](i, j).
for all (i, j) ∈ P . This implies that the images produced 257
with upper and lower constructors can be used as bound- 258
aries for the creation of IV images. 259
Remark. The definition of lower constructor and upper 260
constructor should not be confused with the fuzzy mor- 261
phological operations of dilation and erosion [10], nor 262
with erosion and dilation defined in classical mathemat- 263
ical morphology [16]. 264
Let I ∈ I[0,1] and consider a lower constructor Ln,mT1,T2
and an upper constructor Un,mS1,S2
. Then, In,m ∈ IL([0,1]),defined as
In,m(i, j) = [Ln,mT1,T2
(i, j), Un,mS1,S2
(i, j)] (2)
generates an interval-valued version of the image, that 265
is, an image for which the value of each pixel is in 266
L([0, 1]) [20]. 267
After building the interval valued image In,m fromI Barrenechea et al. propose to construct a fuzzy edgeimage F [In,m] ∈ I[0,1] so that:
F [In,m](i, j) = Un,mS1,S2
[I](i, j) − Ln,mT1,T2
[I](i, j) (3)
for all (i, j) ∈ P . 268
When using lower and upper constructors, the length 269
of the interval associated with a position represents the 270
intensity variation in its neighborhood. Then, in the con- 271
struction of the fuzzy edge image, the length of the 272
Unc
orre
cted
Aut
hor P
roof
C. Guerra et al. / Multichannel generalization of the ULED using OWA operators 5
Algorithm 1. Procedure for the Upper-Lower Edge Detector(ULED).
interval represents the membership degree of each ele-273
ment to the edges. Note that the concept of membership274
degree to the edges is computationally equivalent to that275
of edginess. Besides, from the definitions of upper and276
lower constructors we have that, if there exists at least277
one white pixel and at least one black pixel in the neigh-278
bourhood, the length associated to a pixel is maximal.279
Consequently, that pixel is always considered an edge.280
The main advantage of the ULED is that, depend-281
ing on the lower and upper constructors we use, each282
pixel is associated with a different membership degree283
to the fuzzy edge images (corresponding to its inter-284
val length). This fact enables us to better adjust to the285
application in which we want to use the edge detec-286
tion method. An extension of the ULED was presented287
in [27], namely Directional ULED (DULED). This288
extension endow the use of several non-squared win-289
dows representing specific orientations in the image,290
and is used to generate vectorial representations of the291
edge features. However, in order to narrow down the292
scope of the experiment, in the remainder of this work293
we only consider the original upper and lower con-294
structors based on the t-norm minimum (TM) and the295
t-conorm maximum (SM). The reason is that this cou-296
ple of operators is the only one that guarantees that, if297
the window centered at each (i, j) has a constant inten-298
sity, then the length of the associated interval is zero.299
Therefore, a pixel in a flat (plain tone) region is never300
considered as part of an edge. The procedure of the301
ULED is included in Algorithm 1.302
3.3. Multichannel feature fusion using OWA303
operators304
When an image is composed of several channels305
(either representing color or not), all of them are meant306
to be meaningful, although their interest or reliability307
Fig. 1. Original image 124084 extracted from the BSDS300 test set.
might be heterogeneous. In the specific case of edge 308
detection, this means that edges (or, in a more general 309
way, local intensity variations) might appear at any pos- 310
sible channel. Hence, the information gathered at each 311
of the channels must be considered in the information 312
fusion phase. 313
As cited by Evans and Liu [12], Novak and Shafer 314
estimate that 10% of the edges are only visible in color 315
images [31]. Moreover, in some images the visibility of 316
the edges is reduced in the color to grayscale conver- 317
sion. An example of this fact can be seen in Figs. 1 and 318
2. In Fig. 1 we can observe a color image extracted from 319
the BSDS300 dataset, while in Fig. 2 we can observe the 320
ULED edge features than can be extracted from each 321
of its channels and from their average. In this figure it 322
is evident that the combination of the color channels 323
in a single grayscale image can lead to the disappear- 324
ance of some edges. However, the treatment of each 325
of the channels individually offers a better possibility 326
to appropriately detect the most relevant edges in the 327
image. 328
When fusing edge features at different channels one 329
must consider that, for each pixel, the presence of an 330
strong indicator at a single channel might be an indica- 331
tor of a perceivable edge, regardless of the absence of 332
such hints at other channels. That is, the presence of a 333
strong intensity variation in one of the channels might 334
be enough to consider the pixel as an edge. In fact, one of 335
the motivations to use multichannel images is that each 336
of the channels might capture information that is obliv- 337
ious to the others. Hence, when fusing edge features, 338
the greater values (presumably indicating the presence 339
of an edge) should be dominant over the smaller ones. 340
There exists a variety of options for feature fusion in 341
the literature. In this work we use aggregation functions, 342
which have been extensively studied [22]. Aggregation 343
Unc
orre
cted
Aut
hor P
roof
6 C. Guerra et al. / Multichannel generalization of the ULED using OWA operators
Fig. 2. Channel decomposition of the image in Fig. 1, together with the result of averaging the RGB channels. In the lower row we include theedge features extracted by ULED using n = m = 3 and (T1, T2, S1, S2) = (TM, TM, SM, SM).
functions, according to the most popular definitions [4],344
produce a scalar representation of the values in a vector345
in a monotonic way. Our proposal consists of produc-346
ing an aggregated estimation of the edge at each pixel347
from the value of the edge features obtained at each348
of the channels. Consequently, apart from the mono-349
tonicity, we demand several other characteristics to our350
fusion operators. First, we need our operators to satisfy351
symmetry, since a reordering of the channels should352
not result in a variation of the fused result. Second,353
we need the operators to be compensatory, since the354
resulting estimation should not be above the maximum355
edge feature of below the minimum. Third, we need356
our aggregation operators to give more influence to the357
channels producing the greatest values at each pixel,358
since the existence of an strong edge cue in one single359
channel might indicate the presence of an edge, despite360
it is not appreciable at some other channels.361
Among the families of compensatory, symmetric362
aggregation functions, the one that better fits our pour-363
poses is that of the OWA operators. If using a decreasing364
weighing vector, the greater values to be aggregated365
will always be assigned more relevance, regardless of366
their position in the vector. Hence, our strategy for367
multichannel edge feature fusion will be to aggregate368
the vector of features using an OWA operator with a369
decreasing weighing vector.370
3.4. Multichannel extension of the ULED371
Our proposal to extend the ULED for multiscale372
edge images is as depicted in Algorithm 2. We refer to373
Algorithm 2. Procedure for the Multichannel Upper-Lower EdgeDetector (MULED).
this edge detector as Multichannel ULED (MULED). 374
Note that the method can handle any number of chan- 375
nels, either representing color or any other multispectral 376
information. 377
4. Experimental validation 378
4.1. Aim of the experiment 379
The aim of this experiment is to check the valid- 380
ity of our proposal for edge detection on color images. 381
More specifically, we want to check whether our feature 382
fusion technique results in a significant improvement 383
of the performance compared to the most usual proce- 384
dure to deal with multichannel images, which is fusing 385
Unc
orre
cted
Aut
hor P
roof
C. Guerra et al. / Multichannel generalization of the ULED using OWA operators 7
ULED
ULED
ULED
������
Original image
Smooth image
Binary edges
Fuzzy edges
Binarization
GaussianSmoothing
Cn
C2
C1
Fig. 3. Schematic representation of the procedure used to obtain binary edge images using ULED. The image has been obtained from the BSDS300test set, and each of the channels is obtained from its RGB decomposition.
(averaging) the channels in the image prior to detect386
edges. Note that we do not intend neither to carry out the387
comparison with methods other that the ULED, nor to388
prove that MULED is better than other existing methods389
in the literature. Instead, we try to find out whether mul-390
tichannel feature fusion leads to results better than those391
of fusing the image. Consequently we only compare two392
alternatives for a unique method (ULED). Independent393
studies should be performed if using the presented mul-394
tichannel fusion technique with detectors other than the395
ULED.396
4.2. Experimental dataset397
In this experiment, we have used the Berkeley398
Segmentation Dataset (BSDS). This dataset offers a399
wide variety of natural images, together with several400
hand-made segmentations of each of them. Those seg-401
mentations can be considered as ideal solutions of the402
edge detection problem. It is our intention to test dif-403
ferent edge detectors and see how close their results are404
to the ideal ones. The images in the BSDS have res-405
olution 321 × 481 or 481 × 321, and are provided in406
grayscale. Each of them is associated to a set of 5 to 9407
binary human-made segmentations.408
4.3. Generating binary images409
The procedure in Algorithm 2 produces a fuzzy rep-410
resentation of the edges. However, an edge detector is411
meant to produce a binary representation of edges, pro- 412
vided in the shape of thin lines. Hence, there is a need 413
for binarizing the edges after the generation of the fuzzy 414
representation. In our case this is achieved by using 415
the non-maxima suppression method by Smith [36], in 416
combination with the hysteresis threshold determina- 417
tion technique by Medina-Carnicer et al. [30]. Since 418
the images used for this experiment are natural, they 419
might contain some kind of noise or contamination. 420
Consequently, we need to perform some regularization 421
prior to edge feature extraction, in order to minimize 422
its impact in the final edges. In this case, the regu- 423
larization is carried out using Gaussian filters, which 424
standard deviation is referred to as σr. The schematic 425
representation of our proposal is shown in Fig. 3. 426
4.4. Measuring the performance of an edge 427
detector 428
There exists an open debate on the best way to evalu- 429
ate the performance of an edge detector, which is being 430
boosted by recent works [17, 18, 25]. In this work we use 431
the methodology by Martin et al. [29], which is based on 432
the use of classification measures. This metholodogy is 433
grounded in the fact that edge detection can be seen as 434
a binary classification problem. Hence, it can be eval- 435
uated in terms of success and fallout, comparing the 436
output of an edge detection method with that generated 437
by a human, which we can consider as ground truth. In 438
this way, we build a confusion matrix such as the one 439
Unc
orre
cted
Aut
hor P
roof
8 C. Guerra et al. / Multichannel generalization of the ULED using OWA operators
Edge Non-Edge
Reality
Edge
Non-Edge
Cla
ssifi
catio
n
TP
FN TN
FP
Fig. 4. Confusion matrix for the edge detection problem.
in Fig. 4, with the elements in the main diagonal being440
the ones correctly classified.441
There are some considerations to be taken into442
account, due to the very particular conditions of the443
edge detection problem. More specifically, to the fact444
that edges are not simply a subset of positions of the445
image, since they contain embedded spatial informa-446
tion. For example, it is clear that an edge displaced447
from its true position should not be penalized as much448
as when it would be completely missing. In such a situ-449
ation, a displaced edge should not be penalized as much450
as a false positive and a false negative. In order to solve451
this problem, we use a one-to-one pixel matching algo-452
rithm to map the edge pixels in the candidate edge image453
(generated by an edge detection method) and the ground454
truth. This matching allows for a certain spatial toler-455
ance (in our case, as much as 1% of the diagonal of456
the image), so that an edge pixel can be slightly moved457
from its true position, yet being considered as correctly458
classified. In order to do the pixel-to-pixel matching459
we use the Cost Scaling Algorithm by Goldberg and460
Kennedy [15].461
From the confusion matrix we extract the precisionand recall evaluations, defined as
Prec = TP
TP + FPand Rec = TP
TP + FN. (4)
Precision and recall measures are preferred for mea-suring the performance over other alternatives typicallyused in ROC analysis [41]. The precision and recallmeasures hold good stability properties when the sizeof the image varies [2]. Moreover, they avoid consider-ing TN, which is much larger than the other elementsin a typical edge detection confusion matrix, and hencedistorts the results. Although Prec and Rec measuresillustrate specific aspects of the problem. some scalarevaluation is needed to assess the overall quality of anedge image. We use the F-measure [14], defined as
Fα = Prec · Rec
α Prec +(1 − α) Rec, (5)
where α is a value modulating the relative impact of 462
the Prec and Rec values; We adhere to the commonly 463
used F0.5. Note that F0.5 is the harmonic mean of Prec 464
and Rec. In this way, we evaluate three different facets 465
of the problem: the accuracy (using Prec), the fallout 466
(using Rec) and the overall quality (using F0.5). 467
4.5. Results 468
The results gathered in the experiment are included in 469
Fig. 5, divided by the size of the square neighbourhood 470
(n) considered in the application of the ULED. For each 471
size we consider three values of σr, and for each of them 472
we display the average performance (in terms of F0.5, 473
Prec and Rec), the average ranking (Rank.) and the 474
number of best and worst results (B/W) obtained. Note 475
that, in order to compute the average ranking and the 476
number of best and worst results, we only consider the 477
candidates using the same n and σr. 478
In Fig. 5 we observe that the results obtained by 479
MULED are almost always better than those by the orig- 480
inal ULED. This can be seen in terms of F0.5, as well as 481
in Rank. and B/W. Indeed, the situations in which the 482
ULED outperforms the MULED are restricted to con- 483
figurations with low values of σr. When σr increases, 484
the MULED stands as a better option. 485
The low performance of the MULED when σr = 1.0 486
is explainable from the information fusion technique. 487
When such a regularization setting is used, an important 488
number of imperfections (e.g. speckle) or fine textures 489
are not removed from the image. If using the MULED 490
the presence of these artifacts forces the detection of 491
significant edge hints in, at least, one of the channels. 492
Consequently the MULED can finally classify them as 493
edges (specially if using w1 or w2), generating a large 494
number of false positives. This can be partially solved 495
by mixing the color channels prior to edge detection, as 496
happens in the ULED. Considering this we can assert 497
that the MULED is more sensitive to false detections 498
than ULED, at least regarding the presence of spuri- 499
ous artifacts in non-heavily regularized images. This 500
problem is specially acute when the weighing vector 501
used for feature fusion emphasizes the importance of 502
the greatest feature value. 503
When greater values of σ are used, the noise and 504
imperfections of the image tend to disappear, although 505
this comes coupled to a progressive regularization of 506
the edges. This trade-off has been extensively studied 507
in the literature, giving rise to the Gaussian Scale- 508
Space [23, 26, 43]. When σr ∈ {2.0, 3.0}, the MULED 509
produces better results than the ULED, since it avoids 510
Unc
orre
cted
Aut
hor P
roof
C. Guerra et al. / Multichannel generalization of the ULED using OWA operators 9
Fig. 5. Results gathered in the comparison of the ULED and the MULED. For each possible neighbourhood size and σr we list the averageperformance (F0.5), precision (Prec), recall (Rec), ranking (Rank.) and the number of images of the BSDS for which it is the best (B) andworst (W) performer.
the aforementioned problem of the false positives while511
successfully facing the detection of blurred edges. In512
these situations, the optimism of the MULED with513
respect to the presence of edges render in very good514
performance. Indeed, it can be observed that the perfor-515
mance of the ULED dramatically decreases from σr =516
1.0 to σr = 2.0, while that of the MULED increases517
Note that this can be observed in terms of F0.5, but is 518
also noteworthy in terms of Rank. or B/W. 519
In this experiment the best possible results are 520
reached in the combination of MULED with great val- 521
ues of σr. There is no significant difference in the setting 522
of the neighbourhood size, so n ∈ {3, 5} stand as a 523
safe option. With respect to w, although there are no 524
Unc
orre
cted
Aut
hor P
roof
10 C. Guerra et al. / Multichannel generalization of the ULED using OWA operators
dramatic differences in the average performance (F0.5),525
both Rank. and B/W have to be carefully analyzed in526
order to fix the settings for specific applications. From527
the results in Fig. 5 we recommend w2 and, if using528
a very high value of σr, also w1. Note that w1 obtains529
the best average results, but at the same time performs530
worse than any other OWA operator for a large number531
of images, as we can observe in the results regarding532
B/W. Hence, if a more conservative option is preferred,533
w3 or w4 can arise as a better choice.534
5. Conclusions535
We have proposed a methodology to adapt existing536
edge detectors to multichannel images. Our proposal537
consists of performing monochannel feature extraction538
followed by a novel technique of multichannel feature539
fusion based on Ordered Weighted Averaging (OWA)540
operators. So far our proposal can only be combined541
with edge detections generating scalar features for edge542
characterization. We have combined our proposal with543
the Upper-Lower Edge Detector (ULED), giving rise544
to Multichannel ULED (MULED). Our experimental545
tests with color images have illustrated how applying an546
explicit multichannel approach can lead to results better547
than averaging color aggregation. Our proposal is very548
flexible, but does not involve complicated paradigms,549
and is well grounded in the extensively studied field of550
aggregation operators.551
As future work we propose two extensions to com-552
plete our proposal. First, we intend to create a way to553
automatically train the aggregation operators used for554
feature fusion, so that they can adapt to specific con-555
ditions of the datasets. Second, we propose to allow556
multidimensional edge features to be aggregated, so557
that popular methods such as the Sobel method [37]558
or the Canny method [9] (which use vectorial features)559
can be combined with our multichannel feature fusion560
technique.561
References562
[1] C. Alsina, M.J. Frank and B. Schweizer, Associative Func-563
tions: Triangular Norms and Copulas, World Scientific564
Publishing Company, 2006.565
[2] P. Arbelaez, M. Maire, C. Fowlkes and J. Malik, Contour566
detection and hierarchical image segmentation, IEEE Trans567
On Pattern Analysis and Machine Intelligence 33 (2011),568
898–916.569
[3] E. Barrenechea, H. Bustince Sola, B. De Baets and C. Lopez-570
Molina, Construction of interval-valued fuzzy relations with571
application to the generation of fuzzy edge images, IEEE Trans 572
on Fuzzy Systems 19(5) (2011), 819–830. 573
[4] G. Beliakov, A. Pradera and T. Calvo, Aggregation functions: 574
A guide for Practitioners, volume 221 of Studies in Fuzziness 575
and Soft Computing, Springer, 2007. 576
[5] A. Ben Hamza, Y. He, H. Krim and A. Willsky, A multiscale 577
approach to pixel-level image fusion, Integrated Computer- 578
Aided Engineering 12 (2005), 135–146. 579
[6] J.C. Bezdek, R. Ehrlich and W. Full, FCM: The fuzzy 580