-
5374 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL.
52, NO. 9, SEPTEMBER 2014
Shadow Detection of Man-Made Buildings inHigh-Resolution
Panchromatic Satellite Images
Mohamed I. Elbakary and Khan M. Iftekharuddin, Senior Member,
IEEE
AbstractHigh-resolution satellite imagery is considered
anexcellent candidate for extracting information about the
humanactivities on Earth. The information about residential
develop-ment and suburban area mapping is of interest that can
beobtained from these images. Shadow of structures such as man-made
buildings is one of the main cues for structure detectionin
panchromatic high-resolution satellite imagery. However,
tocorrectly exploit the information of the shadow in an image,
theshadow needs to be detected and isolated first. In this paper,
wepropose a new algorithm for shadow detection and isolation
ofbuildings in high-resolution panchromatic satellite imagery.
Theproposed algorithm is based on tailoring the traditional modelof
the geometric active contours such that the new model of
thecontours is systematically biased toward segmenting the
shadowand the dark regions in the image. The systematic biasing
inthe proposed contour model is accomplished by novel encodingof
the radiometric characteristics of the shadows regions.
Afterdetecting and segmenting the shadow and the dark regions inthe
image, further processing steps are introduced. The
proposedpostprocessing is based on selection of optimal threshold
and aboundary complexity metric to distinguish the true shadows
fromthe clutter. Experimental results are presented to validate
theperformance of the proposed algorithm on real
high-resolutionpanchromatic satellite images.
Index TermsBuilding detection, geometric active contours,image
segmentation, panchromatic imagery, shadow detection.
I. INTRODUCTION AND RELATED WORK
SATELLITE and aerial imaging is a common method toobtain
information about objects on the Earths surface.Object and target
detection is of great interest for many appli-cations, including
rescue operations and defense applications.Recently, extension of
object detection to man-made structure(e.g., buildings) detection
and recognition in aerial images hasattracted attention. The
ability to effectively detect structureshelps in understanding the
scene contents of the image and maybe used for content-based
retrieval in databases and in otherapplications such as residential
development planning, damageevaluation, and military target
detection [1][5]. The shadow isa crucial cue for detecting the
existence of the buildings andother man-made structures in the
overhead images. Shadow
Manuscript received June 16, 2013; revised September 16, 2013;
acceptedOctober 10, 2013.
M. I. Elbakary is with the Electronic Research Institute, Giza
12622, Egypt.K. M. Iftekharuddin is with the Vision Laboratory,
Department of Electrical
and Computer Engineering, Old Dominion University, Norfolk, VA
23529USA.
Color versions of one or more of the figures in this paper are
available onlineat http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TGRS.2013.2288500
of buildings is isolated and then is employed for detection
byintegrating and fusing the geometry of the shadow area withthe
potential geometry of the building or other elevated man-made
structures. However, there is a general lack of shadowdetection
methods for panchromatic imagery and particularlyfor shadow of
man-made buildings in these images in literature.The recent
research for shadow detection mostly focuses ondifferent image
modalities other than panchromatic satelliteimagery.
Shadow detection of moving objects in video streams hasbeen
investigated by many researchers [6], [7]. The main pur-pose is
object detection and tracking. Others have worked onsimultaneous
detection and removal of shadows in the images[8][11]. The authors
in these works propose to alleviate theeffect of the shadow in the
image since the shadow causesloss of the color and hence loss of
feature information forobjects in the shadow areas. Few other works
propose detectionof shadows of still objects [12], [13]. Zhu et al.
[12] uselearning-based approach for shadow detection of still
objectsin grayscale images. The authors train and evaluate their
systemon a database of natural images. In [13], the proposed
algorithmis a successive thresholding scheme to enhance the ratio
mapalgorithm for shadow detection in color aerial images. Tolt et
al.[14] combine data from two different modalities for
shadowdetection. The study employed lidar with position of the
sunto detect shadow and then use that shadow in training
ofsupervised classifier to find shadow in hyperspectral data.
Otheralgorithms are proposed wherein the authors use multiple
bandsfor shadow detection in the aerial images [15][17]. The
mainpurpose for shadow detection in remotely sensed images is
toenhance the classification.
In the color imagery, the primary cue for shadow detection isthe
color in addition to the texture feature of the image contents.In
the multispectral imagery, the radiometric characteristics ofthe
bands are the main source of the information for shadowdetection.
However, detection of shadows in panchromatic im-ages is a
challenging problem. In addition to missing the colorinformation,
many objects and scene contents tend to be dark oralmost dark, and
distinguishing them from the shadow regionsadds to further
difficulty. For shadow detection and isolationof buildings in
panchromatic images, Irvin and Mckeown [18]use the shadow analysis
in images of buildings in a databaseto calculate the mean of the
shadow regions. They then employthe mean plus one standard
deviation as the shadow intensitythreshold for the input image to
detect the shadow regions. Dare[19] uses an optimal method to
obtain a threshold for isolatingthe building shadow regions from
other contents in the inputimages. However, this method requires
that the histogram of
0196-2892 2013 IEEE. Personal use is permitted, but
republication/redistribution requires IEEE permission.See
http://www.ieee.org/publications_standards/publications/rights/index.html
for more information.
-
ELBAKARY AND IFTEKHARUDDIN: SHADOW DETECTION OF MAN-MADE
BUILDINGS 5375
the input image to be bimodal to find the appropriate
threshold,which is not guaranteed in all the input images.
Recently,Luus et al. [20] have employed a threshold, as a ratio
ofthe maximum intensity value after investigating the
imagehistogram, to detect the shadow in the panchromatic
imagery.However, the algorithm in [20] suffers from the clutter,
whichprevents its generalization.
In this paper, we propose a new algorithm for shadow de-tection
of man-made buildings in high-resolution panchromaticsatellite
imagery. Our proposed method is based on modifyingthe geometric
active-contour model, which is first introducedby Chan and Vese
[21]. The geometric active-contour modelproposed in [21] is based
on Mumford and Shahs functionfor image segmentation [22]. The
contour model is representedby the zero-level set of the higher
dimensional function inthe level-set framework. These functions are
able to detect theboundary of regions based on the homogeneity of
local featuressuch as the intensity without depending on the edges
of theregions as motivation force. Few works employed both
thespectral and spatial properties in the images, such as
variance,entropy, seed pixels, and color as the forces to control
thelevel-set curve propagations [23][27]. The advantages of
thismethodology include its low sensitivity to noise and its
abilityto snap the regions and objects without clear edges. In
ourproposed approach, we modify the active contour technique in[21]
such that the new model is systematically biased towarddark
regions, including the shadow in input image. We forcedthe
algorithm in our model to favor the dark regions in theimage by
adding additional term that emphasizes the detectionof shadows in
the image. The novelty of the proposed methodinvolves development
of fully automated algorithms withoutthe need for manual insertion
of initial position of the contoursor a seed in input image. Once
the potential regions of shadoware obtained, we remove the clutter,
such as the vegetationand water bodies, in the candidate shadow
regions by furtherprocessing the results using Otsus thresholding
technique [28]integrated with a geometric filter based on the
boundary com-plexity metric (BCM) [29]. Otsus threshold removes
most ofthe clutter and helps the geometric filter to further remove
theremaining clutter regions with high boundary complexity
(BC).Clutter is naturally associated with high BCM because of
thecomplexity of its irregular boundaries. Consequently,
anothernovel contribution of this paper stems from using an
optimalthreshold for emphasizing the boundary of the regions to be
ap-propriate for distinguishing by the geometric filter.
Specifically,the geometric filter uses low BCM values to detect
shadowsof man-made objects and to remove the clutter. Employingthe
gray-level image only without using the color informationis an
additional advantage of the proposed algorithm. Theexperimental
results on real satellite images show that theproposed algorithm is
robust against the clutter and outperformsthe alternative
algorithms.
The remainder of this paper is organized as follows.Section II
describes the development of the geometric activecontours and its
application to aerial image segmentation. Theproposed algorithm is
introduced in Section III. In Section IV,we present the
experimental results, followed by conclusion inSection V.
II. GEOMETRIC ACTIVE CONTOURS FORAERIAL IMAGE SEGMENTATION
The snake model or the active contour is introduced byKass et
al. [30]. These models are energy-minimizing curves/surfaces that
move in the image domain using image featuresto accurately localize
object contours. Movement of the snakesare guided and influenced by
external and internal forces suchthat the models reach at minimal
energy by segmenting theobject. First, active-contour models are
characterized by a setof parameters, and the evolution of the
contours are performedon a predefined set of control points in the
spatial domainof the image. One of the major drawbacks of the
parametricactive contours is that the methods are influenced by the
initialconditions of the contours in the spatial domain. In other
words,the contours must be originated in the spatial domain
closeenough to the desired feature. Otherwise, the contours
convergeto the undesired objects in the image. An additional
drawback isthat, usually, each contour captures only one object in
the scene.For capturing more than one object, one must initiate a
contourfor each. That means the parametric contours do not
changetopologies during the evolution process. To overcome
thesedrawbacks of the parametric active contours, geometric
activecontours are introduced by Caselles et al. [31] and Malladi
et al.[32]. Geometric active contours are represented by the
zero-level set of a higher dimensional surface such that the
updatingof the surface function is performed in the entire image
domain.Geometric active contours may be either edge-based or
region-based methods. In edge-based methods, the gradient of
theimage is employed as an attraction force to attract the
contourto the edges of the objects in the image [31], [32]. On the
otherhand, the region-based methods employ the region featuressuch
as the gray-level intensity, texture, and other pixels statis-tics
that reflect the homogeneity of spatially localized regions asan
attraction force to these objects [21]. The advantage of thismodel
is low sensitivity to the noise [21]. In addition, geometricactive
contours implicitly handle the topological changes andare not
characterized by a set of parameters. Subsequent workshave
contributed to improved results for region-based models[33],
[24].
For a simple case in the region-based approach such asthe
geometric active-contour method [21], the image domainconsists of
background and foreground regions, which arecharacterized by
homogeneity of the gray level. The gray lev-els in these two
regions are approximately piecewise-constantintensities of
different values uo and ui, and C is the boundarybetween the two
regions. This geometric active-contour methodmay obtain good
results whether the boundaries between re-gions are well
distinguished or not. The energy equation ofthe geometric
active-contour model that extracts the objectboundary is defined as
follows [21]:
Eseg(C) =
inside(C)
|u 1|2 dx dy
+
outside(C)
|u 2|2 dx dy (1)
-
5376 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL.
52, NO. 9, SEPTEMBER 2014
where C denotes the curve of the active contour, u is the
imageintensity, and constants 1 and 2 are the average
intensitiesinside and outside of C, respectively. The curve that
minimizes(1) is the contour that fits to the edges of the object of
interest[see [21] for more details about the level-set formulation
andthe numerical implementation of (1)]. The geometric
active-contour model in [21] has been widely applied to
segmen-tation and detection objects in aerial and satellite
images.Cao et al. [34] detect man-made objects in two stages for
level-set algorithm. In the first stage, Cao et al. employ the
fractalerror image, and then they employ the discrete cosine
transformof the texture in the input image in the second stage. The
tran-sition from the first stage to the second stage is achieved
duringthe level-set evolution of the geometric active-contour model
in[21]. The work in [35] employs spectral information in the
inputimages for multiregion segmentation of multispectral
satelliteimagery based on [21]. The statistical descriptions of the
inputimages regions are integrated with the evolving curve in
Chanand Veses algorithm [21], to detect the man-made objects in
theaerial and satellite images [36]. In order to obtain the
desiredboundary of the buildings in the input image, Ahmadi et
al.[37] propose manual intervention in the framework of [21]
byinserting initial position of the active contours. In
addition,they insert grayscale intensity-based constants into the
energyequation to attract the active contours toward the
buildings.An approach of integrating the prior shapes of the
buildingsinto level-set segmentation of Chan and Veses method
[21]is proposed in [38] and [39]. The shape prior refers to
theparameterization-free description for the building
templates.However, having knowledge of shape priors for buildings
in ad-vance may not be realistic since the shape of buildings
changesdue to weather or to environment. Our proposed methodsin
this paper do not anticipate shape priors.
III. PROPOSED ALGORITHM
The general advantages and wide applications of the geomet-ric
active-contour model [21] for building and other detectionof
man-made structures motivates us to adopt this techniquefor shadow
detection of man-made structures in high-resolutionpanchromatic
satellite images. However, statistical informationand prior shape
information of the shadows are not available inthe overhead
panchromatic images. This is because the shadowareas change during
the day based on the position of the sunin the sky and the
corresponding exposition of geometry ofthe structures that generate
the associated shadow areas. Inaddition, the input panchromatic
images are very similar to thegrayscale, and the color information
and other bands informa-tion are missing. Therefore, we innovate on
the geometric activecontour by plugging additional term in the
energy equation tosystematically bias the contours toward the
boundary of thedark and shadow regions. The additional term is
based onencoding the radiometric characteristics of the dark
regionsrelatively to the neighbors, which is considered a cue
forexistence of the shadows. We further explain this concept inthe
following.
In the geometric active-contour method in [21], (1) min-imizes
energy to fit the contours between background and
foreground regions. Since the shadows in the image cannot
bedescribed by prior spectral information templates or
probabilitydue to missing or changing information, we use the
relativeintensity of the dark regions to the neighbor as a cue
forshadow detection. We reformulate the energy terms to modifythe
active contours during the process of evolution such that theenergy
function is selectively biased to enclose the dark
regions,including the shadows, from their neighbors. The
developmentof the new geometric active-contour model is presented
in thefollowing.
We reformulate (1) in Section II to detect the shadows andthe
dark regions in the input image. Assume an image withdomain R2 and
a level-set representation, i.e., R+.We form an energy functional
Etotal by adding additionalenergy Eshadow that favors the contours
surrounding shadowsand the dark region as follows:
Etotal() = Eseg() + Eshadow() (2)
where Eseg is the energy equation of (1) and Eshadow is
theproposed additional energy term to systematically bias (1)toward
dark regions. The proposed energy Eshadow affects theentire domain
of the input image and can be formulated asfollows:
Eshadow(C) =
inside(C)
|u k |2 dx dy
+
outside(C)
|u |2 dx dy (3)
where is the global mean of the input image, and k is aconstant.
The value of k is systematically chosen to encodethe radiometric
characteristics of shadow regions relative to thesurrounding pixels
in the input image as follows. Comparisonof (3) and (1) shows that
1 is equivalent to k . Therefore,following notations in (1), k is
the average intensity insideC in (3). Note that, for k < 1, the
average intensity for anyimage inside C in (3) is always less than
the global mean ofthe input image (i.e., the average intensity
outside C). In otherwords, the value of k in (3) controls the
average intensity insideC and, hence, the selection of regions by
(3). Consequently,by choosing the appropriate value of k, one can
enforce thecontours in (3) to prefer shadow regions in the
panchromaticimage. To enforce (3) to select enclosing shadow
regions, thevalue of k is chosen such that the average intensity
inside Crelative to that outside it is comparable to the average of
shadowintensity relative to the mean intensity of the image.
Shadowfundamentals suggest that k = 0.7 is a reasonable choice
toencode average shadow intensity values relative to the imagemean.
In other words, shadow pixels have a lower intensity thantheir
neighboring nonshadow pixels, and that the value of k is
areasonable ratio to encode this relation. This choice also
ensuresthat shadow and dark regions are bounded by C in (3) in
thepanchromatic images considered in this paper.
We experiment with different k values in Section IV
tosuccessfully encode the radiometric characteristics of
shadowpixels. The experiment also confirms that k = 0.7
enforces
-
ELBAKARY AND IFTEKHARUDDIN: SHADOW DETECTION OF MAN-MADE
BUILDINGS 5377
contour in (3) and hence systematically biases the model of
(2)to enclose shadow and dark regions for all the input images
inthis paper.
After inserting the regularization terms, (2) can be rewrittenas
follows:
Etotal() = Eseg() + Eshadow()
+ length(C) + area (inside(C)) (4)
where , 0 are constant parameters [21]. The termslength(C) and
area(inside(C)) are the length of contour Cand the area of the
region inside contour C, respectively [21].Equation (4)
specifically considers radiometric characteristicsof the dark
regions when compared with that of the surroundingpixels in the
input panchromatic image. Consequently, our pro-posed geometric
active-contour model in (4) is systematicallybiased to enclose the
regions that exhibit lower average inten-sity values. However,
since the vegetation and water bodies inthe panchromatic images
also show lower intensity than thesurrounding regions, the proposed
active-contour model detectand segment these regions along with the
shadow regions inan image. Therefore, further processing is
necessary to removeclutter regions such as vegetation and water
bodies from thesegmented image. The detail of this clutter removal
process ispresented in Section III-B.
A. Level-Set Formulation of the Proposed ModelFollowing the work
in [21] and [40], the level-set method
is employed to compute energy function over the input
imagedomain . In this method, curve C is represented by the
zero-level set of a function: R+, such that
C = {(x, y) : (x, y) = 0}
inside(C) = {(x, y) : (x, y) > 0}outside(C) = {(x, y) : (x,
y) < 0} .
(5)
Then, curve C is replaced by function by using the
Heavisidefunction, H , and the 1-D Dirac measure as follows:
H() =
{1, if 00, if < 0 (6)
() =d
dH(). (7)
Heaviside function H and Dirac , which is a derivative of H ,are
approximated by the following functions:
H(z) =
1, if z >0, if z < 12
[1 + z +
1 sin
(z)]
, if |z| (8)
(z) =
{0, if |z| >12 +
12 cos
(z), if |z| (9)
respectively following [41] and [42]. Note that the use of
theseapproximations avoids the boundary leakage and the existenceof
the energy term Eshadow avoids local minima. Eshadowfavors certain
regions and globally affects the energy function[38] and therefore
will drive the contours to the global minima.In addition, these
approximations are successfully adopted forsimilar application [36]
and other algorithms [43]. By usingHeaviside function from (8) and
the Dirac function from (9) andfollowing the method of level-set
numerical implementation in[21], [37], and [40], the discretization
of (4) is implemented asshown in (10) at the bottom of the page.
where h and t arethe space iteration step and the time iteration
step, respectively.1, 2, 3, and 4 are constants. In addition, 1 and
2 are theaverages intensities inside and outside C, respectively
[21] [see[21], [37], and [40] for more details about discretization
processand the constants in (10)]. The major steps for
numericalimplementation of the level-set method, as shown in (10),
aresummarized in Algorithm 1.
Algorithm 1. Proposed algorithm for segmentation
1. Initialize 0 at n = 0, where 0 defines the initial
con-tour.
2. Compute of the input image.3. Compute 1 and 2 by using
[21].4. Obtain n+1 by using (10).5. Check whether the solution is
stationary and is not vary-
ing. If yes, stop/ If not, n = n+ 1, and go to step 3.
The result of the using the new contour model in (10),
incomparison to the result of using the model in (1), is
presentedin Fig. 8 in Section IV. The result of the proposed
geometricactive-contour model in (10) is the segmentation of the
dark
n+1i,j ni,jt
= hni,j
h2
x x+n+1i, j(
x+ni,j
)2/h2 +
(ni,j+1 ni,j1
)2/(2h)2
+
h2y
y+n+1i,j(
y+ni,j
)2/h2 +
(ni+1,j ni1,j
)2/(2h)2
1 (ui,j 1n)2
+ 2 (ui,j 2n)2 3 (ui,j kn)2 + 4 (ui,j n)2]
(10)
-
5378 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL.
52, NO. 9, SEPTEMBER 2014
Fig. 1. Region combined with a strip from the surrounding
pixels.
regions, including shadow, in the input image. Therefore,
wefurther process the segmented regions to isolate the shadowregion
from the clutter by employing Otsus threshold [28] withthe BCM in
the following.
B. Shadow Isolation From Clutter
The dark regions obtained by the proposed model for thegeometric
active contours include shadow and clutter such aswater bodies,
vegetation, and dark grounds. In this paper, weemploy the following
steps to remove clutter.
The result of the given modified model in (10) is a set
ofpotential shadow regions R = {R0, R1, . . . , Rn1} that
mayinclude clutter. To remove the clutter from the potential
shadowregions, we propose further processing of the result as
follows.We construct Otsus threshold on the global histogram ofthe
detected regions after adding to each region a strip, withthickness
of three pixels, from the surrounding, as shown inFig. 1. Otsus
threshold is an optimal threshold method toautomatically segment a
bimodal histogram. Our experimentshows that three pixels from the
surrounding for each region areenough to obtain a global bimodal
histogram for the detectedregions. In that bimodal histogram, one
peak is expected toinclude the shadow and similar intensity pixels
in the seg-mented regions. The other peak will include the
surroundingspixels and similar intensity pixels. The global Otsus
thresholdin this step removes the clutter which fall within the
peak ofthe surrounding pixels while at the same time increases
theirregularity and complexity of the boundaries of the
remainingclutter that fall between the two peaks of the bimodal
histogram.High BC helps in removing the clutter by using the
geometricfilter as discussed in the following. The given global
processingsteps are summarized in Fig. 2.
Next, we use a geometric filter to remove the clutter in Rsby
processing each region Rsi in Rs. The geometric filter
isimplemented by comparing the BC measure of each regionwith a
threshold to filter out the regions of values higherthan the
threshold. The BC measure reflects the regularity andthe complexity
of the boundary [29]. Fig. 3 shows examplesof the BC of regular and
irregular regions. We notice thatthe regular boundaries have
smaller BC values than irregularboundaries. Moreover, the BC value
increases as the complexityof the boundary increases. The BC
measure is higher for theclutter because the clutter, such as
vegetation and the water
Fig. 2. Procedure for the proposed global thresholding.
Fig. 3. Examples of regular and irregular boundaries. First row:
regularboundaries with their BC. Second row: irregular boundaries
with their BC.
body regions, have boundaries that tend to be irregular
andcomplicated. On the other hand, the shadows of the buildingsand
other man-made structures have boundaries that tend to beregular,
and in general, zero BC is associated to straight lines,as shown in
the first row in Fig. 3. The value of the threshold forthe
geometric filter is chosen to be 0.15 after extensive testingwith
all the regions in all the input images in this paper to keepthe
shadows of interest and filter out the maximum amount ofclutter. We
did not choose that threshold equals to zero since theshadow
boundary of man-made structures such as building isnot ideal in all
the places but may be associated with complexitythat increases the
BC measure, as shown in the example ofthe first row in Fig. 3. The
operation of the geometric filter issummarized in Fig. 4.
Finally, we declare a region is true shadow after processingthat
region with another local Otsus threshold and then testingits BC
using the above geometric filter. Our reasoning for thisstep is
explained as follows. The two steps discussed in Figs. 2and 4 help
in removing mainly the bright clutter. Here, brightclutter means
that it does not fall within the shadow peak in theglobal bimodal
histogram in procedure GlobalOtsuThreshold()in Fig. 2 and the
clutter that have a high BC measure usingprocedure
GeometricFilter() in Fig. 4. However, the globalprocessing for the
region boundary using algorithms in Figs. 2
-
ELBAKARY AND IFTEKHARUDDIN: SHADOW DETECTION OF MAN-MADE
BUILDINGS 5379
Fig. 4. Procedure for the proposed geometric filter.
Fig. 5. Procedure for the proposed local processing.
and 4 may not be enough to clearly identify the shadow fromthe
clutter. Therefore, we use another local processing step,i.e.,
LocalOtsuThresold(), to strongly emphasize the regionboundaries
such that the geometric filter readily detects thetrue shadow. The
second row in Fig. 3 shows an example forthe expected result from
applying local processing whereinthe BC increased to 0.1584. This
procedure is summarizedin Fig. 5.
In Fig. 5, for each remaining region Rsi , we construct anduse
its Otsus threshold after adding a strip from its
surroundingpixels. The histogram of a region of the shadow of
man-madestructures such as buildings with the surrounding pixels
willconsist of two distinctive peaks for Otsus threshold becausethe
shadow is well defined compared with the surrounding. Onthe other
hand, the two peaks in the histogram of the regionof clutter with
its surrounding are not well defined since theirregularity and the
complexity of its boundary and using itslocal Otsus threshold will
reveal a high BC measure for thegeometric filter. After applying
the local threshold, we employthe geometric filter to remove the
regions that have BC measuremore than 0.15. Fig. 6 shows the
schematic of the wholeproposed algorithm.
In summary, we propose an algorithm for detection
andsegmentation of shadow of man-made buildings in panchro-matic
satellite images. The algorithm is fully automated
andsystematically biased for shadow detection. We remove theclutter
associated with shadow by using an optimal segmenta-tion
methodology integrated with a geometric filter scheme todistinguish
the clutter from the shadow of man-made structures.
Fig. 6. Schematic of the proposed algorithm.
IV. EXPERIMENTAL RESULTS AND DISCUSSIONS
Here, we present the results of our proposed algorithm forshadow
detection of man-made objects in real panchromaticsatellite
imageries for various challenging scenes. The imagesused in this
paper are publicly provided by the U.S. GeologicalSurvey, and each
image consists of four bands (RGBIR) witha resolution of one pixel
per meter in both directions. Forthese color images, we created the
corresponding panchromaticimages for our application. The images
contain various objects,including vegetation, buildings, towers,
water bodies, roads,and other man-made objects such as cars and
trucks. We choosethe scenes such that they contain various
scenarios for buildingscontents such as isolated buildings,
connected buildings, smallbuilding, and large buildings. In
addition, the shadow contentsin the input images vary in the shape
and the size such thatthe shadows might be represented by a few
pixels to a largeregion. In addition, the shadow region can either
be isolated or
-
5380 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL.
52, NO. 9, SEPTEMBER 2014
Fig. 7. (a) Input satellite image used in the first experiment.
(b) Initial contourgenerated in the input image. (c) Ground truth
of the shadow for man-madebuildings. (d) Evolution of geometric
active contours after 20 iterations.
connected to vegetation and/or water bodies. The performanceof
the proposed algorithm is evaluated both by using qualitativeand
quantitative metrics. The quantitative evaluation of theresults is
based on the ground-truth data of the shadows inthe input images,
which are derived manually by labeling theshadow pixels of man-made
structures to zero. Following [21],we choose the values of
parameters in (10): 1=2=1, t=0.1, =1, =1, = 1, and h=1. Moreover,
we set the remain-ing parameter values 3=4=1 similar to 1 and 2 in
[21].
In the first experiment, we use the input image shown inFig.
7(a). The initial contour of the geometric active contouris
generated automatically in the middle of the input image asa
regular circle with a diameter that is proportional with thesize of
the input image, as shown in Fig. 7(b). After few itera-tions, the
active contours start to surround all the dark regionsincluding the
shadow regions. Note that the contours favorand enclose the regions
that are darker than the surroundingsuch as the shadows,
vegetation, and water bodies, as shown inFig. 7(d). Fig. 7(c) shows
the ground truth of the shadow in theinput image in Fig. 7(a). In
addition to the shadow, the contourssurround a clutter such as the
vegetation and the water bodies(e.g., lakes and canals) since these
regions are darker than thesurrounding in the gray-level
images.
Fig. 8(a) shows the shadow regions segmented by usingthe
proposed contour model in (10), and the shadow regionssegmented by
the contour model in (1) is presented in Fig. 8(b).The results of
segmentation in Fig. 8(a) and (b) demonstratethe ability of the new
contour model to distinguish the isolated
Fig. 8. (a)Shadow regions segmented by using the new contour
model in (10).(b) Shadow regions segmented by using the contour
model in (1). (c) Shadowregions segmented by using (10) with k =
0.9. (d) Shadow regions segmentedby using (10) with k = 0.5. (e)
Shadow regions segmented by using (10) withk = 0.3.
shadow and the dark regions in the input image
significantlyefficient than the regular control model in (1). Fig.
8(c)(e)present the segmentation results in (10) with k = 0.9, k =
0.5,and k = 0.3, respectively. Fig. 8(c) shows that k = 0.9 tends
tooversegment shadow regions since the corresponding
segmentsinclude other regions that are not as dark as the shadow.
Forexample, the segmented regions in the top and bottom leftare
segmented, although these are not shadow regions in thepanchromatic
image. On the other hand, k = 0.7 and k = 0.5;the model tends to be
undersegmented and misses some shadowregions, as shown in the
buildings in the middle of Fig. 8(d).The same observation is very
obvious in Fig. 8(e) for k = 0.3.These shadow segmentation results
in Fig. 8(c)(e) confirmthat k = 0.7 is a reasonable choice, as
discussed in Section III.
-
ELBAKARY AND IFTEKHARUDDIN: SHADOW DETECTION OF MAN-MADE
BUILDINGS 5381
Fig. 9. (a) Result from the proposed algorithm. (b) Result from
Daresmethod. (c) Result from the fixed masking method.
To isolate the shadow of man-made buildings from the clutterin
the segmented regions, further processing is introduced,
asdemonstrated in Fig. 6. Fig. 9(a) shows the final result of
theproposed algorithm for the input image in Fig. 7(a). Note
thatour algorithm shows good detection of the shadows of the
man-made buildings in the scene when compared with the groundtruth
in Fig. 7(c). Further note that our algorithm detects theshadow of
the isolated buildings; however, the shadow of theconnected
buildings is segmented as one region. In addition,the algorithm
detects the shadows of large buildings such astowers and that of
most of the small buildings. Our techniqueis able to remove all the
clutter from the image, except the lakethat is connected to the
shadow of the lower tower. Therefore,we can claim that the
algorithm is robust for the clutter, whichis common in overhead
scenes. It is worth noting that the waterbody is hard to separate
from the shadow of the tower sincethe water body is adjacent and
connected to the shadow of thetower, and they exhibit similar
intensity feature characteristics.Hence, the water body and the
shadow of the tower are treatedas one object by the geometric
active-contour method.
For performance comparison, we evaluate the results of
theproposed algorithm with the results of the method used in
[19]for shadow detection of buildings in satellite images.
Moreover,we implement the algorithm in [20] for comparison with
ourtechnique since this algorithm handles a problem of
shadowdetection by using the gray level in high-resolution
overheadsatellite imagery. We refer to the method of [19] as
Daresmethod and the method in [20] as the fixed masking method.
The results of Dares method and the fixed masking method forthe
input image in Fig. 7(a) are presented in Fig. 9(b) and
(c),respectively. Comparing our final result in Fig. 9(a) with
resultsin Fig. 9(b) and (c), our proposed algorithm obviously
per-forms better in overhead building shadow segmentation.
Daresmethod detects all the buildings shadow regions in the
image;however, the boundaries are not clear, and the result
includesnumerous clutter. Similar to our result, Dares method
detectsthe water body connected to the shadow of the lower tower.
Thefixed masking method detects less clutter than Dares methodbut
misses many shadow pixels as well.
To further validate performance, we process the imagesin Figs.
10(a)12(a) using our proposed algorithm. Thesepanchromatic images
cover completely different scenes, and theshadow contents with the
clutter take various challenging sce-narios. The corresponding
results of the manually interpretedshadow maps, which are used as
ground-truth data for theevaluation of the algorithms, are
presented in Figs. 10(b)12(b),respectively.
For the image in Fig. 10(a), we present the results of the
in-termediate processing in the proposed algorithm to show
theeffect of each process in removing the clutter and isolating
thedesired shadow. Fig. 10(c) shows the result of segmentation
byusing (10). In this step, the algorithm clearly segmented thedark
regions beside the shadow since the vegetation regionsand the water
bodies exhibit gray-level characteristics similarto the shadow in
the panchromatic images. In order to removethe clutter, we apply
three steps as follows.
First, we construct the Otsus threshold on the histogram ofthe
detected and segmented regions by the geometric activecontours
after adding for each region a strip from the sur-rounding, as
explained in Section III. Fig. 10(d) presents theresult of using
Otsus threshold with global processing. Thisstep alleviates the
clutter by mainly removing the clutter thatis brighter than the
shadow, and this is clear in the roof of thebuilding in the bottom
of the image and in some bodies of water.The result of this step is
a set of potential shadow regions.
Second, we obtain the BCM for each region and removethe regions
which have BC of more than 0.15, as shown inFig. 10(e). As
discussed in Section III-B, the geometric filteruses the advantages
of irregularity and complexity of the clutterthat results from the
global processing. We find 0.15 to be rea-sonable for all our
experiments in this paper to keep the shadowregions of man-made
structures and to simultaneously filterout the maximum amount of
the clutter. Comparing Fig. 10(d)and(e) shows the effect of using
the value of 0.15, increasingthat value adds more clutter to Fig.
10(e), and reducing it mayremove the desired shadow regions.
Third, to remove the remaining clutter, we distinguish be-tween
them and the shadow by constructing local histogramfor each region
with its surrounding strip. Fig. 10(f) showsthe result of using
Otsus threshold with local processing. Theapplication of local
Otsus threshold of each local histogramclearly emphasizes the
boundaries and removes the pixels thatcan be considered nonshadow
in the perspective of bimodalhistogram for Otsus threshold, as
described in Section III-B.Emphasizing the boundaries help in
obtaining higher valuesfor BCM for the clutter since the clutter
has irregular and
-
5382 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL.
52, NO. 9, SEPTEMBER 2014
Fig. 10. (a) Input satellite image. (b) Ground truth of the
shadow for man-made buildings. (c) Result of segmentation by (10).
(d) Result of using Otsusthreshold with global processing. (e)
Result of using the geometric filter.(f) Result of using Otsus
threshold with local processing. (g) Result of usingthe geometric
filter. (h) Shadow detection result by Dares method. (i)
Shadowdetection by the fixed masking method.
Fig. 11. Comparison of shadow detection results. (a) Original
image.(b) Shadow ground truth. (c) Shadow detection result by the
proposed algo-rithm. (d) Shadow detection result by Dares method.
(e) Shadow detection bythe fixed masking method.
complicated boundaries and in obtaining low values for
thedesired shadow since it has regular boundaries, as we
explainedin Fig. 3. Emphasizing the boundaries is clear in the
shadowof the building in the bottom of the image, and
removingnonshadow pixels is obvious in the remaining clutter.
Then, we employ the geometric filter again for each region
toremove the regions that have a BC value above 0.15 to declarethe
detection of the shadow of buildings. Fig. 10(g) shows theresult of
the proposed algorithm. The result of Dares methodand the fixed
masking method are shown in Fig. 10(h) and (i),respectively.
For the images in Figs. 11(a) and 12(a), the shadow
detectionresults of the proposed algorithm are presented in Figs.
11(c)and 12(c), and the results of Dares method are shown inFigs.
11(d) and 12(d), respectively. The corresponding resultsfrom the
fixed masking method are shown in Figs. 11(e) and12(e). For all the
testing images in Figs. 1012, the proposedalgorithm detects shadow
maps that are quite similar or veryclose to the ground-truth map.
However, the results also include
-
ELBAKARY AND IFTEKHARUDDIN: SHADOW DETECTION OF MAN-MADE
BUILDINGS 5383
Fig. 12. Comparison of shadow detection results. (a) Original
image.(b) Shadow ground truth. (c) Shadow detection result by the
proposed algo-rithm. (d) Shadow detection result by Dares method,
and (e) shadow detectionby the fixed masking method.
the dark cars in the scene since they show well
distinguishableboundary in grayscale images. In comparison, Dares
methodand the fixed masking method fail to separate the shadowsfrom
the surrounding vegetation in the images, and most ofthe vegetation
and the water bodies are detected as shadow, asshown in the
results. Moreover, Dares method and the fixedmasking method detect
more cars than the proposed algorithm.The aforementioned
qualitative evaluation experiments demon-strate that the proposed
algorithm offers better performance,and it is robust against the
clutter.
To further investigate the sensitivity of the
segmentationalgorithm in (10) due to initialization, we repeat an
examplefrom one of experiments with different level-set
initializations,as shown in Fig. 13. The contours can be set at any
place in theimage. Although these initializations are different, we
obtainthe same final segmentation results, as reported in Figs.
912.Therefore, we do not repeat the segmentation results here.
Fig. 13. Different initializations for the level-set function in
the proposedsegmentation algorithm.
These findings confirm our hypothesis that our segmentationmodel
in (2) is not sensitive to the initialization since this
modelconsists of the model in [21], which is not sensitive to
theinitialization [21], [33], [36].
To quantitatively compare the performance, we evaluate
theproposed algorithm using well-known quantitative metrics
[6],[7], [13]. We use these metrics to evaluate the accuracy ofthe
proposed building shadow segmentation to that of Daresmethod and
the fixed masking method. Three types of metricsare adopted as
follows: The first type of metrics is namedproducers accuracies,
which measure the correctness of thealgorithm and indicate how well
the true shadow and non-shadow pixels are correctly classified. The
second type ofmetrics is the users accuracies that measure the
precision ofthe algorithm and indicate the probabilities of
correctly detectedand classified pixels (shadow and nonshadow). The
third typeof metrics is the overall accuracy that measures if the
percent-age correct. The producers accuracy of the shadow s
andproducers accuracy of nonshadow ns are defined by
s =TP
TP + FN(11)
ns =TN
TN + FP(12)
where true positive (TP) is the number of shadow pixels,
whichare correctly detected and identified when compared with
theground truth, and false negative (FN) is the number of thetrue
shadow pixels, which are detected and identified by analgorithm as
nonshadow pixels. True negative (TN) denotesthe number of true
nonshadow pixel that are correctly detected
-
5384 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL.
52, NO. 9, SEPTEMBER 2014
TABLE IDETECTION ACCURACY MEASUREMENTS FOR FIG. 9
TABLE IIDETECTION ACCURACY MEASUREMENTS FOR FIG. 10
TABLE IIIDETECTION ACCURACY MEASUREMENTS FOR FIG. 11
TABLE IVDETECTION ACCURACY MEASUREMENTS FOR FIG. 12
and identified by an algorithm, and false positive (FP) is
thenumber of nonshadow pixels that are detected and identifiedby an
algorithm as true shadow pixels. The users accuracyof shadow s and
the users accuracy of nonshadow ns aredefined by
s =TP
TP + FP(13)
ns =TN
TN + FN. (14)
The overall accuracy is defined by
=TP + TN
TP + TN + FP + FN(15)
where TP + TN is the number of correctly detected and
iden-tified true shadow and nonshadow pixels and TP + TN + FP +FN
is the total number of pixels in the input image.
Tables IIV show the values of the quantitative metricsfor the
results in the experiments of images in Figs. 912.
The results in these tables suggest that the proposed algo-rithm
significantly outperforms Dares method and the fixedmasking method
in three categories (ns, s, ) and producescomparable performance in
ns. Moreover, our method offerslower performance in s when compared
with Dares methodand comparable results, on the average, to the
fixed maskingmethod. The lower performance for s is due to changing
thelabel of some pixels from shadow to nonshadow during theclutter
detection process. The change in pixel labels is causedby higher
brightness values of these pixels in the shadow area,and this, in
turn, is reflected in the increased FN value andhence reduced s
values, respectively. Although the increasedFN value is not
desirable, it does not affect the overall accuracy of the proposed
algorithm, as shown in the tabulated results.Note that a similar
trend is observed in [13].
Our improved geometric active-contour model is derivedfrom the
level-set-based model of [21]. In general, the reini-tialization
step in the level set is usually time-consuming andalso has
expensive computational cost [43]. The reinitialization
-
ELBAKARY AND IFTEKHARUDDIN: SHADOW DETECTION OF MAN-MADE
BUILDINGS 5385
TABLE VPROCESSING TIME OF THE PROPOSED SHADOW DETECTION
APPROACH
step in [21] is optional; hence, all our results are
obtainedwithout reinitialization. In order to study the
computational per-formance, Table V presents the elapsed time for
processing theimages in this paper. All algorithms are implemented
in MatlabR2012a on a laptop computer with an 1.65-GHz AMD E-450APU
processor with 4-GB RAM. Note that the processing timeis highly
dependent on the way Matlab scripts are written, thedetails in the
image contents, and the dimension of the image.In Table V, we
observe that, for most images, at least halfof the total processing
time is consumed in implementing thelevel-set segmentation step.
Furthermore, the processing timefor shadow separation step depends
on the amount of clutter,the spatial distribution of clutter, and
the steps to remove theclutter. Consequently, the processing time
to remove clutter inFig. 11(a) is more than that in Fig. 7(a).
V. CONCLUSION
In this paper, we present a novel algorithm for overheadshadow
detection and extraction from high-resolution panchro-matic
satellite images. The proposed technique segments theshadow of
man-made structures such as buildings by using thegray-level
satellite image without using the color information.The algorithm
is based on employing an improved geometricactive-contour model to
handle a rather difficult problem ofshadow detection in satellite
imagery. Our proposed model ofthe geometric active contours
systematically favors the shadowand the similar dark regions in the
input image. Further pro-cessing steps are introduced to isolate
the shadow from cluttersuch as vegetation and water bodies. Both
qualitative andquantitative experimental results using real images
show thatthe proposed algorithm outperforms a comparable
algorithmfor shadow and man-made structure segmentation. The
resultof the proposed algorithm may be used for overhead
man-madebuilding segmentation in high-resolution satellite
images.
REFERENCES[1] J. Yuan, D. Wang, B. Wu, L. Yan, and R. Li,
LEGION-based automatic
road extraction from satellite imagery, IEEE Trans. Geosci.
RemoteSens., vol. 49, no. 11, pp. 45284538, Nov. 2011.
[2] A. Turlapaty, B. Gokaraju, D. U. Qian, N. H. Younan, and J.
V. Aanstoos,A hybrid approach for building extraction from
spaceborne multi-angularoptical imagery, IEEE J. Sel. Topics Appl.
Earth Observ. Remote Sens.,vol. 5, no. 1, pp. 89100, Feb. 2012.
[3] X. Tong, X. Lin, T. Feng, H. Xie, S. Liu, Z. Hong, and P.
Chen, Useof shadows for detection of earthquake-induced collapsed
buildings inhigh-resolution satellite imagery, ISPRS J. Photogramm.
Remote Sens.,vol. 79, pp. 5367, May 2013.
[4] X. Huang and L. Zhang, Morphological building/shadow index
for build-ing extraction from high-resolution imagery over urban
areas, IEEE J.Sel. Topics. Appl. Earth Observ. Remote Sens., vol.
5, no. 1, pp. 161172,Feb. 2012.
[5] N. Longbotham, C. Chaapel, L. Bleiler, and C. Padwick, Very
highresolution multiangle urban classification analysis, IEEE
Trans. Geosci.Remote Sens., vol. 50, no. 4, pp. 11551170, Apr.
2012.
[6] N. Al-Najdawi, H. E. Bez, J. Singhai, and E. A. Edirisinghe,
A survey ofcast shadow detection algorithms, Pattern Recognit.
Lett., vol. 33, no. 6,pp. 752764, Apr. 2012.
[7] A. Sanin, C. Sanderson, and B. C. Lovell, Shadow detection:
A surveyand comparative evaluation of recent methods, Pattern
Recognit., vol. 45,no. 4, pp. 16841695, Apr. 2012.
[8] Z. Liu, K. Huang, and T. Tan, Cast shadow removal in a
hierarchicalmanner using MRF, IEEE Trans. Circuits Syst. Video
Tech., vol. 22, no. 1,pp. 5666, Jan. 2012.
[9] E. Arbel and H. Hel-Or, Shadow removal using intensity
surfaces andtexture anchor points, IEEE Trans. Pattern Anal. Mach.
Intell., vol. 33,no. 6, pp. 12021216, Jun. 2011.
[10] R. Guo, Q. Dai, and D. Hoiem, Single-image shadow detection
andremoval using paired regions, in Proc. CVPR, Jun. 2025, 2011,pp.
20332040.
[11] L. Lorenzi, F. Melgani, and G. Mercier, A complete
processing chainfor shadow detection and reconstruction in VHR
images, IEEE Trans.Geosci. Remote Sens., vol. 50, no. 9, pp.
34403452, Sep. 2012.
[12] J. Zhu, K. G. G. Samuel, S. Z. Masood, and M. F. Tappen,
Learning torecognize shadows in monochromatic natural images, in
Proc. CVPR,2010, pp. 223230.
[13] K. L. Chung, Y. R. Lin, and Y. H. Huang, Efficient shadow
detectionof color aerial images based on successive thresholding
scheme, IEEETrans. Geosci. Remote Sens., vol. 47, no. 2, pp.
671682, Feb. 2009.
[14] G. Tolt, M. Shimoni, and J. Ahlberg, A shadow detection
method forremote sensing images using VHR hyperspectral and LIDAR
data, inProc. IGARSS, Jul. 2529, 2011, pp. 44234426.
[15] J. Liu, T. Fang, and D. Li, Shadow detection in remotely
sensed imagesbased on self-adaptive feature selection, IEEE Trans.
Geosci. RemoteSens., vol. 49, no. 12, pp. 50925103, Dec. 2011.
[16] K. R. M. Adeline, M. Chen, X. Briottet, S. K. Pang, and N.
Paparoditis,Shadow detection in very high spatial resolution aerial
images: A com-parative study, ISPRS J. Photogramm Remote Sens.,
vol. 80, pp. 2138,Jun. 2013.
[17] A. Makarau, R. Richter, R. Muller, and P. Reinartz,
Adaptive shadowdetection using a blackbody radiator model, IEEE
Trans. Geosci. RemoteSens., vol. 49, no. 6, pp. 20492059, Jun.
2011.
[18] R. Irvin and D. McKeown, Methods for exploiting the
relationship be-tween buildings and their shadows in aerial
imagery, IEEE Trans. Syst.,Man, Cybern., vol. 19, no. 6, pp.
15641575, Nov./Dec. 1989.
[19] P. M. Dare, Shadow analysis in high-resolution satellite
imagery of urbanareas, Photogramm. Eng. Remote Sens., vol. 71, no.
2, pp. 169177,2005.
[20] F. P. S. Luus, F. van den Bergh, and B. T. J. Maharaj, The
effectsof segmentation-based shadow removal on across-date
settlement typeclassification of panchromatic quickbird images,
IEEE J. Sel. TopicsAppl. Earth Observations Remote Sens., vol. 6,
no. 3, pp. 12741285,Jun. 2013.
[21] T. F. Chan and L. A. Vese, Active contours without edges,
IEEE Trans.Image Process., vol. 10, no. 2, pp. 266277, Feb.
2001.
[22] D. Mumford and J. Shah, Optimal approximation by piecewise
smoothfunctions and associated variational problems, Commun. Pure
Appl.Math., vol. 42, no. 5, pp. 577685, Jul. 1989.
[23] C. Li, J. Liu, and M. D. Fox, Segmentation of external
force field for au-tomatic initialization and splitting of snakes,
Pattern Recognit., vol. 38,no. 11, pp. 19471960, Nov. 2005.
[24] J. Lie, M. Lysaker, and X. C. Tai, Binary level set model
and someapplications to Mumford-Shah Image segmentation, IEEE
Trans. ImageProcess., vol. 15, no. 5, pp. 11711181, May 2006.
[25] A. Mansouri, A. Mitiche, and C. Vazquez, Multiregion
competition:A level set extension of region competition to multiple
region image
-
5386 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL.
52, NO. 9, SEPTEMBER 2014
partitioning, Comput. Vis. Image Und., vol. 101, no. 3, pp.
137150,Mar. 2006.
[26] S. Osher and N. Paragios, Geometric Level Set Methods in
Imaging, Visionand Computer Graphics. New York, NY, USA:
Springer-Verlag, 2003.
[27] J. E. Ball and L. M. Bruce, Level set segmentation of
remotely sensedhyperspectral images, in Proc. IEEE Geosci. Remote
Sens. Symp., 2005,vol. 8, pp. 56385642.
[28] N. Otsu, A Threshold selection method from gray-level
histograms,IEEE Trans. Syst., Man., Cybern., vol. SMC-9, no. 1, pp.
6266, Jan. 1979.
[29] H. Su, A. Bouridane, and D. Crookes, Scale adaptive
complexity measureof 2-D shapes, in Proc. IEEE ICPR, 2006, pp.
134137.
[30] M. Kass, A. Witkin, and D. Terzopoulos, Snakes: Active
contour mod-els, Int. J. Comput. Vis., vol. 1, no. 4, pp. 321331,
Jan. 1988.
[31] V. Caselles, F. Catte, T. Coll, and F. Dibos, A geometric
model for activecontours in image processing, Numer. Math., vol.
66, no. 1, pp. 131,1993.
[32] R. Malladi, J. A. Sethian, and B. C. Vemuri, Shape modeling
with frontpropagation: A level set approach, IEEE Trans. Pattern
Anal. MachIntell., vol. 17, no. 2, pp. 158175, Feb. 1995.
[33] L. A. Vese and T. F. Chan, A multiphase level set framework
for imagesegmentation using the mumford and shah model, Int. J.
Comput. Vis.,vol. 50, no. 3, pp. 271293, Dec. 2002.
[34] G. Cao, X. Yang, and Z. Mao, A two-stage level set
evolution scheme forman-made objects detection in aerial images, in
Proc. CVPR, San Diego,CA, USA, 2005, pp. 474479.
[35] O. Besbes, Z. Belhad, and N. Boujemaa, Adaptive satellite
images seg-mentation by level set multiregion competition, INRIA,
Paris, France,Tech. Rep. No. 5855, 2006.
[36] K. Karantzalos and D. Argialas, A region-based level set
segmentationfor automatic detection of man-made objects from aerial
and satelliteimages, Photogramm. Eng. Remote Sens., vol. 75, no. 6,
pp. 667677,Jun. 2009.
[37] S. Ahmadi, M. J. V. Zoej, H. Ebadi, H. A. Moghaddam, andA.
Mohammadzadeh, Automatic urban building boundary extractionfrom
high resolution aerial images using an innovative model of
activecontours, Int. J. Appl. Earth Observ. Geoinf., vol. 12, no.
3, pp. 150157,Jun. 2010.
[38] K. Karantzalos and N. Paragios, Automatic model-based
building detec-tion from single panchromatic high resolution
images, in Proc. ISPRSArchives, 2008, pp. 127132.
[39] K. Karantzalos and N. Paragios, Recognition-driven two
dimensionalcompeting prior towards automatic and accurate building
detection,IEEE Trans. Geosci. Remote Sens., vol. 47, no. 1, pp.
133144, Jan. 2009.
[40] S. Osher and J. A. Sethian, Fronts propagating with
curvature-dependentspeed: Algorithms based on Hamilton-Jacobi
formulation, J. Comput.Phys., vol. 79, no. 1, pp. 1249, Nov.
1988.
[41] H. K. Zhao, T. Chan, B. Merriman, and S. Osher, Variational
level setmethod approach to multiphase motion, J. Comput. Phys.,
vol. 127, no. 1,pp. 179195, Aug. 1996.
[42] C. Li, C. Xu, C. Gui, and M. D. Fox, Distance regularized
level setevolution and its application to image segmentation, IEEE
Trans. Image.Process., vol. 19, no. 12, pp. 32433254, Dec.
2010.
[43] K. Zhang, L. Zhang, H. Song, and D. Zhang,
Reinitialization-free levelset evolution via reaction diffusion,
IEEE Trans. Image Process., vol. 22,no. 1, pp. 258271, Jan.
2013.
Mohamed I. Elbakary received the B.S. and M.S.degrees in
electronic engineering from the Universityof El Mansoura, El
Mansoura, Egypt, in 1991 and1996, respectively, and the Ph.D.
degree in electricaland computer engineering from the University
ofArizona, Tucson, AZ, USA, in 2005.
Until 2000, he worked as a Researcher Assistantwith the
Department of Computers and Systems,Electronic Research Institute
(ERI), Giza, Egypt.After obtaining the Ph.D. degree, he joined the
re-search group for hyperspectral image processing in
the Department of Electrical and Computer Engineering,
University of SouthAlabama, Mobile, AL, USA, where he worked until
2008. He worked asa Visiting Assistant Professor and a Research
Scholar with the Departmentof Computer Systems, Taif University,
Taif, Saudi Arabia, and with the De-partment of Electrical and
Computer Engineering, Old Dominion University,Norfolk, VA, USA,
respectively. He is currently an Assistant Professor withERI. His
research interests include image processing and analysis,
patternrecognition, image registration, superresolution imaging
(2-D/3-D), intelligentsystems, computer vision, and
multispectral/hyperspectral image processingand analysis.
Khan M. Iftekharuddin (SM02) received the B.Sc.degree from the
Bangladesh Institute of Technology,Dhaka, Bangladesh, in 1989 and
the M.S. and Ph.D.degrees both in electrical engineering from the
Uni-versity of Dayton, Dayton, OH, USA, in 1991 and1995,
respectively.
He is currently a Professor and Chair with theDepartment of
Electrical and Computer Engineering,the Director of the Vision
Laboratory, a memberof a biomedical engineering program at Old
Do-minion University, Norfolk, VA, USA. He is the
author of a book, several book chapters, and more than 150
refereed journaland conference papers. His research has been funded
by different agenciessuch as the National Science Foundation, the
National Institute of Health,Air Force Office of Scientific
Research, Army Research Office, Air ForceResearch Laboratory, U.S.
Department of Transportation, U.S. Departmentof Energy, Whitaker
Foundation, Assisi Foundation, and FedEx Institute ofTechnology.
His research interests include computational modeling of
intel-ligent systems and reinforcement learning, stochastic medical
image analysisfor tumor phenotype extraction, intersection of
bioinformatics and medicalimage analysis, distortion-invariant
automatic target recognition, biologicallyinspired human and
machine centric recognition, recurrent networks for
visionprocessing, probabilistic vision for robotics, emotion
detection from speech anddiscourse, sensor signal acquisition and
modeling, and optical computing andinterconnection.
Dr. Iftekharuddin is a Fellow of the International Society for
Optics and Pho-tonics (SPIE) and a Senior Member of The Optical
Society (OSA). He servesas an Associate Editor for multiple
journals, including Optical Engineering andComputer Methods in
Biomechanics and Biomedical Engineering: Imaging
andVisualization.
/ColorImageDict > /JPEG2000ColorACSImageDict >
/JPEG2000ColorImageDict > /AntiAliasGrayImages false
/CropGrayImages true /GrayImageMinResolution 300
/GrayImageMinResolutionPolicy /OK /DownsampleGrayImages true
/GrayImageDownsampleType /Bicubic /GrayImageResolution 300
/GrayImageDepth -1 /GrayImageMinDownsampleDepth 2
/GrayImageDownsampleThreshold 1.50000 /EncodeGrayImages true
/GrayImageFilter /DCTEncode /AutoFilterGrayImages false
/GrayImageAutoFilterStrategy /JPEG /GrayACSImageDict >
/GrayImageDict > /JPEG2000GrayACSImageDict >
/JPEG2000GrayImageDict > /AntiAliasMonoImages false
/CropMonoImages true /MonoImageMinResolution 1200
/MonoImageMinResolutionPolicy /OK /DownsampleMonoImages true
/MonoImageDownsampleType /Bicubic /MonoImageResolution 600
/MonoImageDepth -1 /MonoImageDownsampleThreshold 1.50000
/EncodeMonoImages true /MonoImageFilter /CCITTFaxEncode
/MonoImageDict > /AllowPSXObjects false /CheckCompliance [ /None
] /PDFX1aCheck false /PDFX3Check false /PDFXCompliantPDFOnly false
/PDFXNoTrimBoxError true /PDFXTrimBoxToMediaBoxOffset [ 0.00000
0.00000 0.00000 0.00000 ] /PDFXSetBleedBoxToMediaBox true
/PDFXBleedBoxToTrimBoxOffset [ 0.00000 0.00000 0.00000 0.00000 ]
/PDFXOutputIntentProfile (None) /PDFXOutputConditionIdentifier ()
/PDFXOutputCondition () /PDFXRegistryName () /PDFXTrapped
/False
/Description > /Namespace [ (Adobe) (Common) (1.0) ]
/OtherNamespaces [ > /FormElements false /GenerateStructure
false /IncludeBookmarks false /IncludeHyperlinks false
/IncludeInteractive false /IncludeLayers false /IncludeProfiles
false /MultimediaHandling /UseObjectSettings /Namespace [ (Adobe)
(CreativeSuite) (2.0) ] /PDFXOutputIntentProfileSelector
/DocumentCMYK /PreserveEditing true /UntaggedCMYKHandling
/LeaveUntagged /UntaggedRGBHandling /UseDocumentProfile
/UseDocumentBleed false >> ]>> setdistillerparams>
setpagedevice