building extraction from high-resolution optical spaceborne images using the integration of support...

12
International Journal of Applied Earth Observation and Geoinformation 34 (2015) 58–69 Contents lists available at ScienceDirect International Journal of Applied Earth Observation and Geoinformation jo ur nal home p age: www.elsevier.com/locate/ jag Building extraction from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping Mustafa Turker a,, Dilek Koc-San b a Hacettepe University, Department of Geomatics Engineering, 06800 Cankaya-Ankara, Turkey b Department of Space Sciences and Technologies, Akdeniz University, 07058 Antalya, Turkey a r t i c l e i n f o Article history: Received 15 February 2014 Accepted 19 June 2014 Keywords: Building extraction SVM classification Hough transformation Perceptual grouping High-resolution imagery a b s t r a c t This paper presents an integrated approach for the automatic extraction of rectangular- and circular- shape buildings from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping. The building patches are detected from the image using the binary SVM classification. The generated normalized digital surface model (nDSM) and the normalized difference vegetation index (NDVI) are incorporated in the classi- fication process as additional bands. After detecting the building patches, the building boundaries are extracted through sequential processing of edge detection, Hough transformation and perceptual group- ing. Those areas that are classified as building are masked and further processing operations are performed on the masked areas only. The edges of the buildings are detected through an edge detection algorithm that generates a binary edge image of the building patches. These edges are then converted into vec- tor form through Hough transform and the buildings are constructed by means of perceptual grouping. To validate the developed method, experiments were conducted on pan-sharpened and panchromatic Ikonos imagery, covering the selected test areas in Batikent district of Ankara, Turkey. For the test areas that contain industrial buildings, the average building detection percentage (BDP) and quality percent- age (QP) values were computed to be 93.45% and 79.51%, respectively. For the test areas that contain residential rectangular-shape buildings, the average BDP and QP values were computed to be 95.34% and 79.05%, respectively. For the test areas that contain residential circular-shape buildings, the average BDP and QP values were found to be 78.74% and 66.81%, respectively. © 2014 Elsevier B.V. All rights reserved. Introduction Automatic urban-building extraction from space imagery is a challenging problem. Building boundary information is needed for a variety of applications, such as geographic information systems (GIS) database updating, cartography, urban monitoring, 3D city modeling, disaster management and land use analysis. Nowadays, the commercial high-resolution satellite images with multispec- tral bands provide a potential for the extraction of buildings in urban areas. Building extraction from space imagery has been car- ried out manually for decades. However, manual object extraction is slow, requires qualified operators and therefore is a costly task. Corresponding author. Tel.: +90 312 2976990; fax: +90 312 2976167. E-mail addresses: [email protected], [email protected] (M. Turker), [email protected] (D. Koc-San). Thus, automatic extraction of buildings is becoming of increasing practical importance. Automatic building extraction from high-resolution space imagery has been addressed by many researchers. The approach presented by Segl and Kaufmann (2001) for the detection of small objects in high-resolution satellite imagery is based on the shape characteristics. Benediktsson et al. (2003) proposed to use of morphological transformations for the classification and fea- ture extraction from the Indian Remote Sensing 1C (IRS-1C) and Ikonos satellite images of urban areas. Lee et al. (2003) used a class-guided approach to extract buildings from Ikonos images. After obtaining approximate location and shapes of potential build- ings, precise delineation was carried out in the panchromatic image using segmentation and squaring. The attempt of Tupin and Roux (2003) for building detection was to the simultaneous use of synthetic aperture radar (SAR) and optical images. The method developed by Haverkamp (2004) is based on edge maps http://dx.doi.org/10.1016/j.jag.2014.06.016 0303-2434/© 2014 Elsevier B.V. All rights reserved.

Upload: dilek

Post on 31-Jan-2017

214 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: Building extraction from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping

BuH

Ma

b

a

ARA

KBSHPH

I

ca(mtturi

(

h0

International Journal of Applied Earth Observation and Geoinformation 34 (2015) 58–69

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation andGeoinformation

jo ur nal home p age: www.elsev ier .com/ locate / jag

uilding extraction from high-resolution optical spaceborne imagessing the integration of support vector machine (SVM) classification,ough transformation and perceptual grouping

ustafa Turkera,∗, Dilek Koc-Sanb

Hacettepe University, Department of Geomatics Engineering, 06800 Cankaya-Ankara, TurkeyDepartment of Space Sciences and Technologies, Akdeniz University, 07058 Antalya, Turkey

r t i c l e i n f o

rticle history:eceived 15 February 2014ccepted 19 June 2014

eywords:uilding extractionVM classificationough transformationerceptual groupingigh-resolution imagery

a b s t r a c t

This paper presents an integrated approach for the automatic extraction of rectangular- and circular-shape buildings from high-resolution optical spaceborne images using the integration of support vectormachine (SVM) classification, Hough transformation and perceptual grouping. The building patches aredetected from the image using the binary SVM classification. The generated normalized digital surfacemodel (nDSM) and the normalized difference vegetation index (NDVI) are incorporated in the classi-fication process as additional bands. After detecting the building patches, the building boundaries areextracted through sequential processing of edge detection, Hough transformation and perceptual group-ing. Those areas that are classified as building are masked and further processing operations are performedon the masked areas only. The edges of the buildings are detected through an edge detection algorithmthat generates a binary edge image of the building patches. These edges are then converted into vec-tor form through Hough transform and the buildings are constructed by means of perceptual grouping.To validate the developed method, experiments were conducted on pan-sharpened and panchromaticIkonos imagery, covering the selected test areas in Batikent district of Ankara, Turkey. For the test areas

that contain industrial buildings, the average building detection percentage (BDP) and quality percent-age (QP) values were computed to be 93.45% and 79.51%, respectively. For the test areas that containresidential rectangular-shape buildings, the average BDP and QP values were computed to be 95.34% and79.05%, respectively. For the test areas that contain residential circular-shape buildings, the average BDPand QP values were found to be 78.74% and 66.81%, respectively.

© 2014 Elsevier B.V. All rights reserved.

ntroduction

Automatic urban-building extraction from space imagery is ahallenging problem. Building boundary information is needed for

variety of applications, such as geographic information systemsGIS) database updating, cartography, urban monitoring, 3D city

odeling, disaster management and land use analysis. Nowadays,he commercial high-resolution satellite images with multispec-ral bands provide a potential for the extraction of buildings in

rban areas. Building extraction from space imagery has been car-ied out manually for decades. However, manual object extractions slow, requires qualified operators and therefore is a costly task.

∗ Corresponding author. Tel.: +90 312 2976990; fax: +90 312 2976167.E-mail addresses: [email protected], [email protected]

M. Turker), [email protected] (D. Koc-San).

ttp://dx.doi.org/10.1016/j.jag.2014.06.016303-2434/© 2014 Elsevier B.V. All rights reserved.

Thus, automatic extraction of buildings is becoming of increasingpractical importance.

Automatic building extraction from high-resolution spaceimagery has been addressed by many researchers. The approachpresented by Segl and Kaufmann (2001) for the detection ofsmall objects in high-resolution satellite imagery is based on theshape characteristics. Benediktsson et al. (2003) proposed to useof morphological transformations for the classification and fea-ture extraction from the Indian Remote Sensing 1C (IRS-1C) andIkonos satellite images of urban areas. Lee et al. (2003) used aclass-guided approach to extract buildings from Ikonos images.After obtaining approximate location and shapes of potential build-ings, precise delineation was carried out in the panchromatic

image using segmentation and squaring. The attempt of Tupinand Roux (2003) for building detection was to the simultaneoususe of synthetic aperture radar (SAR) and optical images. Themethod developed by Haverkamp (2004) is based on edge maps
Page 2: Building extraction from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping

d Earth Observation and Geoinformation 34 (2015) 58–69 59

dwicamsmsrssudrassÜifbsh(mi(iams

ff(elntoaeHbfJfd(aeer

bpsbsagaotfi

M. Turker, D. Koc-San / International Journal of Applie

erived from the Ikonos panchromatic image. The edge pixelsere chained together to obtain the individual sides of build-

ngs and to reconstruct the buildings the logical grouping of thesehains were found. Jin and Davis (2005) presented an integratedpproach that employs structural, contextual and spectral infor-ation for automatic extraction of buildings in high-resolution

atellite imagery. Ünsalan and Boyer (2005) introduced an auto-atic system for street network and house detection from Ikonos

atellite images. Kim et al. (2006) proposed a semi-automatic algo-ithm to extract building lines from monoscopic high-resolutionpace images. They focused on extracting lines from rectangular-hape buildings of a relatively large size. Sohn and Dowman (2007)tilized a data-driven approach on the optical imagery and a model-riven approach on airborne laser scanning data for extractingectilinear lines of buildings. The approach proposed by Agüerand Liu (2009) for delineating greenhouses from high-resolutionatellites QuickBird and Ikonos comprises the steps of image clas-ification, automatic vectorization, and delineation. Sırmac ek andnsalan (2009) detected urban areas and buildings in satellite

mages through an approach that uses scale invariant feature trans-orm (SIFT) and graph theoretical tools. The approach proposedy Lhomme et al. (2009) for building extraction using very-high-patial-resolution images is based on discriminating features thatave specific size and shape within pixel clusters. Huang and Zhang2011) presented a morphological building index (MBI) for auto-

atic building extraction from high-resolution remote sensingmagery. In a recent study conducted by Koc-San and Turker2012) a model-based approach was presented for automatic build-ng database updating from high-resolution space imagery. Theirpproach utilizes an existing building database as a building-odel-library for detecting the buildings from high resolution

pace imagery.In the literature, several research studies based on Hough trans-

ormation have been reported for the mapping of geological lineareatures (Karnieli et al., 1996), for the mapping of agricultural plotsRuiz et al., 2011), for the detection of crop orientation (Chanussott al., 2005), for the discrimination of the farmland types grass-and and cropland (Helmholz et al., 2007), for the detection ofatural gas seepages (Van der Werff et al., 2006), for the detec-ion of greenhouses (Agüera and Liu, 2009), and for the detectionf buildings (Lee et al., 2003; Croitoru and Doytsher, 2004; Jungnd Schramm, 2004; Cui et al., 2012; Grigillo et al., 2012). Leet al. (2003) proposed a building squaring approach based on theough transformation for the detection of the rectilinear buildingoundaries. Croitoru and Doytsher (2004) employed a Hough trans-ormation to extract straight lines to express building structures.ung and Schramm (2004) proposed a windowed Hough trans-ormation to detect rectangles and applied it on aerial images toetect rectangular buildings. The approach proposed by Cui et al.2012) combines the robustness of the Hough transformation with

graph search algorithm for complex building description andxtraction from high-resolution remotely sensed imagery. Grigillot al. (2012) applied the Hough transformation simultaneously withegion growing for the extraction of buildings of suburban areas.

In this study, we present an integrated approach to automaticuilding extraction from high-resolution spaceborne imagery. Thearticular attention was the extraction of rectangular- and circular-hape (circle, ring, C and S) buildings. Our extraction method isased on the integration of support vector machine (SVM) clas-ification, Hough transformation, and perceptual grouping. Thedded value of the integration of these approaches is that the inte-rated strategies offer a better solution compared with individual

pproaches as the integration of complementary methods helps tovercome the drawbacks of the individual methods. First, we detecthe building patches from the image using a binary SVM classi-cation. To increase the reliability of classification the nDSM and

Fig. 1. The schematic workflow of the overall process of the developed buildingextraction method.

NDVI are used as additional bands during image classification. Then,to extract boundaries of the buildings, we sequentially performedge detection, Hough transformation, and perceptual grouping.To implement the approach we developed a program in MATLAB®

v. 7.1 programming environment. We tested our approach on apan-sharpened and panchromatic Ikonos imagery, covering theselected test areas in Batikent district of Ankara, Turkey. The testareas contain residential and industrial buildings with differentshapes, sizes, and usage, as well as various urban-area character-istics. Tests demonstrate the potential of our developed automaticurban-building extraction method. The novelty of this work is thatwe exploit the combined processing and analysis of SVM classi-fication, Hough transform, and perceptual grouping in buildingextraction from high resolution spaceborne images. We detect thebuilding patches as the regions of interest areas (ROIs) using thebinary SVM classification. By this way, the irrelevant objects areexcluded from further processing operations allowing us to focuson the building areas only. With the exclusion of the unnecessaryedges form the processing operations the reliability of buildingboundary extraction is increased. Further, we propose a circularHough transformation-based approach for the extraction of C- andS-shape buildings.

The remainder of this paper is structured in four parts. First, wedescribe the proposed method, which comprises the main stepsof building detection and building boundary delineation. Then, weprovide the experimental tests. In this section we describe the studyarea, the data used, and the data processing and performance evalu-ation. We then provide the results obtained from the experimentaltests. The last section comprises the conclusions drawn from thepresented method.

Method

The schematic workflow of the whole process of the proposedbuilding extraction method is shown in Fig. 1. The method requiresa high-resolution optical spaceborne image with multi-spectralbands, digital surface model (DSM) and digital terrain model (DTM)

Page 3: Building extraction from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping

6 ed Ear

dboe(setefo

B

SitostitTimebto(2

B

aibvbt

D

ufCtHytygTi�iob

H

p

0 M. Turker, D. Koc-San / International Journal of Appli

ata. The image is orthorectified using the generated DSM. Theuilding patches are detected through binary SVM classificationf the image. During classification, the derived normalized differ-nce vegetation index (NDVI) and normalized digital surface modelnDSM) are used as additional bands. The nDSM is calculated byubtracting DTM from DSM. To extract building boundaries, thedges of the building patches are detected using an edge detec-ion algorithm that generates a binary edge image. The detecteddges are then converted into vector form through Hough trans-orm. The buildings are constructed through perceptual groupingf the detected lines.

uilding detection

To detect building patches, the image is classified using binaryVM classification. The data sets NDVI and nDSM are incorporatedn the classification process as additional bands during classifica-ion. SVM is a non-parametric classification technique that is basedn statistical learning theory (Vapnik, 1995). Initially, the trainingamples are collected for each class to represent the feature vec-ors. Then, an optimal separating hyperplane between the classess identified by concentrating on the support vectors, which arehe training samples located at the edge of the class distributions.he position of the hyperplane is optimized during the SVM learn-ng process to separate the data having maximum margin and

inimum misclassification. If the classes in the dataset are lin-arly inseparable, a kernel function is used to solve the problemy transforming the feature vectors into a higher-dimensional fea-ure space. The detailed information and mathematical formulationf the SVM classification are provided in the literature extensivelyFoody and Mathur, 2004; Huang et al., 2002; Vapnik, 1995; Wang,005).

uilding delineation

To delineate building boundaries, the detected building patchesre processed one patch at a time. To do that each building patchs labeled using a label matrix operation, in which the pixels thatelong to patch being processed are assigned a unique integeralue, while the background pixels are set the value of 0. Next, theoundaries of each building patch are delineated through Houghransform and the developed boundary tracing based operations.

elineation of rectangular-shape buildingsThe general steps include smoothing the building patch image

sing the Gaussian smoothing filter, edge detection, Hough trans-ormation, and perceptual grouping. To detect edges we use theanny Edge Detection algorithm (Canny, 1986). Then, we extracthe line segments through Hough transformation. The basis of theough transformation is given in equation (1), where (�, �) and (x,) respectively refer to Hough and image domains, while, ı referso Dirac delta function. Each point (x, y) in the original image F(x,) is transformed into a sinusoid � = x cos(�) − y sin(�) and H(�, �)ives the total number of sinusoids that intersect at point (�, �).herefore, it gives the total number of points making up the linen the original image. By choosing a line-cut threshold T for H(�,) and using the inverse Hough transformation, the original image

s filtered so that the lines that contain at least T points are keptnly. A detailed description about the Hough transformation cane found in Hough (1962) and Duda and Hart (1972).∫ ∞ ∫ ∞

(�, �) =

−∞ −∞F(x, y)ı(� − x cos(�) − y sin(�))dxdy (1)

After extracting the line segments, we group them based onerceptual grouping that can be described as the capability to

th Observation and Geoinformation 34 (2015) 58–69

impose structural organization on sensory data, so as to groupsensory primitives arising from a common underlying cause.Gestalt-psychologists described the principles of perceptual group-ing (Gestalt is German for “pattern” or “shape”) nearly a hundredyears ago (Koffka, 1935; Köhler, 1929; Wertheimer, 1923). A set ofprinciples were suggested as significant in the perceptual group-ing of images in our real world by the Gestalt-psychologists. Theseprinciples are proximity, continuity, similarity, closure, symmetryand the new properties (Rock and Palmer, 1990) of common regionand connectedness. In this study, we concentrated on the relations’proximity, continuation, symmetry and closure.

The steps of the developed grouping procedure are shown inFig. 2. We compute the centroid (C) of the building patch beingconsidered. The collinear Hough lines are merged based on a con-dition that the distance between the line segments are shorter thanthe predefined threshold value. We assume that the dominant linerepresents a true edge of the building and therefore, we take thelongest Hough line as the first dominant line. In Fig. 2(b), label D1represents the detected first dominant line. After detecting the firstdominant line in one direction, the second dominant line is selectedfrom those line segments that are perpendicular (90◦ ± 10◦) to firstdominant line. In Fig. 2(c), label D2 represents the second domi-nant line that is perpendicular to first dominant line. The detectionof these two lines is important as they represent the adjacent longand short edges of the building to be extracted. Next, the point ofintersection (I1) of the detected dominant lines is calculated andthey are extended or shortened along their directions to snap themto the intersection point (Fig. 2(d)). This point of intersection rep-resents one of the corners of the building. In order to delineate theboundaries of the building we need to construct the second cor-ner in the opposite direction of the first corner. To achieve this,the distance between the intersection point and the centroid iscalculated and the second corner (I2) is located in the oppositedirection with the same distance from the centroid (Fig. 2(e)). Afterthat, passing through the second corner (I2), two new lines, oneis parallel to first dominant line (D1) and the other is parallel tosecond dominant line (D2), are drawn. Finally, the building bound-aries are constructed through integrating the detected four lines(Fig. 2(f)).

Delineation of circular-shape buildingsThe Hough transformation was originally designed to detect

straight lines. Later it was extended to detect curves (Ballard, 1981;Davies, 1988; Duda and Hart, 1972; Ioannou et al., 1999; Parrot andTaud, 1992), ellipses (Yip et al., 1992) and arbitrarily-shape objects(Pao et al., 1992). In this study, we employ Circular Hough transfor-mation to delineate circular-shape (circle- ring- C- and S-shapes)buildings. Circular Hough transformation is similar to the Houghtransformation for lines, however it utilizes the parametric formfor a circle. Each edge point (x, y) generates a circle in a 3D parame-ter space with radius r and the circle can be expressed with Eq. (2),where a and b are the x and y coordinates of the center of the circle.

(x − a)2 + (y − b)2 = r2 (2)

A circle with radius r and center (a, b) can be described with theparametric Eqs. (3) and (4).

x = a + r cos(�) (3)

y = b + r sin(�) (4)

When the angle � sweeps through the full range of 360◦, thepoints (x, y) trace the perimeter of a circle. To find the circles using

Circular Hough transformation, a circle with the desired radius isdrawn for each edge point in the parameter space. The accumulatorarray is incremented for the coordinates that belong to the perime-ter of the drawn circle. This process is performed for all edges. At the
Page 4: Building extraction from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping

M. Turker, D. Koc-San / International Journal of Applied Earth Observation and Geoinformation 34 (2015) 58–69 61

ped pe

ecr1

absTmeccbc

aHabTetat

Fig. 2. The steps of the develo

nd of this process the highest number in the accumulator spaceorresponds to the center of the circle in the image space. The algo-ithm for the Circular Hough transformation is as follows (Cross,988; Parrot and Taud, 1992; Taud and Parrot, 1992).

Find the edges on the candidate building patch image

For each edge point

Draw a circle with radius r which have a center in the edgepoint

Increment accumulator array for coordinates that the perime-ter of circle passes through

Find the local maxima

Map the found parameters (r, a, b) back to original image

The parameters that we used include segment no, radius rangend tolerance value. The parameter segment no refers to the num-er of segments (points) on the detected circle. As the number ofegments increases the circle to be generated becomes smoother.he parameter radius range is used to determine the minimum andaximum radius range values of the circular buildings. The param-

ter tolerance value is used to determine the tolerance value for theoncentric circles. The ring-, C- and S-shape buildings may containoncentric circles/semi-circles. Therefore, a tolerance value muste determined to detect the multiple circles that have the sameenter point.

To delineate circle- and ring-shape buildings, the sequentialpplications of the aforementioned edge detection and Circularough transformation would be enough. The delineation of C-nd S-shape buildings requires further processing operations toe carried out on the results of Circular Hough transformation.he general steps of the developed C- and S-shape building delin-

ation procedures include (a) detecting the concentric semi-circleshrough processing the building patch pixels, (b) finding the startnd end points of these circles, (c) grouping the points that belongo semi-circles, and (d) delineating the building boundaries.

rceptual grouping procedure.

We demonstrate the developed approach in Fig. 3. In Fig. 3(a),the dark gray pixels represent a building patch on which theCircular Hough points of the concentric outer and inner circles areoverlaid. Both the inner and outer circles contain 24 points, whichmeans that the parameter segment no is set to 24. The inner andouter boundaries of a C-shape building resemble concentric twosemi-circles. Our effort is to detect these circles from the buildingpatch pixels that correspond to circle points. Therefore, those Circu-lar Hough points that do not overlap with the building patch pixelsare deleted and excluded from further processing operations. Thedeleted non-building Circular Hough points that belong to outerand inner circles are shown in light color in Fig. 3(b). Next, the startand end points of the concentric semi-circles are detected. To doso, for the outer and inner semi-circles, the distances (d) betweenthe successive Hough points are calculated and the start and endpoints are determined on a condition that the distance betweentwo successive points is the greatest (Fig. 3(c)). To illustrate, d12represents the greatest distance between two successive Houghpoints of the outer semi-circle (Fig. 3(c)). Therefore, the end pointsof d12 are taken as the start and end points of the outer semi-circle.In Fig. 3(d), the points SOC and EOC, respectively represent thedetected start and end points of the outer semi-circle, while thepoints SIC and EIC, respectively represent the detected start andend points of the inner semi-circle. Next, for each semi-circle, theconsecutive Hough points are connected starting from the startpoint and finishing at the end point. In Fig. 3(d), the connectedHough points between SOC and EOC represent the delineated outersemi-circle, while the connected points between SIC and EIC rep-resent the delineated inner semi-circle. The building delineationprocedure is completed by connecting the start point of the outersemi-circle (SOC) to the start point of the inner semi-circle (SIC) andconnecting the end point of the outer semi-circle (EOC) to the endpoint of the inner semi-circle (EIC). The final delineated boundary

of the C-shape building is shown in Fig. 3(f).

The detection of S-shape buildings also requires furtherprocessing operations to be carried out on the output of the Cir-cular Hough transformation. The steps include (a) detecting the

Page 5: Building extraction from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping

62 M. Turker, D. Koc-San / International Journal of Applied Earth Observation and Geoinformation 34 (2015) 58–69

oped

codsctdp

spcetttafit1oacaNscatfTtsc

Fig. 3. The schematic demonstration of the devel

oncentric semi-circles that correspond to double curves of S shapef the building patch pixels, (b) deleting those Hough points thato not overlap with the building patch pixels, (c) identifying thetart and end points of all concentric semi-circles, (d) detecting thelosest points of the outer and inner semi-circles with different cen-roids and merging these points to make a connection between theouble curves of the S shape, and (e) grouping the correspondingoints and delineating the boundaries.

We demonstrate the developed approach in Fig. 4. Fig. 4(a)hows the patch image of a S-shape building with the superim-osed Circular Hough points of the outer and inner circles. As itan be seen in Fig. 4(a), the Circular Hough transformation gen-rates two concentric circles for each of the double curves ofhe S shape. After performing the Circular Hough transformation,he semi-circles are generated by deleting those Hough pointshat do not overlap with the building patch pixels, and the startnd end points of these semi-circles (S1, S2, E1, E2) are identi-ed as previously described. In Fig. 4(c), S1IC and E1IC representhe start and end points of the inner semi-circle with centroid

(+1), while S1OC and E1OC represent the start and end pointsf the outer semi-circle with the same centroid. Similarly, S2ICnd E2IC represent the start and end points of the inner semi-ircle with centroid 2 (+2), while S2OC and E2OC represent the startnd end points of the outer semi-circle with the same centroid.ext, the closest points of the inner semi-circle 1 and the outer

emi-circle 2 are detected. The closest points of the inner semi-ircle 2 and the outer semi-circle 1 are also detected. These pointsre shown in green color in Fig. 4(d). Then, for each semi-circlehose points that fall inside the building patch are deleted startingrom the detected closest points and going in clockwise direction.

he deleted points are shown in pink color in Fig. 4(d). Finally,he boundaries of the S-shape building are constructed throughequentially grouping the remaining Hough points of the semi-ircles (Fig. 4(e)).

approach for the extraction of C-shape buildings.

Experimental setup

Study area and data used

The approach was tested in Batikent district of Ankara, Turkey.Batikent is a planned and regularly developed settlement areawhere the buildings are with different shapes, sizes, and usage,such as the residential and industrial facilities. Most of the resi-dential buildings are rectangular in shape, while several exist incircular shape. The roofs of most of the buildings are in brick color,while several exist in gray and white colors. The residential build-ings are low-rise (one or two storeys), middle-rise (three-to-fivestoreys) and high-rise (more than five storeys). The area also con-tains two industrial zones, where buildings are rectangular in shapewith the gray-, white-, and blue-color roofs. Almost all industrialbuildings are two or three storeys and their sizes are much largerwhen compared to sizes of the residential buildings.

The data used include Ikonos pan-sharpened (PS) and stereopanchromatic images and 1:1000-scale existing digital vectordataset that contains contour lines and 3D points. The Ikonosimages acquired on August 4, 2002 were along track stereo imagesin “Geo” data format. The 1:1000-scale existing digital vectordataset was used to assess the results.

Data processing and performance evaluation

A Digital Terrain Model (DTM) was generated from the existing1:1000-scale digital vector dataset. A DSM was generated from theIkonos stereo panchromatic image pairs. To generate DSM we used48 ground control points (GCP) which were measured on the site

using differential global positioning system (DGPS) technique. Ofthese points, 24 were used as GCPs and the remaining 24 were setaside as independent check points (ICPs). The accuracy of the gen-erated DSM was computed to be 0.7 m (Koc-San, 2009). The nDSM
Page 6: Building extraction from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping

M. Turker, D. Koc-San / International Journal of Applied Earth Observation and Geoinformation 34 (2015) 58–69 63

oped

woo

ootreSsttTnpbtlm

Fig. 4. The schematic demonstration of the devel

as calculated by subtracting DTM from DSM and the above groundbjects were separated by applying a 3-m threshold to nDSM basedn the assumption that the buildings are higher than 3 m.

The Ikonos imagery that we used has a nominal collection anglef 81.16643◦, which is near to nadir. However, to remove the effectf possible relief displacement the Ikonos PS image was orthorec-ified using the computed DSM. The NDVI was calculated using theed and near infra red (NIR) bands of the Ikonos PS image. The gen-rated nDSM and NDVI layers were used as additional bands inVM classification of the Ikonos PS image. To perform SVM clas-ification the training samples were collected from two classeshat are building and non-building. The class non-building includeshe subclasses vegetation, road, bare land, shadow and pavement.he training data set contained 1000 pixels (500 building and 500on-building), and the validation of the classification outputs waserformed using 4000 test pixels (2000 building and 2000 non-uilding). To represent the impartial reference information, the

est pixels and the training pixels were collected from differentocations. For performing the binary SVM classification, the deter-

ination of the penalty parameter (C) value, the selection of the

approach for the extraction of S-shape buildings.

kernel method and the parameters related to the selected kernelmethod are also important. In this study, the value for the penaltyparameter (C) was taken to be 1000. This parameter is a form ofregularization parameter that defines the trade-off between thenumber of noisy training samples and classifier complexity. As thekernel method, the radial basis function (RBF) was selected. Theparameter � (gamma term) was determined as the inverse of thenumber of bands in the input image. A detailed description aboutthe SVM classification that we carried out is given in a previousstudy conducted by the authors (Koc-San and Turker, 2012). Toremove artifacts we applied the opening and closing morphologicaloperations with an isotropic structuring element on the classifiedimage. We determined the threshold value as 50 pixels (50 m2 forthe Ikonos PS image) based on the assumption that patches smallerthan 50 m2 do not represent buildings. Next, the detected edgeswere converted into vector form using the above described Houghtransformation procedure. For the circular Hough transformation,

the number of segments was taken to be 32. Finally, the build-ings were constructed through the developed perceptual groupingalgorithms.
Page 7: Building extraction from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping

64 M. Turker, D. Koc-San / International Journal of Applied Earth Observation and Geoinformation 34 (2015) 58–69

Table 1The quantity assessment results of the extracted buildings.

Building type Reference building (RB) Extracted building (EB) EB/RB (%)

Industrial rectangular 77 76 98.70Residential rectangular Detached 107 105 98.13

Semi-detached 77 74 96.10Terraced 123 122 99.19

Residential circular 10 10 100.00

Total 394 387 98.22

Table 2The quality assessment results of the extracted buildings in urban blocks. (BF: branching factor, MF: miss factor, BDP: building detection percentage, QP: quality percentage).

Building type Urban block no. BF MF BDP QP

Industrial rectangular 1 0.14 0.06 94.17 82.882 0.29 0.09 91.83 72.67

Average 0.19 0.07 93.45 79.51

Residential rectangular

Detached 1 0.20 0.08 92.44 77.972 0.24 0.04 96.38 78.443 0.13 0.07 93.11 83.324 0.29 0.03 97.22 75.705 0.18 0.04 95.73 81.676 0.40 0.03 96.79 69.70

Semi-detached 1 0.14 0.06 94.57 83.612 0.29 0.05 95.50 74.573 0.16 0.05 95.28 82.484 0.15 0.03 97.36 85.115 0.24 0.02 97.90 79.48

Terraced 1 0.29 0.02 98.26 76.432 0.29 0.04 96.46 75.163 0.15 0.13 88.39 77.754 0.21 0.04 96.28 80.055 0.28 0.04 96.50 75.996 0.17 0.03 97.09 83.237 0.20 0.04 95.73 80.16

Average 0.22 0.05 95.34 79.05

Residential circular 1 0.55 0.11 90.37 60.502 0.05 0.12 89.14 85.483 0.19 1.47 68.50 61.004 0.15 0.35 73.88 66.52

Average 0.23 0.27 78.74 66.81

Fig. 5. The extracted building boundaries for urban blocks that contain the industrial-rectangular buildings.

Page 8: Building extraction from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping

M. Turker, D. Koc-San / International Journal of Applied Earth Observation and Geoinformation 34 (2015) 58–69 65

Fig. 6. The extracted building boundaries for urban blocks that contain the detached residential-rectangular buildings.

Fig. 7. The extracted building boundaries for the semi-detached residential-rectangular buildings.

Page 9: Building extraction from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping

66 M. Turker, D. Koc-San / International Journal of Applied Earth Observation and Geoinformation 34 (2015) 58–69

or the

(coM

R

atsbawbT9ra6rte

Fig. 8. The extracted building boundaries f

The statistical measures branching factor (BF), miss factorMF), building detection percentage (BDP), and quality per-entage (QP) were computed to evaluate the performancef the developed building extraction procedure (Shufelt andcKeown, 1993).

esults

The results were assessed using the reference building bound-ries. Based on a visual evaluation we can make a general statementhat most of the buildings were delineated successfully. Table 1ummarizes the assessment results of the developed automaticuilding extraction approach in this study. As it can be seen, thepproach correctly extracts 387 buildings out of the total 394ith the computed overall accuracy of 98.22%. For each urban

lock, the computed BF, MF, BDP and QP values are presented inable 2. The average BDP values were computed to be 93.45%,5.34%, and 78.74% for the industrial-rectangular, residential-ectangular, and residential-circular buildings, respectively. Theverage QP values were computed to be 79.51%, 79.05%, and

6.81% for the industrial-rectangular, residential-rectangular, andesidential-circular buildings, respectively. The results illustratehat our approach is quite successful for the automatic buildingxtraction from high resolution space imagery.

Fig. 9. The extracted boundaries for the r

terraced residential-rectangular buildings.

The extracted building boundaries for urban blocks that con-tain the industrial-rectangular buildings are shown in Fig. 5. Asit can be seen in Fig. 5, the boundaries of most of the buildingswere correctly extracted. Of the total of 77 buildings that fall withinthis test area, 76 were delineated successfully with the computedaccuracy of 98.70%. In Block #1 (Fig. 5(a)), two closely located build-ings, which are shown in a yellow circle in the upper left of theimage, were erroneously delineated as a joined single building. It isevident that the computed BDP values (above 91.83%) and QP val-ues (above 72.67%) were quite high for the industrial-rectangularbuilding category (Table 2).

The extracted building boundaries for six urban blocks that con-tain detached residential-rectangular buildings are shown in Fig. 6.Of the urban blocks, Block #1 (Fig. 6(a)) contains two storey low-rise buildings, Block #6 (Fig. 6(f)) contains eight storey high-risebuildings, and the other blocks (Blocks #2, #3, #4, and #5) con-tain four and five storey middle-rise buildings. As it can be seenin Fig. 6(a)–(f), the boundaries of most of the buildings were cor-rectly extracted. Of the total 107 buildings that fall within this testarea, 105 were delineated correctly with the computed accuracy of

98.13% (Table 1). However, the failure of the approach is evident fortwo buildings that are shown in yellow circles in Fig. 6(c) and (f).Obviously, these buildings were missed. However, we found thatthese buildings were quite small. For all urban blocks, the computed

esidential circular-shape buildings.

Page 10: Building extraction from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping

M. Turker, D. Koc-San / International Journal of Applied Earth Observation and Geoinformation 34 (2015) 58–69 67

fail in

Bt8

rtmAapstToaravc

Fig. 10. The illustration of the cases for which our approach may

DP values were quite high staying above 92.44%. Whereas,he QP values were relatively low staying between 69.70% and3.32%.

The extracted boundaries for the semi-detached residential-ectangular buildings are shown in Fig. 7. Of the five urban blockshat fall within this category, Blocks #1, #2, #3, and #4 contain

iddle-rise buildings, while Block #5 contains high-rise buildings.s it can be seen in Fig. 7, our approach correctly extracted almostll buildings. Of the total 77 buildings, 74 were correctly extractedroviding a high success rate of 96.10% (Table 1). However, threemall buildings, one is in the upper right of Block #2 (Fig. 7(b)) andhe others are in the upper left of Block #4, (Fig. 7(d)), were missed.hese buildings are shown in yellow circles in Fig. 7(b) and (d). Webserved that overlap exists between the spectral reflectance char-cteristics of these buildings and the surrounding pavements and

oads. The quantitative results given in Table 2 illustrate that ourpproach is also quite successful (BDP values above 94.57% and QPalues above 74.57%) for the extraction of buildings that fall in thisategory.

detecting building patches and extracting building boundaries.

The extracted boundaries for the terraced residential-rectangular buildings are shown in Fig. 8. Our approach wasvery successful as it correctly extracted 122 buildings out of thetotal 123, providing a very high accuracy rate. One failure of theapproach is evident in the extraction of a semi-occluded smallbuilding located in the central part of Block #7 (Fig. 8(g)). Further,in Block #1 two buildings that are located closely but are notadjoined were also erroneously delineated as a joint building(Fig. 8(a)). A terraced building in Block #4 that is shown in yellowcircle in Fig. 8(d) was extracted as two separate buildings due to thevariation of the spectral reflectance on the roof of the building. Thequantitative results presented in Table 2 show that our approachis also very successful for the extraction of the buildings that fallin this category.

The extracted boundaries for the residential circular-shape

buildings are shown in Fig. 9. As it can be seen, our approach isable to delineate the circular-shape buildings, even the complexones. It is evident that the building located in the lower part ofthe image was extracted quite successfully (Fig. 9(a)). Whereas,
Page 11: Building extraction from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping

6 ed Ear

aaiiaitcosaeslcr

asswwboobtbettaOoitcadei(inmbmti

C

erpiousbatdl

8 M. Turker, D. Koc-San / International Journal of Appli

noticeable deviation is evident between the extracted and thectual boundaries of the building located in the upper part of themage. However, the spectral reflectance of this building is sim-lar to the spectral reflectance of the surrounding concrete floor,nd we believe that the spectral similarity might have caused thencorrect detection of the building patch for this building. Evenhe use of nDSM did not help extract the upper side of this cir-ular building in a correct manner. As it can be seen in Fig. 9(b),ur approach was quite successful for the extraction of the ring-hape building. Similarly, the C- and S-shape buildings in Fig. 9(c)nd (d) were also extracted quite successfully. Although the gen-ral forms of these buildings are C- and S-shape, they have angularides. Due to the limitations of the developed approach, the angu-ar sides of these buildings were not able to be delineated as in realase and therefore, the extracted boundaries do not fully match theeal boundaries.

In several cases our approach may fail to detect building patchesnd extract building boundaries (Fig. 10). In the first case, themall buildings may not be extracted correctly (Fig. 10(a)). In thistudy, we made an assumption that the size of building patchesere larger than 50 m2. Therefore, the small building in Fig. 10(a)as excluded from further processing as its patch size stayed

elow the defined threshold value of 50 m2. In the second case,ur approach fails to extract those buildings that have semi-ccluded roofs. As it can be seen in Fig. 10(b), the roofs of twouildings in the middle part of the image were occluded by therees and therefore, our approach did not correctly extract theiroundaries. We can state that our approach is quite successful inxtracting the semi-occluded buildings substantially. However, ifhe building is small and a significant portion of it is occluded,hen our approach may fail. In the third case (Fig. 10(c)), ourpproach may also fail to extract those buildings located closely.f course, this case is also related with the spatial resolutionf the image used, which imposes restrictions on the separabil-ty of closely located buildings. In the fourth case (Fig. 10(d)),he buildings that have reflectance characteristics similar to otherlasses may not be extracted correctly. Although the developedpproach is quite successful to extract buildings that representifferent spectral reflectance characteristics, in certain cases thextraction process may fail. In the fifth case (Fig. 10(e)), the adjoin-ng buildings may not be extracted separately. In the sixth caseFig. 10(f)), the buildings that have reflectance characteristics sim-lar to reflectance characteristics of the surrounding area mayot be extracted correctly. In the last case (Fig. 10(g)), if theutual lines of the buildings are not parallel and/or the angles

etween the adjacent lines are not perpendicular, the buildingsay not be extracted correctly. This is due to the restrictions of

he approach and the assumptions made for perceptual group-ng.

onclusions

This study presents an integrated approach for the automaticxtraction of rectangular- and circular-shape buildings from high-esolution optical spaceborne images. To do so, first the buildingatches are detected as the regions of interest areas from the

mage through binary SVM classification. By this way, the irrelevantbjects are excluded from further processing operations allowings to focus on the building areas only. In other words, the unneces-ary edges are not processed and therefore the reliability of buildingoundary extraction is increased. During image classification nDSM

nd NDVI are used as additional bands. The building boundaries arehen extracted through the sequential processing of Canny edgeetection, Hough transformation, and perceptual grouping of the

ine segments that are detected through Hough transformation.

th Observation and Geoinformation 34 (2015) 58–69

The performance of the approach was tested on residential andindustrial areas using the Ikonos satellite imagery. After exten-sive testing, we observed that our approach is able to extract theboundaries of most of the buildings (having various shapes, sizesand spectral reflectance values) in a correct manner. For the testareas that cover industrial buildings, the average BDP value wascomputed to be 93.45%, while the QP value was computed to be79.51%. For the test areas that cover residential rectangular-shapebuildings, the average BDP and QP values were found to be 95.34%and 79.05%, respectively. For the test areas that cover residentialcircular-shape buildings, the average BDP and QP values were com-puted to be 78.74% and 66.81%, respectively.

Experimental testing has shown that if the adjacent lines of thebuildings are perpendicular to each other and the mutual linesare parallel, or if the buildings are circular- or curved-shape, ourapproach that integrates SVM classification, Hough transformation,and perceptual grouping can be applied in any area with simi-lar success. However, fails may still occur when the buildings areclosely located, the edges are too vague, the roofs are occludedor there is a significant difference between the size and shapeof the building patch detected and the building itself. Therefore,further research is needed to refine the current method. Futureimprovements in the method could involve the inclusion of theshape characteristics of the buildings and improving the percep-tual grouping rules that we used to construct building boundaries.It is also worthwhile evaluating the effectiveness of the approachfor building extraction from higher resolution space images.

Acknowledgement

This research was supported by theState Planning Organization(DPT) Grant: BAP-08-11-DPT2002K120510.

References

Agüera, F., Liu, J.G., 2009. Automatic greenhouse delineation from QuickBird andIkonos satellite images. Comput. Electron. Agric. 66, 191–200.

Ballard, D.H., 1981. Generalizing the Hough transform to detect arbitrary shapes.Pattern Recogn. 13, 111–122.

Benediktsson, J.A., Pesaresi, M., Arnason, K., 2003. Classification and feature extrac-tion for remote sensing images from urban areas based on morphologicaltransformations. IEEE Trans. Geosci. Remote Sens. 41, 1940–1949.

Canny, J., 1986. Computational approach to edge detection. IEEE Trans. Pattern Anal.Mach. Intell. 8, 679–698.

Chanussot, J., Bas, P., Bombrun, L., 2005. Airborne remote sensing of vineyards forthe detection of dead vine trees. In: Proceedings IEEE International Geoscience& Remote Sensing Symposium, Seoul, Korea, pp. 3090–3093.

Croitoru, A., Doytsher, Y., 2004. Right-angle rooftop polygon extraction in regular-ized urban areas: cutting the corners. Photogramm. Rec. 19, 311–341.

Cross, A.M., 1988. Detection of circular geological features using the Hough Trans-form. Int. J. Remote Sens. 9, 1519–1528.

Cui, S., Yan, Q., Reinartz, P., 2012. Complex building description and extraction basedon Hough transformation and cycle detection. Remote Sens. Lett. 3, 151–159.

Davies, E.R., 1988. A modified Hough scheme for general circle location. PatternRecogn. Lett. 7, 37–43.

Duda, R.O., Hart, P.E., 1972. Use of the Hough transformation to detect lines andcurves in pictures. Commun. Assoc. Comput. Mach. 15, 11–15.

Foody, G.M., Mathur, A., 2004. A relative evaluation of multiclass image classificationby support vector machines. IEEE Trans. Geosci. Remote Sens. 42, 1335–1343.

Grigillo, D., Fras, M.K., Petrovic, D., 2012. Automated building extraction fromIKONOS images in suburban areas. Int. J. Remote Sens. 33, 5149–5170.

Haverkamp, D., 2004. Automatic building extraction from Ikonos imagery. In:Proceedings of ASPRS 2004 Conference, May 23–28, Denver, 8 pp., on CD-ROM.

Helmholz, P., Gerke, M., Heipke, C., 2007. Automatic discrimination of farmlandtypes using Ikonos imagery. In: Proceedings Photogrammetric Image Analysis07, Munich, Germany, pp. 81–86.

Hough, P.V.C., 1962. Methods and means for recognizing complex patterns. U.S.Patent 3; 069; 654.

Huang, X., Zhang, L., 2011. A multidirectional and multiscale morphological indexfor automatic building extraction from multispectral GeoEye-1 imagery. Pho-

togramm. Eng. Remote Sens. 77, 721–732.

Huang, C., Davis, L.S., Townshend, J.R.G., 2002. An assessment of support vectormachines for land cover classification. Int. J. Remote Sens. 23, 725–749.

Ioannou, D., Huda, W., Laine, A.F., 1999. Circle recognition through a 2D Houghtransform and radius histogramming. Image Vis. Comput. 17, 15–26.

Page 12: Building extraction from high-resolution optical spaceborne images using the integration of support vector machine (SVM) classification, Hough transformation and perceptual grouping

d Ear

J

J

K

K

K

K

KKL

L

P

P

R

R

Berlin, Heidelberg.Wertheimer, M., 1923. Laws of organization in perceptual forms. First published

M. Turker, D. Koc-San / International Journal of Applie

in, X., Davis, C.H., 2005. Automated building extraction from high-resolutionsatellite imagery in urban areas using structural, contextual, and spectral infor-mation. EURASIP J. Appl. Signal Process. 14, 2196–2206.

ung, C.R., Schramm, R.,2004. Rectangle detection based on a windowed Houghtransform. In: Proceedings of the Computer Graphics and Image Processing,17–20 October 2004. IEEE Computer Society, Curitiba, pp. 113–120.

arnieli, A., Meisels, A., Fisher, L., Arkin, Y., 1996. Automatic extraction and evalua-tion of geological linear features from digital remote sensing data using a Houghtransform. Photogramm. Eng. Remote Sens. 62, 525–531.

im, T., Lee, T.Y., Kim, K.O., 2006. Semiautomatic building line extraction from Ikonosimages through monoscopic line analysis. Photogramm. Eng. Remote Sens. 72,541–549.

oc-San, D., (Unpublished Ph.D. thesis) 2009. Approaches for automatic urban build-ing extraction and updating from high resolution satellite imagery. Middle EastTechnical University, Turkey.

oc-San, D., Turker, M., 2012. A model-based approach for automatic buildingdatabase updating from high resolution space imagery. Int. J. Remote Sens. 33,4193–4218.

offka, K., 1935. Principles of Gestalt Psychology. Harcourt, New York.öhler, W., 1929. Gestalt Psychology. Liveright, New York.ee, D.S., Shan, J., Bethel, J.S., 2003. Class-guided building extraction from Ikonos

imagery. Photogramm. Eng. Remote Sens. 69, 143–150.homme, S., He, D.C., Weber, C., Morin, D., 2009. A new approach to building

identification from very-high-spatial-resolution images. Int. J. Remote Sens. 30,1341–1354.

ao, D.C.W., Li, H.F., Jayakumar, R., 1992. Shapes recognition using the straight lineHough transform: theory and generalization. IEEE Trans. Pattern Anal. Mach.Intell. 14, 1076–1089.

arrot, J.-F., Taud, H., 1992. Detection and classification of circular structures on SPOTimages. IEEE Trans. Geosci. Remote Sens. 30, 996–1005.

ock, I., Palmer, S., 1990. The legacy of Gestalt psychology. Sci. Am. (December),84–90.

uiz, L.A., Recio, J.A., Fernandez-Sarria, A., Hermosilla, T., 2011. A feature extractionsoftware tool for agricultural object-based image analysis. Comput. Electron.Agric. 76, 284–296.

th Observation and Geoinformation 34 (2015) 58–69 69

Segl, K., Kaufmann, H., 2001. Detection of small objects from high-resolutionpanchromatic satellite imagery based on supervised image segmentation. IEEETrans. Geosci. Remote Sens. 39, 2080–2083.

Shufelt, J.A., McKeown, D.M., 1993. Fusion of monocular cues to detect man-madestructures in aerial imagery. CVGIP: Image Understand. 57, 307–330.

Sırmac ek, B., Ünsalan, C., 2009. Urban-area and building detection usingSIFT keypoints and graph theory. IEEE Trans. Geosci. Remote Sens. 47,1156–1167.

Sohn, G., Dowman, I., 2007. Data fusion of high-resolution satellite imagery andLiDAR data for automatic building extraction. ISPRS J. Photogramm. RemoteSens. 62, 43–63.

Taud, H., Parrot, J.-F., 1992. Detection of circular structures on satellite images. Int.J. Remote Sens. 13, 319–335.

Tupin, F., Roux, M., 2003. Detection of building outlines based on the fusion of SARand optical features. ISPRS J. Photogramm. Remote Sens. 58, 71–82.

Ünsalan, C., Boyer, K.L., 2005. A system to detect houses and residential streetnetworks in multispectral satellite images. Comput. Vis. Image Understand. 98,423–461.

Van der Werff, H.M.A., Bakker, W.H., Van der Meer, F.D., Siderius, W., 2006. Com-bining spectral signals and spatial patterns using multiple Hough transforms:an application for detection of natural gas seepages. Comput. Geosci. 32,1334–1343.

Vapnik, V.N., 1995. The Nature of Statistical Learning Theory. Springer Verlag, NewYork, ISBN 0-387-94559-8.

Yip, R.K.K., Tam, P.K.S., Leung, D.N.K., 1992. Modification of Hough transform forcircles and ellipses detection using a 2-dimensional array. Pattern Recogn. 25,1007–1022.

Wang, L., 2005. Support Vector Machines: Theory and Applications. Springer-Verlag,

as Untersuchungen zur Lehre von der Gestalt II. Psychologische Forschung 4,301–350 (Translation published in W. Ellis (1938). A Source Book of GestaltPsychology, Routledge & Kegan, London, pp. 71–88).