ImageProcessing-BasedRecognitionofWallDefectsUsing ...downloads.hindawi.com/journals/cin/2018/7913952.pdf · ResearchArticle ImageProcessing-BasedRecognitionofWallDefectsUsing MachineLearningApproachesandSteerableFilters

Research ArticleImage Processing-Based Recognition of Wall Defects UsingMachine Learning Approaches and Steerable Filters

Nhat-Duc Hoang

Lecturer Faculty of Civil Engineering Institute of Research and Development Duy Tan University P809-03 Quang TrungDa Nang Vietnam

Correspondence should be addressed to Nhat-Duc Hoang hoangnhatducdtueduvn

Received 26 August 2018 Revised 12 October 2018 Accepted 18 October 2018 Published 15 November 2018

Guest Editor Zoran Peric

Copyright copy 2018 Nhat-Duc Hoang)is is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Detection of defects including cracks and spalls on wall surface in high-rise buildings is a crucial task of buildingsrsquo maintenance Ifleft undetected and untreated these defects can significantly affect the structural integrity and the aesthetic aspect of buildingsTimely and cost-effective methods of building condition survey are of practicing need for the building owners and maintenanceagencies to replace the time- and labor-consuming approach of manual survey )is study constructs an image processingapproach for periodically evaluating the condition of wall structures Image processing algorithms of steerable filters andprojection integrals are employed to extract useful features from digital images )e newly developed model relies on the Supportvector machine and least squares support vector machine to generalize the classification boundaries that categorize conditions ofwall into five labels longitudinal crack transverse crack diagonal crack spall damage and intact wall A data set consisting of 500image samples has been collected to train and test the machine learning based classifiers Experimental results point out that theproposed model that combines the image processing and machine learning algorithms can achieve a good classification per-formance with a classification accuracy rate 8533 )erefore the newly developed method can be a promising alternative toassist maintenance agencies in periodic building surveys

1 Introduction

During the construction and maintenance of high-risebuildings it is very crucial to attain good surface qualityof structures due to safety and esthetics aspects Because ofthe combined effects of aging weather conditions andhuman activities the condition of building structures de-teriorates over time [1] If left untreated damages such ascracks and spalls obviously cause inconvenience for thebuildingrsquos occupants deteriorate the structural integrity andlead to a significant reduction of the value of the poorlymaintained asset )erefore identifying defective areas thatappear on surface structure is one of the main tasks inperiodic survey buildings

In high-rise buildings the components that have largesurface areas typically include walls (both concrete walls andbrick walls covered by mortar) and slabs )e assessment ofconcrete slabs is usually performed during the construction

phase It is because during the operation phase the surface ofslab structure is concealed by floor coverings such as ceramicor stone tiles )erefore this study focuses on the visualassessment of wall structures

In Vietnam periodic surveys on building conditionare usually performed by visual assessment of human in-spectors )is fact is also common in other countries be-cause visual changes in structures can directly point out thepotential problems of building structures [2] Forinstances cracks can be indicators of structural problemsincluding building settlement and degradation of buildingmaterials particularly for concrete walls spalls can becaused by the corrosion of embedded reinforcement bars[3ndash5]

In the current practice of building condition assessmentthe damages on the surface of building structures are usuallyinspected by qualified technicians )ese technicians oftenutilize contact-type equipment including profilometer and

HindawiComputational Intelligence and NeuroscienceVolume 2018 Article ID 7913952 18 pageshttpsdoiorg10115520187913952

measuring tape for identifying the defective areas [6] Al-though the manual procedure can help to obtain accuratecondition of the structure it also has several disadvantagesFirst the surveying process is strongly affected by theknowledge experience and subjective judgment of humaninspectors therefore this issue can lead to inconsistencyof the assessment outcome Second the process of visualassessment measurement data processing and reportcan be very time consuming especially for high-rise build-ings with large surface areas needed to be inspectedperiodically

Accordingly it is immensely beneficial for the buildingowners and maintenance agencies if the manual inspectionprocess can be replaced by a more productive and consistentmethod of surveying [7] Among automated methods forbuilding condition evaluation machine vision based ap-proaches are widely employed due to their ease of access toequipment their fast computing processes and the rapidadvancements of image processing techniques [8ndash14]

Hutchinson and Chen [15] presented a statistical-basedmethod for evaluating concrete damage including cracksand spalls and relied on a Bayesian method for recognizingcracks automatically from images Chen et al [16] employedthe first derivative of a Gaussian filter to analyze multi-temporal images for measuring cracks Zhu [17] put forwardan intelligent method using three circular filters to detect airpockets appearing on the surfaces of concrete

)e level set method and morphological algorithms forimage processing have been used by Chen and Hutchinson[18] to identify and analyze cracks in laboratory environmentLee et al [19] proposed a model that integrates various imageprocessing operations (brightness adjustment binarisationand shape analysis) to facilitate the accuracy of crack de-tection in addition a neural network-based model wasimplemented to classify crack patterns of cracks Valenccedila et al[20] introduced a method based on multispectral imageanalysis to evaluate and delineate defective areas

An image processing-based approach for detectingbugholes on concrete surface has been established by Liu andYang [21] the employed techniques are contrast enhance-ment and Otsu thresholding Kim et al [22] compareddifferent image binarisation algorithms for identifyingcracks in concrete structures Hoang [23] employed the Otsumethod and a gray intensity modification approach forbinarizing images and isolating cracks

Silva and Lucena [24] demonstrated the capability ofdeep learning approach for concrete crack detection Dor-afshan et al [25] has recently compared the performances ofdeep convolutional neural networks and edge detectionmethods for the task of recognizing concrete cracks )enotable advantage of deep learning approaches is that theirfeature extraction operators are autonomously constructedduring the model training phase [26] However methodsbased on deep learning often necessitate a considerableamount of training samples and require a large computa-tional cost

As can be seen from the existing literature most of theprevious works have dedicated in constructing models forthe classification of crack and noncrack conditions Few

studies have constructed an integrated image processingmodel for detect cracks and spalls )e objective of thecurrent study is to combine image processing techniquesand advanced machine learning algorithms into an in-tegrated model that is capable of detecting and categorizingthe defective areas on wall structures By using an intelligentmodel that can recognize and categorize types of cracks andspalling areas simultaneously the task of periodicbuilding condition survey can be executed in amore effectivemanner

)e feature extraction phase of the new model relies onimage processing algorithms of steerable filters and pro-jection integrals It is because these two algorithms of imageprocessing have demonstrated their usefulness in recog-nizing defects in pavements [27 28] Based on the extractedfeatures machine learning algorithms includingsupport vector machine (SVM) and least squares supportvector machine (LSSVM) are employed to classify inputimages into five labels longitudinal cracks transversecracks diagonal cracks spalls and intact walls )e reasonfor the selection of these machine learning approachesis that their outstanding performances in classificationtasks have been reported in the literature [29ndash33]

Based on the aforementioned features the main con-tributions of the newly constructed model can be summa-rized as follows

(i) Multiple types of cracks (longitudinal crackstransverse cracks diagonal cracks) existing in wallstructures can be detected and categorized in anintegrated model )is can be considered to bea significant improvement since most of the existingmodels can only produce the prediction outcome ofcrack or noncrack conditions [13 23ndash25]

(ii) Although projection integrals have been utilized instructural defect classification [27 28] diagonalprojection integrals which have been rarelyexploited in concrete surface crack categorizationare employed in this study to specifically deal withdiagonal cracks

(iii) Cracks and spall damages can be recognized si-multaneously in an integrated model which has alsorarely been achieved in the current literature

(iv) Neural networks have been extensively usedin concrete crack categorization [19 34 35]However the applications of SVM and LSSVM inthis task are still limited and a comparative workneeds to be performed to evaluate the potential ofthese two advance machine learning algorithmsin dealing with the problem of wall defectrecognition

)e rest of the paper is organized as follows the researchmethodology is reviewed briefly in the next section followedby the description of the newly constructed automatic ap-proach for wall defect detection the fourth section reportsexperimental results of this study followed by the conclu-sion in the final section

2 Computational Intelligence and Neuroscience

2 Research Methodology

21 Image Processing Approaches

211 Steerable Filter (SF) SF [36] is an orientation-selectiveconvolution kernel used widely used for feature extractionIn image processing field oriented filters are oftenemployed in various vision and image processing tasksincluding edge detection and texture analysis [37] Sincecracks and spalls on wall surface have distinctive edges andpatterns the utilization of SF can be helpful to recognizethese defects

SF is based on the computation of directional derivativesof Gaussians accordingly these filters can be used toconstruct local orientation maps of a digital image [37] SF isessentially a linear combination of Gaussian second de-rivatives For an image I (xy) a 2-D Gaussian at a certainpixel is computed in the following formula [38]

G(x y r) 12πr

radic expminus x2 + y2( 1113857

2r2 (1)

where r is a free parameter which denotes the Gaussianfunction variance

)e expression of the SF formulation with an orientationof θ is given as follows

F(xyrθ) Gxx cos2(θ) +2Gxy cos(θ)sin(θ) + Gyy sin

2(θ)

(2)

where Gxx Gxy and Gyy denote the Gaussian second de-rivatives and their expression are given in the followingequations

Gxx(x y r) x2 minus r2( 1113857exp minus x2 + y2( 11138572r2( 1113857

2π

radicr5

Gyy(x y r) y2 minus r2( 1113857exp minus x2 + y2( 11138572r2( 1113857

2π

radicr5

Gxy(x y r) Gxy(x y r) xy exp minus x2 + y2( 11138572r2( 1113857

2π

radicr5

(3)

Notably if the value of the parameter r which is thevariance of the Gaussian function variance is fixed the finalresponse map of an image is obtained by combining theoutcomes of individual SFs with different values of θ In thisstudy the values of θ vary from 0deg to 360deg with an interval of30deg )e responses of SFs of images containing defects aredemonstrated in Figure 1 with different values of theGaussian function variance

In addition the final response map created by SFs for animage I is calculated using the equation below

R(x y) F(x y σ θ)lowast I(x y) (4)

where ldquolowastrdquo is the symbol of the convolution operator

212 Projection Integral (PI) PI is a widely used method forimage analysis )is approach is particularly useful for

shape and texture categorization and has been extensivelyemployed for face and facial recognition [39 40] In thefield of civil engineering this image analysis method hasbeen successfully applied in pavement crack classificationtasks [27 38 41] as well as pavement pothole recognition[28]

)e two PIs along the horizontal and vertical axes of animage are denoted as horizontal PI (HPI) and vertical PI(VPI) )ese two PIs are computed according to the fol-lowing equations

HPI(y) 1113944iisinxy

I(i y)

VPI(x) 1113944jisinyx

I(x j)(5)

where xy and yx are the set of horizontal pixels at the lo-cation y and the set of vertical pixels at the location x re-spectively (Figure 2)

Since HPI and VPI are incapable of detecting diagonalcracks [42] the two diagonal PIs (DPIs) of an image areemployed )e directions of the two DPs denoted as DPI1and DPI2 are illustrated in Figure 3 )e two DPIs arecalculated as follows

DPI1(x y) 1113944xyisinD1

I(x y)


I(x y)(6)

where D1 and D2 are the set of pixels along the two diagonaldirections of an image (as illustrated in Figure 3)

As illustrated in Figure 4 images containing diagonalcracks have PIs in which there are exceptionally high in-tensities in the two DPIs In addition as can be shown inFigure 5 HPI and VPI have the strongest responses of PI inimages having longitudinal and transverse cracks On theother hand an image containing a spall damage results in PIswhich do not have a significant peak of signal and the av-erage value of its SF response is higher than that of an imagewithout defective areas

22 SupportVectorMachine andLeast Squares SupportVectorMachine Support vector machine (SVM) is a machinelearning based classifier which is established on the basis ofthe statistical learning theory [43] )e aim of this learningalgorithm is to find a predictive function based on thecollected data set )e standard version of SVM is designedto cope with binary or two-class pattern-recognitionproblems )rough the model construction phase SVMconstructs a hyperplane to classify data points so that thedistance from it to the nearest data sample of each class labelis maximized [44]

Moreover this algorithm relies on the kernel trick tobetter deal with nonlinearly separable cases Using the kerneltrick the data points are mapped from an original inputspace to a high-dimensional feature space so that linearseparability is easier to achieve (Figure 6) Superior

Computational Intelligence and Neuroscience 3

Original image SF response r = 10 SF response r = 15 SF response r = 20

(a)


(b)


(c)


(d)


(e)

Figure 1 SF responses (a) longitudinal crack (b) transverse crack (c) diagonal crack (d) spall damage and (e) intact wall


classification performance of SVM has been widely reportedin a large number of previous studies [45ndash49]

Given a set of training data points xk yk1113864 1113865Nk1 with input

data xk isin Rn and a set of class labels yk isin minus1 +1 the modelconstruction phase of SVM is equivalent to solving thefollowing optimization problem

minimize Jp(w e) 12w

Tw + c

12

1113944

N

k1e2k

subject to yk wTφ xk( 1113857 + b1113872 1113873ge 1minus ek k 1 N ek ge 0

(7)

where w isin Rn and b isin R denote the model parameters ek gt 0represents a slack variable c denotes a penalty constantwhich determines severity of learning error and φ(x) de-notes a nonlinear mapping from the input space to thefeature space

One of the advantages of SVM is that it does not requireexpressing the mapping function φ(x) explicitly Due to theconcept of kernel trick the model identification only ne-cessitates the computation of the kernel function K() whichis the dot product of φ(x) )e kernel function is shownbelow

K xk xl( 1113857 φ xk( 1113857Tφ xl( 1113857 (8)

Radial basis function (RBF) is often selected to be used inSVM [30] its formula is given as follows

K xk xl( 1113857 exp minusxk minusxl

2

2σ2⎛⎝ ⎞⎠ (9)

where σ denotes the kernel function parameterTo solve the aforementioned constrained optimization

the Lagrangian is given as follows

L(wbeαv) Jp(we)minus1113944N

k1αk yk w

Tφ xk( 1113857 + b1113872 1113873minus1+ ek1113966 1113967

minus1113944N

k1vkek

(10)

where αk ge 0 vk ge 0 denote Lagrange multipliers for k 1 2 N

Accordingly the conditions for optimality are stated asfollows

zL

zw 0⟶ w 1113944

N

k1αkykφ xk( 1113857

zL

zb 0⟶ 1113944

N

k1αkyk 0

zL

zek

0⟶ 0le αk le c k 1 N

⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(11)

y xy

(a)

x

yx

(b)

Figure 2 Illustration of PIs of an image (a) HPI and (b) VPI

y DPI1 DPI2

x

Figure 3 Illustration of DPIs of an image


Based on the equations found by the conditions foroptimality it is able to attain the following dual quadraticprogramming problem

maxα

JD(α) minus12sumN

kl1ykylφ xk( )Tφ xl( )αkαl +sum

N

k1αk

Subject to sumN

k1αkyk 0 0le αk le c k 1 N

(12)

In addition the kernel function is applied in the fol-lowing manner

ω ykylφ xk( )Tφ xl( ) ykylK xk xl( ) (13)

Finally the classication model based on SVM can bederived as follows

y(x) sign sumSV

k1αkykK xk xl( ) + b (14)

where SV is the number of support vectors which aretraining data points that have αk gt 0

Least squares support vector machine (LSSVM) [50] isa least square version of the original SVM Instead of solvinga quadratic programming problem required by SVM themodel construction phase of LSSVM is equivalent to solvinga system of linear equation erefore the computationalexpense of LSSVM can be much lower than that of thestandard SVM

To construct a LSSVM based classier it is needed tosolve the following minimization problem

minimize Jp(w e) 12wTw + c

12sumN

k1e2k

subject to yk wTφ xk( ) + b( ) 1minus ek k 1 N

(15)

where w isin Rn and b isin R are also the model parametersek isin R denote error variables cgt 0 is called a regularizationconstant

0 100 200 300 400Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

SF responseOriginal image

(a)

HPIVPI

DPI1DPI2


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIsOriginal image SF response

(b)

Figure 4 PIs of diagonal cracks




0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(a)SF responseOriginal image


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(b)SF responseOriginal image


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(c)SF responseOriginal image


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)

Figure 5 PIs of image samples (a) longitudinal crack (b) transverse crack (c) spall damage and (d) intact wall


Subsequently the Lagrangian is applied in the followingway

L(w b e α) Jp(w e)minus 1113944

N

k1αk yk w

Tφ xk( 1113857 + b1113872 1113873minus 1 + ek1113966 1113967

(16)

where αk is the kth Lagrange multiplier φ(xk) representsa nonlinear mapping function

Relied on the KKTconditions for optimality it is able toconvert the aforementioned constrained optimizationproblem to a linear system [51] )e classifier based onLSSVM is compactly expressed as follows

y(x) sign 1113944N

k1αkyiK xk xl( 1113857 + b⎛⎝ ⎞⎠ (17)

where αk and b are found by solving a linear system Similarto the standard SVM K(xk xl) denotes the kernel function

3 Image Sample Collection

Because the machine learning algorithms of SVM andLSSVM are supervised learning approach a set of wallimages with the corresponding ground truth labels must beprepared in advance of the model construction and classi-fication phases To establish the required data set images ofwalls have been collected during field surveys at high-risebuildings in Da Nang city (Vietnam)

To ease the computational process the size of each imagesample is fixed to be 200 times 200 pixels Moreover there arefive classes of wall condition namely longitudinal cracks(LC) transverse crack (TC) diagonal cracks (DC) spalldamage (SD) and intact wall (IW) )e number of imagesamples in each class is 100 )us the collected image dataset contains 500 samples and is illustrated in Figure 7 It isalso noted that all the image samples have been preprocessedby the median filter with a window size of 5 times 5 pixels )ispreprocessing step aims at suppressing the noise existing inthe collected digital image [52] In addition the data set isdivided into two folders that contain the training set ofimages (90) and the testing set of images (10))e first set

is used in the phase of model construction and the second setis utilized to verify the generalization capability of the walldefect classification model

4 The Proposed Hybrid Approach of ImageProcessing and Machine Learning forDetection of Wall Defects

)is section describes the proposed model used for auto-matic classification of wall defects )e overall modelstructure is illustrated in Figure 8 )e model can be dividedinto two separated modules

(i) Feature extraction phase that employs the imageprocessing methods of SFs and IPs

(ii) Machine learning based classification phase thatrelies on the SVM and LSSVM algorithms

In the first step of feature extraction SFs are employed tocompute a salient defect map from a digital image )eminimum and maximum angles of SFs are 0deg and 360degrespectively )e parameter r of SFs is selected from a set of[10 15 20 25 30] )e value of r which results in thehighest accuracy for the training data set is selected as theoptimal one Based on the map generated by SFs PIs in-cluding HPI VPI DPI1 and DPI2 are then computed tocharacterize the texture of the captured images As earliermentioned each image of the data set has the size of 200 times

200 pixels )us each PI contains 200 sampled points andthe total number of features to be analyzed by the machinelearning algorithm is 200 times 4 800 Herein 4 is the numberof PIs

Obviously the original PIs can be very rough and havemany peaks and valleys due to local fluctuations of the grayintensity of an image Moreover the large number of inputfeatures can create difficulty for the machine learning al-gorithms of SVM and LSSVM due to the curse of di-mensionality [53] )erefore it is beneficial to smooth theoriginal PIs by the use of moving average method [42] Indetail the average value ofWPI consecutive values along thePIs is calculated to create smoothed PIs with fewer datapoints For example if WPI 10 then the total number of

x1

Kernel mapping

Φ(xu)

Φ(xv)

Φ(x)

The original input space The high-dimensional feature spacex2 Φ(xl)

Label 1

Label 2

Label 2

Label 1

Hyperplane

Figure 6 )e learning phase of SVM


features in the contracted PIs is reduced from 800 to 80 ifWPI 20 then the machine learning algorithms only have todeal with a data set having 40 features

)e process of feature number reduction using movingaverage for an image is demonstrated in Figure 9 As can beseen from this example the smoothed PIs even with thewindow size WIP 20 still preserve crucial features of risesand ebbs of the original PIs )erefore the value ofWIP 20is selected for the feature extraction step Accordingly theset of 40 features extracted from PIs is used as input pattern

to categorize the four labels of wall defect (LC TC DC andSD) and the label of intact wall (IW)

Furthermore to obtain the two DPIs of DPI1 and DPI2the original map of the SF response has been rotated with theangles of +45 and minus45 [42] Accordingly the two DPIs arederived from the computation of the HPIs of the two rotatedSF maps )e process of computing DPIs of images con-taining diagonal cracks is demonstrated in Figure 10 Inaddition the overall feature extraction module is depicted inFigure 11 Based on the 40 input features (IF) extracted from

(a)

(b)

(c)

(d)

(e)

Figure 7 )e collected image samples (a) longitudinal crack (b) transverse crack (c) diagonal crack (d) spall damage and (e) intact wall


the PIs the machine learning algorithms of SVM andLSSVM are employed to perform the model learningand classification of images in the second module of themodel

It is noted that the standard versions of SVM andLSSVM are designed two-class pattern recognition

problems Hence the one-versus-one (OvO) strategy [54]has been used with SVM and LSSVM to make them capableof dealing with the five-class recognition tasks at handPrevious works have confirmed the advantages of OvOstrategy in coping with multiclass classification problems[53 55 56]

Image acquisition

Data set

Model training

Model prediction

Wall defect classification

Steerable filters

Integral projections

Training data set (90)

Testing data set (10)

HPI

VPI

DPIs

Figure 8 )e proposed model structure


(a)

0 100 200 300 400 500 600 700 800Pixel interval

002040608

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(b)

0 10 20 30 40 50 60 70 80Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI 1DPI 2

(c)

0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)

Figure 9 PIs of an image (a) the original image (b) the original PIs (WPI 1) (c) the smoothed PIs (WPI 10) and (d) the smoothed PIs(WPI 20)



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

SF

Original image

Rotated SF 1 Rotated SF 2

HPIVPI

DPI1DPI2

(a)


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

SF Rotated SF 1 Rotated SF 2

HPIVPI

DPI1DPI2

Original image

(b)

Figure 10 )e process of computing DPIs (a) a minus45deg diagonal crack and (b) a + 45deg diagonal crack


5 Experimental Results

Because the detection of wall defects is formulated as a ve-class pattern recognition problem classication accuracy rate(CAR) computed for each individual class and for all of theclasses is employed CAR for the class i is computed as follows

CARi RiCRiA

times 100() (18)

where RiC and RiA denote the number of data samples in class

ith being correctly classied and the total number of datainstances in this class respectively It is reminded that thatthere are ve class labels in the data set longitudinal crack(LC) transverse crack (TC) diagonal crack (DC) spalldamage (SD) and intact wall (IW)

e overall classication accuracy rate (CAR) for all theve class labels is simply computed as follows

sFSdetatoRFSegamitupnI

sIPlanigiroehTsIPdehtoomsehT

The extracted feature used for pattern classification

0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

0 100 200 300 400 500 600 700 800Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

IF1 IF2 IF3 IF4 IF5 IF36 IF37 IF38 IF39 IF400015 0013 0012 0013 0015 hellip 0053 0042 0024 0025 0004

HPIVPIDPI1DPI2

HPIVPIDPI1DPI2

Figure 11 e whole feature extraction process

8127

8307

8513

81738040

LSSVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(a)

7780

8013

8433

8133

7833

SVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(b)

Figure 12 Model performance with dierent values of the parameter r (a) LSSVM and (b) SVM


Table 1 Result comparison

Statistics CAR ()Classification models

LSSVM SVM BPANN CT LDA NBC

Average

CARLC 8533 7900 7933 6700 7433 7300CARTC 8300 7567 7733 6867 7467 7267CARDC 9000 9467 6367 6267 5300 5467CARSD 7833 8433 6933 5533 3133 2600CARIW 8900 8800 8133 5767 7467 7233CARO 8513 8433 7420 6227 6160 5973

Std


LSSVM SVM BPANN CT LDA NBCClassification models

55

60

65

70

75

80

85

CAR

()

CARo

Figure 13 Result comparison based on CARo


3035404550556065707580859095

100

CAR

()

CARLC

CARTC

CARDC

CARSD

CARIW

Figure 14 Classification accuracy comparison based on CAR of each class (LD TC DC SD and IW)


CAROverall sum5

i1

CARi5

(19)

As stated earlier the data set including 500 samples isemployed to train and test the wall defect classicationmodel is data set is randomly separated into two subsetsdata for model training (90) and data for testing (10)Because a single run of model training and testing may nothelp to reveal the true performance of a classier this studyperforms a repeated subsampling process that includes 20times of model training and prediction In each time ofrunning 10 of the data set is randomly extracted to formthe testing data subset the rest of the data set is reserved formodel testing phase

As described in the formulation of the two machinelearning algorithms of SVM and LSSVM these two algo-rithms both require a proper setting of their tuning pa-rameters In the case of SVM the tuning parameters are thepenalty coecient and the kernel function parameter In thecase of LSSVM the regularization and the kernel functionparameters need to be selected appropriately In this sectionthe grid search method described in the previous work of[57] is employed for automatically setting those tuningparameters of SVM and LSSVM In addition the featureselection stage requires the setting of the parameter r in thecomputation of SFs e model performance with dierentvalues of the parameter r for the cases of SVM and LSSVMis reported in Figure 12 As can be observed from this gure

r 2 results in the highest CAROverall values for both SVMand LSSVM

Besides the two machine learning algorithms of SVM andLSSVM the backpropagation articial neural network(BPANN) classication tree (CT) linear discriminant anal-ysis (LDA) and naive Bayes classier (NBC) are alsoemployed as benchmark classiers It is also noted that CTLDA and NBC are also equipped with the OvO strategy todeal with the current ve-class recognition problem of walldefect classication e SVM and the benchmark algorithmsimplemented in MATLAB with the help of the statistics andmachine learning toolbox [58] In addition the LSSVMmodelis constructed by the built-in functions provided in thetoolbox developed by De Brabanter et al [59]

In addition the training phases of BPANN and CTrequire a proper setting of their model hyper-parametersSimilar to the cases of SVM and LSSVM the model hyper-parameters of those models that result in the best perfor-mance for the testing set are selected In the case of the DTmodel the suitable value of the minimal number of ob-servations per tree leaf is found to be 2 e appropriatestructure of the BPANNmodel consists of 35 neurons in thehidden layers furthermore the scaled conjugate gradientalgorithm with the maximum number of training epochs 3000 is used to train the neural network model Moreoverthe processes of selecting an appropriate value of the tuningparameter r used in constructing the SF maps of thebenchmark models of BPANN CT LDA and NBC are


50

55

60

65

70

75

80

85

90

95

CAR O

Figure 15 Box plots of overall CARs

Table 2 p values of the Wilcoxon signed-rank test

LSSVM SVM BPANN CT LDA NBCLSSVM 000000 060134 000001 000000 000000 000000SVM 060134 000000 000002 000000 000000 000000BPANN 000001 000002 000000 000000 000000 000000CT 000000 000000 000000 000000 006728 000444LDA 000000 000000 000000 006728 000000 002230NBC 000000 000000 000000 000444 002230 000000



0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)

SF responses Projection integrals


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05

1PIsRotated SF 1 Rotated SF 2SF

Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Figure 16 Examples of incorrect classications


similar to those of the SVM and LSSVM models Based onresult comparisons the suitable value of the tuning pa-rameter r used with BPANN CT LDA and NBC is also 2

)e performances of the machine learning classifiersused for wall damage recognition are summarized in Table 1Observed from this table LSSVM has achieved the bestpredictive performance in terms of CARo (8513) followedby SVM (CARo 8433) BPANN (CARo 7420) CT(CARo 6227) LDA (CARo 6100) and NBC (CARo

5973))us it can be seen that classifiers based on LSSVMand SVM are more suitable for the task of wall defectclassification than other machine learning and statisticalapproaches of BPANN CT LDA and NBC

Figures 13 and 14 graphically compare the results ob-tained from the prediction models in terms of CARo andCAR of each individual class label It is shown that the CARsof LC (8533) TC (8300) and IW (8900) obtainedfrom LSSVM are higher than those yielded by SVM (79007567 and 8800 for the LC TC and IW respectively)However results of SVM for the classes of DC (9467) andSD (8433) are better than those of LSSVM (9000 and7833 for the classes of DC and SD respectively)

In addition the classification results of all the models interms of CARo are displayed by the box plots in Figure 15Moreover to better demonstrate the statistical differencebetween each pair of classifiers employed in the task of walldefect recognition theWilcoxon signed-rank test (WSRT) isutilized in this section WSRT is a nonparametric statisticalhypothesis test that is widely used for verifying the statisticaldifference of model performances [57] With the significancelevel of the test 005 if p value computed from the test issmaller than 005 it can be confirmed that the performancesof the two selected classifiers are statistically different )e p

values obtained from WSRT for each pair of classifiers arereported in Table 2 )e outcomes shown in this tableconfirm that LSSVM and SVM are significantly better thanother benchmark models in the task of recognizing walldamages In addition with p values 060134 there is nostatistical difference between the performances of LSSVMand SVM

Although the LSSVM model has delivered the highestCARs this model also commits wrong classification cases)ese misclassifications are investigated and examples ofthem are illustrated in Figure 16 In Figures 16(a) and 16(b)images with the ground truth label of LC and TC have beenassigned the label of IW )e reasons for these mis-classifications are that the crack objects are too thinmoreover there is a line of stain existing in Figure 16(b))ecase in Figure 16(c) shows an image with its ground truthclass of DC which has been categorized as the class of TC)e possible reason of this phenomenon is that the crackobject appears too close to corner of the image therefore thesignals of the SF responses of the two DPIs are not signif-icantly stronger than those of other PIs In Figure 16(d) animage with complex background texture has caused themachine to misclassify an image with the ground truthcategory of SD Moreover an object of stain (Figure 16(e))leads to the classification of an image into the class of LCwhile it actually belongs to the class of IW

6 Conclusion

)is study has proposed an automatic approach for periodicsurvey of concrete wall structures )e newly constructedapproach mainly consists of the feature extraction step andthe pattern classification step In the first step image pro-cessing techniques of SF and PI have been employed tocharacterize the texture and the pattern existing in imagesIn the second step machine learning algorithms of SVM andLSSVM have been used to analyze the features extracted bythe image processing techniques and to assign input imagesinto one of the five class labels of LC TC DC SD and IWExperimental results using a repeated random subsamplingwith 20 runs show that the predictive performances ofLSSVM (CARo 8513) and SVM (CARo 8433) aresuperior to other benchmark models of BPANN CT LDAand NBC )ese facts confirm that the proposed approachcan be a time- and cost-effective solution for the task ofbuilding periodic survey )e future developments of thecurrent study include the integration of other advancedimage processing methods (eg image segmentationcolortexture analyses) to enhance the CARs via the re-duction of falsely classified cases

Data Availability

)e data used to support the findings of this study areavailable from the corresponding author upon request

Conflicts of Interest

)e author declares that there are no conflicts of interestsregarding the publication of this research work

Supplementary Materials

)e supplementary file contains the data set used in thisstudy In this file the first 500 columns are the input features(X1 to X40) of the data (which are the smoothed projectionintegrals) the last column is the class labels (1 longitudinalcrack 2 transverse crack 3 diagonal crack 4 spalldamage and 5 intact wall) (Supplementary Materials)

References

[1] W Zhang Z Zhang D Qi and Y Liu ldquoAutomatic crackdetection and classification method for subway tunnel safetymonitoringrdquo Sensors vol 14 no 10 pp 19307ndash19328 2014

[2] Y-J Cha W Choi G Suh S Mahmoudkhani andO Buyukozturk ldquoAutonomous structural visual inspectionusing region-based deep learning for detecting Multipledamage typesrdquo Computer-Aided Civil and InfrastructureEngineering vol 33 no 9 pp 731ndash747 2018

[3] T Dawood Z Zhu and T Zayed ldquoMachine vision-basedmodel for spalling detection and quantification in subwaynetworksrdquo Automation in Construction vol 81 pp 149ndash1602017

[4] S German I Brilakis and R DesRoches ldquoRapid entropy-based detection and properties measurement of concretespalling with machine vision for post-earthquake safety


assessmentsrdquo Advanced Engineering Informatics vol 26no 4 pp 846ndash858 2012

[5] M-K Kim H Sohn and C-C Chang ldquoLocalization andquantification of concrete spalling defects using terrestriallaser scanningrdquo Journal of Computing in Civil Engineeringvol 29 no 6 article 04014086 2015

[6] P Tang D Huber and B Akinci ldquoCharacterization of laserscanners and algorithms for detecting flatness defects onconcrete surfacesrdquo Journal of Computing in Civil Engineeringvol 25 no 1 pp 31ndash42 2011

[7] H Kim E Ahn M Shin and S-H Sim ldquoCrack and noncrackclassification from concrete surface images using machinelearningrdquo Structural Health Monitoring vol 0 article1475921718768747 2018

[8] R Davoudi G R Miller and J N Kutz ldquoStructural loadestimation using machine vision and surface crack patternsfor shear-critical RC beams and slabsrdquo Journal of Com-puting in Civil Engineering vol 32 no 4 article 040180242018

[9] J-Y Jung H-J Yoon and H-W Cho ldquoA study on crackdepth measurement in steel structures using image-basedintensity differencesrdquo Advances in Civil Engineeringvol 2018 Article ID 7530943 10 pages 2018

[10] C Koch K Georgieva V Kasireddy B Akinci andP Fieguth ldquoA review on computer vision based defect de-tection and condition assessment of concrete and asphalt civilinfrastructurerdquo Advanced Engineering Informatics vol 29no 2 pp 196ndash210 2015

[11] C Koch S G Paal A Rashidi Z Zhu M Konig andI Brilakis ldquoAchievements and challenges in machine vision-based inspection of large concrete structuresrdquo Advances inStructural Engineering vol 17 no 3 pp 303ndash318 2014

[12] H Pragalath S Seshathiri H Rathod B Esakki andR Gupta ldquoDeterioration assessment of infrastructure usingfuzzy logic and image processing algorithmrdquo Journal ofPerformance of Constructed Facilities vol 32 no 2 article04018009 2018

[13] Y-S Yang C-l Wu T T C Hsu H-C Yang H-J Lu andC-C Chang ldquoImage analysis method for crack distributionand width estimation for reinforced concrete structuresrdquoAutomation in Construction vol 91 pp 120ndash132 2018

[14] Z Zhu S German and I Brilakis ldquoDetection of large-scaleconcrete columns for automated bridge inspectionrdquo Auto-mation in Construction vol 19 no 8 pp 1047ndash1055 2010

[15] T C Hutchinson and Z Chen ldquoImproved image analysis forevaluating concrete damagerdquo Journal of Computing in CivilEngineering vol 20 no 3 pp 210ndash216 2006

[16] L-C Chen Y-C Shao H-H Jan C-W Huang andY-M Tien ldquoMeasuring system for cracks in concrete usingmultitemporal imagesrdquo Journal of Surveying Engineeringvol 132 no 2 pp 77ndash82 2006

[17] Z Zhu and I Brilakis ldquoDetecting air pockets for architecturalconcrete quality assessment using visual sensingrdquo ElectronicJournal of Information Technology in Construction vol 13pp 86ndash102 2008

[18] Z Chen and T C Hutchinson ldquoImage-based framework forconcrete surface crack monitoring and quantificationrdquo Ad-vances in Civil Engineering vol 2010 Article ID 21529518 pages 2010

[19] B Y Lee Y Y Kim S-T Yi and J-K Kim ldquoAutomatedimage processing technique for detecting and analysingconcrete surface cracksrdquo Structure and Infrastructure Engi-neering vol 9 no 6 pp 567ndash577 2013

[20] J Valenccedila L M S Gonccedilalves and E Julio ldquoDamage as-sessment on concrete surfaces using multi-spectral imageanalysisrdquo Construction and Building Materials vol 40pp 971ndash981 2013

[21] B Liu and T Yang ldquoImage analysis for detection of bugholeson concrete surfacerdquo Construction and Building Materialsvol 137 pp 432ndash440 2017

[22] H Kim E Ahn S Cho M Shin and S-H Sim ldquoComparativeanalysis of image binarization methods for crack identifica-tion in concrete structuresrdquo Cement and Concrete Researchvol 99 pp 53ndash61 2017

[23] N-D Hoang ldquoDetection of surface crack in building struc-tures using image processing technique with an improvedOtsu method for image thresholdingrdquo Advances in CivilEngineering vol 2018 Article ID 3924120 10 pages 2018

[24] W R L de Silva and D S d Lucena ldquoConcrete cracks de-tection based on deep learning image classificationrdquo Pro-ceedings vol 2 no 8 p 489 2018

[25] S Dorafshan R J )omas and M Maguire ldquoComparison ofdeep convolutional neural networks and edge detectors forimage-based crack detection in concreterdquo Construction andBuilding Materials vol 186 pp 1031ndash1045 2018

[26] Y-J Cha W Choi and O Buyukozturk ldquoDeep learning-based crack damage detection using convolutional neuralnetworksrdquo Computer-Aided Civil and Infrastructure Engi-neering vol 32 no 5 pp 361ndash378 2017

[27] A Cubero-Fernandez F J Rodriguez-Lozano R VillatoroJ Olivares and J M Palomares ldquoEfficient pavement crackdetection and classificationrdquo EURASIP Journal on Image andVideo Processing vol 2017 no 39 2017

[28] N-D Hoang ldquoAn artificial intelligence method for asphaltpavement pothole detection using least squares support vectormachine and neural network with steerable filter-based fea-ture extractionrdquo Advances in Civil Engineering vol 2018Article ID 7419058 12 pages 2018

[29] G M Hadjidemetriou P A Vela and S E ChristodoulouldquoAutomated pavement patch detection and quantificationusing support vector machinesrdquo Journal of Computing in CivilEngineering vol 32 no 1 article 04017073 2018

[30] B T Pham A Jaafari I Prakash and D T Bui ldquoA novelhybrid intelligent model of support vector machines and theMultiBoost ensemble for landslide susceptibility modelingrdquoBulletin of Engineering Geology and the Environment 2018 Inpress

[31] H Shin and J Paek ldquoAutomatic task classification via supportvector machine and crowdsourcingrdquo Mobile InformationSystems vol 2018 Article ID 6920679 9 pages 2018

[32] M Wang Y Wan Z Ye and X Lai ldquoRemote sensing imageclassification based on the optimal support vector machineand modified binary coded ant colony optimization algo-rithmrdquo Information Sciences vol 402 pp 50ndash68 2017

[33] Y Zhou W Su L Ding H Luo and P E D Love ldquoPre-dicting safety risks in deep foundation pits in subway in-frastructure projects support vector machine approachrdquoJournal of Computing in Civil Engineering vol 31 no 5 article04017052 2017

[34] G K Choudhary and S Dey ldquoCrack detection in concretesurfaces using image processing fuzzy logic and neuralnetworksrdquo in Proceedings of 2012 IEEE Fifth InternationalConference on Advanced Computational Intelligence (ICACI)pp 404ndash411 Nanjing China October 2012

[35] A Mohan and S Poobal ldquoCrack detection using imageprocessing a critical review and analysisrdquo Alexandria Engi-neering Journal vol 57 no 2 2017


[36] W T Freeman and E H Adelson ldquoSteerable filters for earlyvision image analysis and wavelet decompositionrdquo in Pro-ceedings of ird International Conference on Computer Vi-sion pp 406ndash415 Osaka Japan December 1990

[37] E P Simoncelli and H Farid ldquoSteerable wedge filtersrdquo inProceedings of IEEE International Conference on ComputerVision pp 189ndash194 Cambridge MA USA June 1995

[38] N-D Hoang and Q-L Nguyen ldquoA novel method for asphaltpavement crack classification based on image processing andmachine learningrdquo Engineering with Computers 2018 Inpress

[39] A Chouchane M Belahcene and S Bourennane ldquo3D and 2Dface recognition using integral projection curves based depthand intensity imagesrdquo International Journal of IntelligentSystems Technologies and Applications vol 14 no 1pp 50ndash69 2015

[40] A Hernandez-Matamoros A Bonarini E Escamilla-Her-nandez M Nakano-Miyatake and H Perez-Meana ldquoFacialexpression recognition with automatic segmentation of faceregions using a fuzzy based classification approachrdquoKnowledge-Based Systems vol 110 pp 1ndash14 2016

[41] S Li Y Cao and H Cai ldquoAutomatic pavement-crack de-tection and segmentation based on steerable matched filteringand an active contour modelrdquo Journal of Computing in CivilEngineering vol 31 no 5 article 04017045 2017

[42] N-D Hoang and Q-L Nguyen ldquoAutomatic recognition ofasphalt pavement cracks based on image processing andmachine learning approaches a comparative study on clas-sifier performancerdquo Mathematical Problems in Engineeringvol 2018 Article ID 6290498 16 pages 2018

[43] V N Vapnik Statistical Learning eory John Wiley amp SonsInc Hoboken NJ USA 1998 ISBN-10 0471030031

[44] L H Hamel Knowledge Discovery with Support Vector Ma-chines John Wiley amp Sons Inc Hoboken NJ USA 2009ISBN 978-0-470-37192-3

[45] W L Al-Yaseen Z A Othman and M Z A Nazri ldquoMulti-level hybrid support vector machine and extreme learningmachine based on modified K-means for intrusion detectionsystemrdquo Expert Systems with Applications vol 67 pp 296ndash303 2017

[46] W Chen H R Pourghasemi and S A Naghibi ldquoA com-parative study of landslide susceptibility maps produced usingsupport vector machine with different kernel functions andentropy data mining models in Chinardquo Bulletin of EngineeringGeology and the Environment vol 77 no 2 pp 647ndash6642018

[47] H Hasni A H Alavi P Jiao and N Lajnef ldquoDetection offatigue cracking in steel bridge girders a support vectormachine approachrdquo Archives of Civil and Mechanical Engi-neering vol 17 no 3 pp 609ndash622 2017

[48] B Kalantar B Pradhan S A Naghibi A Motevalli andS Mansor ldquoAssessment of the effects of training data selectionon the landslide susceptibility mapping a comparison be-tween support vector machine (SVM) logistic regression (LR)and artificial neural networks (ANN)rdquo Geomatics NaturalHazards and Risk vol 9 no 1 pp 49ndash69 2018

[49] A Saxena and S Shekhawat ldquoAmbient air quality classifi-cation by grey wolf optimizer based support vector machinerdquoJournal of Environmental and Public Health vol 2017 ArticleID 3131083 12 pages 2017

[50] J A K Suykens and J Vandewalle ldquoLeast squares supportvector machine classifiersrdquo Neural Processing Letters vol 9no 3 pp 293ndash300 1999

[51] J Suykens J V Gestel J D Brabanter B D Moor andJ Vandewalle Least Square Support Vector Machines WorldScientific Publishing Co Pte Ltd Singapore 2002 ISBN-13978-9812381514

[52] H Rababaah ldquoAsphalt pavement crack classificationa comparative study of three ai approaches multilayer per-ceptron genetic algorithms and self-organizing mapsrdquo MS)esis Indiana University South Bend South Bend IN USA2005

[53] C M Bishop Pattern Recognition and Machine Learning(Information Science and Statistics Springer Berlin Ger-many ISBN-10 0387310738 2011

[54] A Rocha and S K Goldenstein ldquoMulticlass from binaryexpanding one-versus-all one-versus-one and ECOC-basedapproachesrdquo IEEE Transactions on Neural Networks andLearning Systems vol 25 no 2 pp 289ndash302 2014

[55] K-B Duan J C Rajapakse and M N Nguyen One-Versus-One and One-Versus-All Multiclass SVM-RFE for Gene Se-lection in Cancer Classification pp 47ndash56 Springer BerlinHeidelberg Berlin Heidelberg 2007

[56] C-W Hsu and C-J Lin ldquoA comparison of methods formulticlass support vector machinesrdquo IEEE Transactions onNeural Networks vol 13 pp 415ndash425 2002

[57] N-D Hoang and D T Bui ldquoPredicting earthquake-inducedsoil liquefaction based on a hybridization of kernel Fisherdiscriminant analysis and a least squares support vectormachine a multi-dataset studyrdquo Bulletin of Engineering Ge-ology and the Environment vol 77 no 1 pp 191ndash204 2018

[58] Matwork Statistics and Machine Learning Toolbox UserrsquosGuide Matwork Inc Natick MA USA 2017 httpswwwmathworkscomhelppdf_docstatsstatspdf

[59] K De Brabanter P Karsmakers F Ojeda et al LS-SVMlabToolbox Userrsquos Guide Version 18 2010


Computer Games Technology

International Journal of

Hindawiwwwhindawicom Volume 2018

Hindawiwwwhindawicom

Journal ofEngineeringVolume 2018

Advances in

FuzzySystems


Volume 2018


ReconfigurableComputing



Applied Computational Intelligence and Soft Computing

thinspAdvancesthinspinthinsp

thinspArtificial Intelligence

Hindawiwwwhindawicom Volumethinsp2018


Civil EngineeringAdvances in


Electrical and Computer Engineering

Journal of

Journal of

Computer Networks and Communications


Hindawi

wwwhindawicom Volume 2018

Advances in

Multimedia


Biomedical Imaging



Engineering Mathematics


RoboticsJournal of



Computational Intelligence and Neuroscience


Mathematical Problems in Engineering

Modelling ampSimulationin EngineeringHindawiwwwhindawicom Volume 2018

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018


Human-ComputerInteraction

Advances in


Scientic Programming

Submit your manuscripts atwwwhindawicom

measuring tape for identifying the defective areas [6] Al-though the manual procedure can help to obtain accuratecondition of the structure it also has several disadvantagesFirst the surveying process is strongly affected by theknowledge experience and subjective judgment of humaninspectors therefore this issue can lead to inconsistencyof the assessment outcome Second the process of visualassessment measurement data processing and reportcan be very time consuming especially for high-rise build-ings with large surface areas needed to be inspectedperiodically

Accordingly it is immensely beneficial for the buildingowners and maintenance agencies if the manual inspectionprocess can be replaced by a more productive and consistentmethod of surveying [7] Among automated methods forbuilding condition evaluation machine vision based ap-proaches are widely employed due to their ease of access toequipment their fast computing processes and the rapidadvancements of image processing techniques [8ndash14]

Hutchinson and Chen [15] presented a statistical-basedmethod for evaluating concrete damage including cracksand spalls and relied on a Bayesian method for recognizingcracks automatically from images Chen et al [16] employedthe first derivative of a Gaussian filter to analyze multi-temporal images for measuring cracks Zhu [17] put forwardan intelligent method using three circular filters to detect airpockets appearing on the surfaces of concrete

)e level set method and morphological algorithms forimage processing have been used by Chen and Hutchinson[18] to identify and analyze cracks in laboratory environmentLee et al [19] proposed a model that integrates various imageprocessing operations (brightness adjustment binarisationand shape analysis) to facilitate the accuracy of crack de-tection in addition a neural network-based model wasimplemented to classify crack patterns of cracks Valenccedila et al[20] introduced a method based on multispectral imageanalysis to evaluate and delineate defective areas

An image processing-based approach for detectingbugholes on concrete surface has been established by Liu andYang [21] the employed techniques are contrast enhance-ment and Otsu thresholding Kim et al [22] compareddifferent image binarisation algorithms for identifyingcracks in concrete structures Hoang [23] employed the Otsumethod and a gray intensity modification approach forbinarizing images and isolating cracks

Silva and Lucena [24] demonstrated the capability ofdeep learning approach for concrete crack detection Dor-afshan et al [25] has recently compared the performances ofdeep convolutional neural networks and edge detectionmethods for the task of recognizing concrete cracks )enotable advantage of deep learning approaches is that theirfeature extraction operators are autonomously constructedduring the model training phase [26] However methodsbased on deep learning often necessitate a considerableamount of training samples and require a large computa-tional cost

As can be seen from the existing literature most of theprevious works have dedicated in constructing models forthe classification of crack and noncrack conditions Few

studies have constructed an integrated image processingmodel for detect cracks and spalls )e objective of thecurrent study is to combine image processing techniquesand advanced machine learning algorithms into an in-tegrated model that is capable of detecting and categorizingthe defective areas on wall structures By using an intelligentmodel that can recognize and categorize types of cracks andspalling areas simultaneously the task of periodicbuilding condition survey can be executed in amore effectivemanner

)e feature extraction phase of the new model relies onimage processing algorithms of steerable filters and pro-jection integrals It is because these two algorithms of imageprocessing have demonstrated their usefulness in recog-nizing defects in pavements [27 28] Based on the extractedfeatures machine learning algorithms includingsupport vector machine (SVM) and least squares supportvector machine (LSSVM) are employed to classify inputimages into five labels longitudinal cracks transversecracks diagonal cracks spalls and intact walls )e reasonfor the selection of these machine learning approachesis that their outstanding performances in classificationtasks have been reported in the literature [29ndash33]

Based on the aforementioned features the main con-tributions of the newly constructed model can be summa-rized as follows

(i) Multiple types of cracks (longitudinal crackstransverse cracks diagonal cracks) existing in wallstructures can be detected and categorized in anintegrated model )is can be considered to bea significant improvement since most of the existingmodels can only produce the prediction outcome ofcrack or noncrack conditions [13 23ndash25]

(ii) Although projection integrals have been utilized instructural defect classification [27 28] diagonalprojection integrals which have been rarelyexploited in concrete surface crack categorizationare employed in this study to specifically deal withdiagonal cracks

(iii) Cracks and spall damages can be recognized si-multaneously in an integrated model which has alsorarely been achieved in the current literature

(iv) Neural networks have been extensively usedin concrete crack categorization [19 34 35]However the applications of SVM and LSSVM inthis task are still limited and a comparative workneeds to be performed to evaluate the potential ofthese two advance machine learning algorithmsin dealing with the problem of wall defectrecognition

)e rest of the paper is organized as follows the researchmethodology is reviewed briefly in the next section followedby the description of the newly constructed automatic ap-proach for wall defect detection the fourth section reportsexperimental results of this study followed by the conclu-sion in the final section






G(x y r) 12πr


2r2 (1)




2(θ)

(2)



2π

radicr5


2π

radicr5


2π

radicr5

(3)









I(i y)


I(x j)(5)




I(x y)


I(x y)(6)







(a)


(b)


(c)


(d)


(e)







Tw + c

12

1113944

N

k1e2k


(7)



K xk xl( 1113857 φ xk( 1113857Tφ xl( 1113857 (8)



2

2σ2⎛⎝ ⎞⎠ (9)




k1αk yk w

Tφ xk( 1113857 + b1113872 1113873minus1+ ek1113966 1113967

minus1113944N

k1vkek

(10)



zL

zw 0⟶ w 1113944

N


zL

zb 0⟶ 1113944

N

k1αkyk 0

zL

zek


⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(11)

y xy

(a)

x

yx

(b)


y DPI1 DPI2

x




maxα

JD(α) minus12sumN


N

k1αk

Subject to sumN


(12)




y(x) sign sumSV






12sumN

k1e2k


(15)



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2


(a)

HPIVPI

DPI1DPI2


0

02

04

06

08

1

Ave

rage

pix

el v

alue


(b)





0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)





N

k1αk yk w

Tφ xk( 1113857 + b1113872 1113873minus 1 + ek1113966 1113967

(16)



y(x) sign 1113944N

k1αkyiK xk xl( 1113857 + b⎛⎝ ⎞⎠ (17)













x1

Kernel mapping

Φ(xu)

Φ(xv)

Φ(x)


Label 1

Label 2

Label 2

Label 1

Hyperplane







(a)

(b)

(c)

(d)

(e)






Image acquisition

Data set

Model training

Model prediction


Steerable filters




HPI

VPI

DPIs



(a)

0 100 200 300 400 500 600 700 800Pixel interval

002040608

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(b)

0 10 20 30 40 50 60 70 80Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI 1DPI 2

(c)

0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)




0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

SF

Original image


HPIVPI

DPI1DPI2

(a)


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPI

DPI1DPI2

Original image

(b)





CARi RiCRiA

times 100() (18)







0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

0 100 200 300 400 500 600 700 800Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPIDPI1DPI2

HPIVPIDPI1DPI2


8127

8307

8513

81738040

LSSVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(a)

7780

8013

8433

8133

7833

SVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(b)






Average


Std



55

60

65

70

75

80

85

CAR

()

CARo



3035404550556065707580859095

100

CAR

()

CARLC

CARTC

CARDC

CARSD

CARIW



CAROverall sum5

i1

CARi5

(19)







50

55

60

65

70

75

80

85

90

95

CAR O






0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)



Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in








G(x y r) 12πr


2r2 (1)




2(θ)

(2)



2π

radicr5


2π

radicr5


2π

radicr5

(3)









I(i y)


I(x j)(5)




I(x y)


I(x y)(6)







(a)


(b)


(c)


(d)


(e)







Tw + c

12

1113944

N

k1e2k


(7)



K xk xl( 1113857 φ xk( 1113857Tφ xl( 1113857 (8)



2

2σ2⎛⎝ ⎞⎠ (9)




k1αk yk w

Tφ xk( 1113857 + b1113872 1113873minus1+ ek1113966 1113967

minus1113944N

k1vkek

(10)



zL

zw 0⟶ w 1113944

N


zL

zb 0⟶ 1113944

N

k1αkyk 0

zL

zek


⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(11)

y xy

(a)

x

yx

(b)


y DPI1 DPI2

x




maxα

JD(α) minus12sumN


N

k1αk

Subject to sumN


(12)




y(x) sign sumSV






12sumN

k1e2k


(15)



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2


(a)

HPIVPI

DPI1DPI2


0

02

04

06

08

1

Ave

rage

pix

el v

alue


(b)





0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)





N

k1αk yk w

Tφ xk( 1113857 + b1113872 1113873minus 1 + ek1113966 1113967

(16)



y(x) sign 1113944N

k1αkyiK xk xl( 1113857 + b⎛⎝ ⎞⎠ (17)













x1

Kernel mapping

Φ(xu)

Φ(xv)

Φ(x)


Label 1

Label 2

Label 2

Label 1

Hyperplane







(a)

(b)

(c)

(d)

(e)






Image acquisition

Data set

Model training

Model prediction


Steerable filters




HPI

VPI

DPIs



(a)

0 100 200 300 400 500 600 700 800Pixel interval

002040608

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(b)

0 10 20 30 40 50 60 70 80Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI 1DPI 2

(c)

0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)




0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

SF

Original image


HPIVPI

DPI1DPI2

(a)


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPI

DPI1DPI2

Original image

(b)





CARi RiCRiA

times 100() (18)







0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

0 100 200 300 400 500 600 700 800Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPIDPI1DPI2

HPIVPIDPI1DPI2


8127

8307

8513

81738040

LSSVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(a)

7780

8013

8433

8133

7833

SVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(b)






Average


Std



55

60

65

70

75

80

85

CAR

()

CARo



3035404550556065707580859095

100

CAR

()

CARLC

CARTC

CARDC

CARSD

CARIW



CAROverall sum5

i1

CARi5

(19)







50

55

60

65

70

75

80

85

90

95

CAR O






0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)



Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in





(a)


(b)


(c)


(d)


(e)







Tw + c

12

1113944

N

k1e2k


(7)



K xk xl( 1113857 φ xk( 1113857Tφ xl( 1113857 (8)



2

2σ2⎛⎝ ⎞⎠ (9)




k1αk yk w

Tφ xk( 1113857 + b1113872 1113873minus1+ ek1113966 1113967

minus1113944N

k1vkek

(10)



zL

zw 0⟶ w 1113944

N


zL

zb 0⟶ 1113944

N

k1αkyk 0

zL

zek


⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(11)

y xy

(a)

x

yx

(b)


y DPI1 DPI2

x




maxα

JD(α) minus12sumN


N

k1αk

Subject to sumN


(12)




y(x) sign sumSV






12sumN

k1e2k


(15)



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2


(a)

HPIVPI

DPI1DPI2


0

02

04

06

08

1

Ave

rage

pix

el v

alue


(b)





0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)





N

k1αk yk w

Tφ xk( 1113857 + b1113872 1113873minus 1 + ek1113966 1113967

(16)



y(x) sign 1113944N

k1αkyiK xk xl( 1113857 + b⎛⎝ ⎞⎠ (17)













x1

Kernel mapping

Φ(xu)

Φ(xv)

Φ(x)


Label 1

Label 2

Label 2

Label 1

Hyperplane







(a)

(b)

(c)

(d)

(e)






Image acquisition

Data set

Model training

Model prediction


Steerable filters




HPI

VPI

DPIs



(a)

0 100 200 300 400 500 600 700 800Pixel interval

002040608

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(b)

0 10 20 30 40 50 60 70 80Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI 1DPI 2

(c)

0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)




0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

SF

Original image


HPIVPI

DPI1DPI2

(a)


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPI

DPI1DPI2

Original image

(b)





CARi RiCRiA

times 100() (18)







0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

0 100 200 300 400 500 600 700 800Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPIDPI1DPI2

HPIVPIDPI1DPI2


8127

8307

8513

81738040

LSSVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(a)

7780

8013

8433

8133

7833

SVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(b)






Average


Std



55

60

65

70

75

80

85

CAR

()

CARo



3035404550556065707580859095

100

CAR

()

CARLC

CARTC

CARDC

CARSD

CARIW



CAROverall sum5

i1

CARi5

(19)







50

55

60

65

70

75

80

85

90

95

CAR O






0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)



Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in








Tw + c

12

1113944

N

k1e2k


(7)



K xk xl( 1113857 φ xk( 1113857Tφ xl( 1113857 (8)



2

2σ2⎛⎝ ⎞⎠ (9)




k1αk yk w

Tφ xk( 1113857 + b1113872 1113873minus1+ ek1113966 1113967

minus1113944N

k1vkek

(10)



zL

zw 0⟶ w 1113944

N


zL

zb 0⟶ 1113944

N

k1αkyk 0

zL

zek


⎧⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎪⎩

(11)

y xy

(a)

x

yx

(b)


y DPI1 DPI2

x




maxα

JD(α) minus12sumN


N

k1αk

Subject to sumN


(12)




y(x) sign sumSV






12sumN

k1e2k


(15)



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2


(a)

HPIVPI

DPI1DPI2


0

02

04

06

08

1

Ave

rage

pix

el v

alue


(b)





0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)





N

k1αk yk w

Tφ xk( 1113857 + b1113872 1113873minus 1 + ek1113966 1113967

(16)



y(x) sign 1113944N

k1αkyiK xk xl( 1113857 + b⎛⎝ ⎞⎠ (17)













x1

Kernel mapping

Φ(xu)

Φ(xv)

Φ(x)


Label 1

Label 2

Label 2

Label 1

Hyperplane







(a)

(b)

(c)

(d)

(e)






Image acquisition

Data set

Model training

Model prediction


Steerable filters




HPI

VPI

DPIs



(a)

0 100 200 300 400 500 600 700 800Pixel interval

002040608

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(b)

0 10 20 30 40 50 60 70 80Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI 1DPI 2

(c)

0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)




0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

SF

Original image


HPIVPI

DPI1DPI2

(a)


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPI

DPI1DPI2

Original image

(b)





CARi RiCRiA

times 100() (18)







0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

0 100 200 300 400 500 600 700 800Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPIDPI1DPI2

HPIVPIDPI1DPI2


8127

8307

8513

81738040

LSSVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(a)

7780

8013

8433

8133

7833

SVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(b)






Average


Std



55

60

65

70

75

80

85

CAR

()

CARo



3035404550556065707580859095

100

CAR

()

CARLC

CARTC

CARDC

CARSD

CARIW



CAROverall sum5

i1

CARi5

(19)







50

55

60

65

70

75

80

85

90

95

CAR O






0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)



Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in





maxα

JD(α) minus12sumN


N

k1αk

Subject to sumN


(12)




y(x) sign sumSV






12sumN

k1e2k


(15)



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2


(a)

HPIVPI

DPI1DPI2


0

02

04

06

08

1

Ave

rage

pix

el v

alue


(b)





0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)





N

k1αk yk w

Tφ xk( 1113857 + b1113872 1113873minus 1 + ek1113966 1113967

(16)



y(x) sign 1113944N

k1αkyiK xk xl( 1113857 + b⎛⎝ ⎞⎠ (17)













x1

Kernel mapping

Φ(xu)

Φ(xv)

Φ(x)


Label 1

Label 2

Label 2

Label 1

Hyperplane







(a)

(b)

(c)

(d)

(e)






Image acquisition

Data set

Model training

Model prediction


Steerable filters




HPI

VPI

DPIs



(a)

0 100 200 300 400 500 600 700 800Pixel interval

002040608

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(b)

0 10 20 30 40 50 60 70 80Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI 1DPI 2

(c)

0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)




0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

SF

Original image


HPIVPI

DPI1DPI2

(a)


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPI

DPI1DPI2

Original image

(b)





CARi RiCRiA

times 100() (18)







0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

0 100 200 300 400 500 600 700 800Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPIDPI1DPI2

HPIVPIDPI1DPI2


8127

8307

8513

81738040

LSSVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(a)

7780

8013

8433

8133

7833

SVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(b)






Average


Std



55

60

65

70

75

80

85

CAR

()

CARo



3035404550556065707580859095

100

CAR

()

CARLC

CARTC

CARDC

CARSD

CARIW



CAROverall sum5

i1

CARi5

(19)







50

55

60

65

70

75

80

85

90

95

CAR O






0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)



Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in






0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2



0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)





N

k1αk yk w

Tφ xk( 1113857 + b1113872 1113873minus 1 + ek1113966 1113967

(16)



y(x) sign 1113944N

k1αkyiK xk xl( 1113857 + b⎛⎝ ⎞⎠ (17)













x1

Kernel mapping

Φ(xu)

Φ(xv)

Φ(x)


Label 1

Label 2

Label 2

Label 1

Hyperplane







(a)

(b)

(c)

(d)

(e)






Image acquisition

Data set

Model training

Model prediction


Steerable filters




HPI

VPI

DPIs



(a)

0 100 200 300 400 500 600 700 800Pixel interval

002040608

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(b)

0 10 20 30 40 50 60 70 80Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI 1DPI 2

(c)

0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)




0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

SF

Original image


HPIVPI

DPI1DPI2

(a)


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPI

DPI1DPI2

Original image

(b)





CARi RiCRiA

times 100() (18)







0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

0 100 200 300 400 500 600 700 800Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPIDPI1DPI2

HPIVPIDPI1DPI2


8127

8307

8513

81738040

LSSVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(a)

7780

8013

8433

8133

7833

SVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(b)






Average


Std



55

60

65

70

75

80

85

CAR

()

CARo



3035404550556065707580859095

100

CAR

()

CARLC

CARTC

CARDC

CARSD

CARIW



CAROverall sum5

i1

CARi5

(19)







50

55

60

65

70

75

80

85

90

95

CAR O






0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)



Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in






N

k1αk yk w

Tφ xk( 1113857 + b1113872 1113873minus 1 + ek1113966 1113967

(16)



y(x) sign 1113944N

k1αkyiK xk xl( 1113857 + b⎛⎝ ⎞⎠ (17)













x1

Kernel mapping

Φ(xu)

Φ(xv)

Φ(x)


Label 1

Label 2

Label 2

Label 1

Hyperplane







(a)

(b)

(c)

(d)

(e)






Image acquisition

Data set

Model training

Model prediction


Steerable filters




HPI

VPI

DPIs



(a)

0 100 200 300 400 500 600 700 800Pixel interval

002040608

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(b)

0 10 20 30 40 50 60 70 80Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI 1DPI 2

(c)

0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)




0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

SF

Original image


HPIVPI

DPI1DPI2

(a)


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPI

DPI1DPI2

Original image

(b)





CARi RiCRiA

times 100() (18)







0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

0 100 200 300 400 500 600 700 800Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPIDPI1DPI2

HPIVPIDPI1DPI2


8127

8307

8513

81738040

LSSVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(a)

7780

8013

8433

8133

7833

SVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(b)






Average


Std



55

60

65

70

75

80

85

CAR

()

CARo



3035404550556065707580859095

100

CAR

()

CARLC

CARTC

CARDC

CARSD

CARIW



CAROverall sum5

i1

CARi5

(19)







50

55

60

65

70

75

80

85

90

95

CAR O






0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)



Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in








(a)

(b)

(c)

(d)

(e)






Image acquisition

Data set

Model training

Model prediction


Steerable filters




HPI

VPI

DPIs



(a)

0 100 200 300 400 500 600 700 800Pixel interval

002040608

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(b)

0 10 20 30 40 50 60 70 80Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI 1DPI 2

(c)

0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)




0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

SF

Original image


HPIVPI

DPI1DPI2

(a)


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPI

DPI1DPI2

Original image

(b)





CARi RiCRiA

times 100() (18)







0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

0 100 200 300 400 500 600 700 800Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPIDPI1DPI2

HPIVPIDPI1DPI2


8127

8307

8513

81738040

LSSVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(a)

7780

8013

8433

8133

7833

SVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(b)






Average


Std



55

60

65

70

75

80

85

CAR

()

CARo



3035404550556065707580859095

100

CAR

()

CARLC

CARTC

CARDC

CARSD

CARIW



CAROverall sum5

i1

CARi5

(19)







50

55

60

65

70

75

80

85

90

95

CAR O






0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)



Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in







Image acquisition

Data set

Model training

Model prediction


Steerable filters




HPI

VPI

DPIs



(a)

0 100 200 300 400 500 600 700 800Pixel interval

002040608

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(b)

0 10 20 30 40 50 60 70 80Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI 1DPI 2

(c)

0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

HPIVPI

DPI1DPI2

(d)




0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

SF

Original image


HPIVPI

DPI1DPI2

(a)


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPI

DPI1DPI2

Original image

(b)





CARi RiCRiA

times 100() (18)







0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

0 100 200 300 400 500 600 700 800Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPIDPI1DPI2

HPIVPIDPI1DPI2


8127

8307

8513

81738040

LSSVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(a)

7780

8013

8433

8133

7833

SVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(b)






Average


Std



55

60

65

70

75

80

85

CAR

()

CARo



3035404550556065707580859095

100

CAR

()

CARLC

CARTC

CARDC

CARSD

CARIW



CAROverall sum5

i1

CARi5

(19)







50

55

60

65

70

75

80

85

90

95

CAR O






0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)



Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in





0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

SF

Original image


HPIVPI

DPI1DPI2

(a)


0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPI

DPI1DPI2

Original image

(b)





CARi RiCRiA

times 100() (18)







0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

0 100 200 300 400 500 600 700 800Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPIDPI1DPI2

HPIVPIDPI1DPI2


8127

8307

8513

81738040

LSSVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(a)

7780

8013

8433

8133

7833

SVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(b)






Average


Std



55

60

65

70

75

80

85

CAR

()

CARo



3035404550556065707580859095

100

CAR

()

CARLC

CARTC

CARDC

CARSD

CARIW



CAROverall sum5

i1

CARi5

(19)







50

55

60

65

70

75

80

85

90

95

CAR O






0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)



Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in






CARi RiCRiA

times 100() (18)







0 5 10 15 20 25 30 35 40Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs

0 100 200 300 400 500 600 700 800Pixel interval

0

02

04

06

08

1

Ave

rage

pix

el v

alue

PIs


HPIVPIDPI1DPI2

HPIVPIDPI1DPI2


8127

8307

8513

81738040

LSSVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(a)

7780

8013

8433

8133

7833

SVM

1 15 2 25 3r

75

80

85

90

Mod

el p

erfo

rman

ce

(b)






Average


Std



55

60

65

70

75

80

85

CAR

()

CARo



3035404550556065707580859095

100

CAR

()

CARLC

CARTC

CARDC

CARSD

CARIW



CAROverall sum5

i1

CARi5

(19)







50

55

60

65

70

75

80

85

90

95

CAR O






0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)



Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in







Average


Std



55

60

65

70

75

80

85

CAR

()

CARo



3035404550556065707580859095

100

CAR

()

CARLC

CARTC

CARDC

CARSD

CARIW



CAROverall sum5

i1

CARi5

(19)







50

55

60

65

70

75

80

85

90

95

CAR O






0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)



Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in




CAROverall sum5

i1

CARi5

(19)







50

55

60

65

70

75

80

85

90

95

CAR O






0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)



Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in





0

05

1

Aver

age p

ixel

valu

e

PIs

HPIVPI

DPI1DPI2

SF

Image

(a)

(b)

(c)

(d)

(e)



Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05


Aver

age p

ixel

valu

e

HPIVPI

DPI1DPI2


0

05











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in











6 Conclusion


Data Availability






References





































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in



































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in


































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in









Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in




Documents

ImageProcessing-BasedRecognitionofWallDefectsUsing ...downloads.hindawi.com/journals/cin/2018/7913952.pdf · ResearchArticle ImageProcessing-BasedRecognitionofWallDefectsUsing MachineLearningApproachesandSteerableFilters