ImageProcessing-BasedDetectionofPipeCorrosionUsing...

Research ArticleImage Processing-Based Detection of Pipe Corrosion UsingTexture Analysis and Metaheuristic-Optimized MachineLearning Approach

Nhat-Duc Hoang 1 and Van-Duc Tran 2

1Lecturer Faculty of Civil Engineering Institute of Research and Development Duy Tan University R809ndashNo03 Quang TrungDa Nang 550000 Vietnam2Lecturer International School Duy Tan University 254 Nguyen Van Linh Danang 550000 Vietnam

Correspondence should be addressed to Nhat-Duc Hoang hoangnhatducdtueduvn

Received 27 March 2019 Revised 21 May 2019 Accepted 17 June 2019 Published 11 July 2019

Academic Editor Juan A Gomez-Pulido

Copyright copy 2019 Nhat-Duc Hoang and Van-Duc Tran is is an open access article distributed under the Creative CommonsAttribution License which permits unrestricted use distribution and reproduction in anymedium provided the original work isproperly cited

To maintain the serviceability of buildings the owners need to be informed about the current condition of the water supply andwaste disposal systems erefore timely and accurate detection of corrosion on pipe surface is a crucial task e conventionalmanual surveying process performed by human inspectors is notoriously time consuming and labor intensive Hence this studyproposes an image processing-based method for automating the task of pipe corrosion detection Image texture includingstatistical measurement of image colors gray-level co-occurrence matrix and gray-level run length is employed to extract featuresof pipe surface Support vector machine optimized by differential flower pollination is then used to construct a decision boundarythat can recognize corroded and intact pipe surfaces A dataset consisting of 2000 image samples has been collected and utilized totrain and test the proposed hybrid model Experimental results supported by the Wilcoxon signed-rank test confirm that theproposed method is highly suitable for the task of interest with an accuracy rate of 9281us the model proposed in this studycan be a promising tool to assist building maintenance agents during the phase of pipe system survey

1 Introduction

In high-rise building maintenance an important objective isconcerned with the integrity of the water supply system andprevention of water contamination Cast iron is widely usedin water supply and waste disposal systems due to the ad-vantage of high strength Since stainless steel pipes often fallout of favor in domestic pipework because of their highexpenses [1] corrosion is a widely observed type of struc-tural damage

Corrosion (see Figure 1) can be defined as a chemicalprocess caused by chemical and electrochemical reactionsis phenomenon is typically observed in environmentalconditions featuring a high level of moisture ere aredifferent kinds of corrosion such as general corrosion whichoccurs as uniformly distributed nonprotective flakes of rustand pitting which is a localized point of corrosive attack [2]

Corrosion brings about the destruction of metal pipeworksurface and consequently leads to reduction in pipe servicelife and increase in building maintenance cost [3] In certaincase this defect may strongly affect the health of buildingoccupants due to deterioration of water quality uscorrosion should be identified timely by means of periodicsurveys to ensure the integrity of pipe systems and establishcost-effective maintenance strategies

In Vietnam as well as in many other countries manualmethods performed by human inspectors are commonlyemployed for condition assessment of water supplywastedisposal systems As clearly pointed out by Liu et al [4] andAtha and Jahanshahi [5] these manual approaches are laborintensive and time consuming Corroded regions can beneglected in positions of pipe system that are difficult toreach and observe visually Moreover the processes of dataprocessing and reporting are also very tedious for human

HindawiComputational Intelligence and NeuroscienceVolume 2019 Article ID 8097213 13 pageshttpsdoiorg10115520198097213

technicians erefore there is a practical need to come upwith a more productive and accurate method of pipe con-dition survey

Although there is a wide range of existing pipe in-spection approaches (such as magnetic flux leakage ul-trasonic testing and external corrosion direct assessment)all of these methods have limitations including highequipment cost restricted range of inspection and in-capability of detecting small pitting regions [3] Consid-ering the large amount of pipe systems needed to besurveyed and the limited access to sophisticated equipmentin developing countries there is an urgent need for aproductive and low-cost solution for periodic surveys ofpipe system condition Recently digital image processinghas gained a great attention within the field of structuralheath monitoring [6 7]

Particularly image processing techniques can be effec-tively employed to investigate the outer surface for detectingdefects on pipes or other metal structures including cor-rosion and cracks [8] Itzhak et al [9] relied on statisticmeasurement of image pixels to quantify pitting corrosionChoi and Kim [10] identified corrosion based on themorphology of the corroded surface features of image colortexture and shape are employed for corrosion recognitionA model for classifying corroded and noncorroded surfacesusing texture descriptors obtained from gray level co-oc-currence matrix and image color has been proposed inMedeiros et al [11]

A method based on watershed segmentation has beenemployed in [12] for rating of corrosion defects the per-centage area of corroded region was used for determiningthe grade of defects Idris and Jafar [13] used image filter-based image enhancement and neural network for corrosioninspection Son et al [14] proposed a model based on de-cision tree algorithm for identifying rusted surface area of

steel bridge A model based on image color analysis andK-means clustering for bridge rust identification has beenconstructed and verified by Liao and Lee [15]

Petricca et al [16] compared standard computer visiontechniques and deep neural network for rust and nonrustdetection Deep neural networks have also been employedfor corrosion detection by Liu et al [4] and Atha andJahanshahi [5] Safari and Shoorehdeli [17] applied artificialneural network Gabor filter and entropy filter for pipedefect detection Cheriet et al [18] incorporated expertknowledge and field data to construct a knowledge-basedsystem for assessing corrosive damage on metallic pipeconduits Gibbons et al [19] relied on a Gaussian mixturemodel for probabilistic classification of corroded andnoncorroded areas Bondada et al [3] detected and quan-titatively assessed corrosion damages on pipelines bycomputing the mean of saturation value of image pixels byimage analysis the corroded areas on pipelines can besegmented

From the above literature it can be seen that imageprocessing and machine learning have been a feasible al-ternative for replacing the tedious process of manual surveyBased on a recent review work of Ahuja and Shukla [20]there is an increasing trend of applying computer visiontechniques for corrosion detection Moreover due to theimportance of the research theme exploring other imageprocessing and machine learning methods used for pipecorrosion detection can be highly meaningful in both aca-demic and practical aspects

As reported in the literature although image textureanalysis has been applied few previous studies haveemployed a combination of image texture descriptors forpipe corrosion recognition Hence this study is an attemptto fill this gap in the literature by proposing a method usedfor analyzing texture of water pipe surface that integrates

(a) (b)

(c) (d)

Figure 1 Corroded areas on pipe surface

2 Computational Intelligence and Neuroscience

statistical measurement of color channels gray-level co-occurrence matrix and gray-level run length matrix Basedon the features extracted by the above texture descriptorsthe support vector machine (SVM) [21] is employed tocategorize image samples into two classes noncorrosion(negative) and corrosion (positive) SVM is utilized in thisstudy due to the fact that it has been confirmed to be a robusttool for pattern classification in various studies [22ndash24] Inaddition to optimize the training process of SVM-basedcorrosion detection model differential flower pollination(DFP) metaheuristic is employed A dataset consisting of2000 image samples has been collected to train and verify theproposed method

e rest of the study is organized as follows Section 2reviews the research material and methods used to con-struct the water pipe corrosion detection approach Sec-tion 3 reports experimental results and discussionsSection 4 provides several concluding remarks of thisstudy

2 Material and Methods

21 Image Texture Analysis Identifying corroded areasbased on two-dimensional image samples is a challengingtask due to the complex and deceptive features of pipesurfaces containing various irregular objects such as dirtand paints erefore using information provided by onepixel is definitely not sufficient for corrosion detection It isbecause a pixel having similar color values can belong toboth categories of noncorrosion and corrosion Hencetexture information extracted from a certain region of pipesurface can be used for recognizing the defect of interestis section of the study describes the employed texturedescriptors used for computing the features of water pipesurface

211 Statistical Properties of Color Channels Herein thestatistical properties of three color channels (red green andblue) of an image sample can be employed to representimage texture us an image is described in a RGB colorspace It is noted that besides RGB there are other colorspaces such as HSV which can also be useful in the task ofcorrosion detection However in this study we rely on theoriginal RGB color model obtained from the employeddigital camera Let I be a variable representing the colorlevels of an image sample e first-order histogram P(I) iscalculated as follows [25]

Pc(I) NIc

W times H (1)

where c denotes a color channelNIc is the number of pixelswith intensity value I of the channel c and H and Wrepresent the height and width of an image samplerespectively

us the mean (μc) standard deviation (σc) skewness(δc) kurtosis (ηc) entropy (ρc) and range (Δc) of color valueare calculated as follows

μc 1113944NLminus1

i0Iic times Pc(I)

σc

1113944

NLminus1

i0Iic minus μc1113872 1113873

2times Pc(I)

11139741113972

δc 1113936

NLminus1i0 Iic minus μc1113872 1113873

3times Pc(I)

σ3c

ηc 1113936


4times Pc(I)

σ4c

ρc minus1113944NLminus1

i0Pc(I) times log2 Pc(I)( 1113857

Δc Max Ic( 1113857minusMin Ic( 1113857

(2)

where NL 256 denotes the number of discrete color values

212 Gray-Level Co-Occurrence Matrix (GLCM) eGLCM [26] is also a commonly used texture descriptor Toemploy this technique a color image must first be convertedto a gray scale one e GLCM discriminates different imagetextures based on the repeated occurrence of some gray-levelpatterns existing in the texture [27] Let δ (r θ) be a vectorin the polar coordinates of an image sample For each δ thejoint probability of the pairs of gray levels that occur at thetwo points separated by the relationship δ is computed [28]is joint probability is compactly displayed in a GLCM Pδwithin which Pδ(i j) represents the probability of the twogray levels of i and j occurring according to δ e originalPδ(i j) is often normalized via the following equation

PNδ (i j)

Pδ(i j)

SP (3)

where PNδ (i j) denotes the normalized GLCM and SP is the

number of pixelsBased on the suggestion of Haralick et al [29] four

GLCMs with r 1 and θ 0deg 45deg 90deg and 135deg can beestablished Accordingly angular second moment (AM)contrast (CO) correlation (CR) and entropy (ET) for eachmatrix can be computed to serve as texture descriptors asfollows [28 29]

AM 1113944

Ng

i11113944

Ng

j1P

Nδ (i j)

2

CO 1113944

Ngminus1

k0k2

1113944

Ng

i1|iminusj|k

1113944

Ng

j1P

Nδ (i j)

CR 1113936

Ngi11113936

Ngj1i times j times PN

δ (i j)minus μXμY

σXσY

ET minus1113944

Ng

i11113944

Ng

j1P

Nδ (i j)log P

Nδ (i j)1113872 1113873

(4)

Computational Intelligence and Neuroscience 3

where Ng is the number of gray-level values and μX μY σXand σY are the means and standard deviations of the marginaldistribution associated with a normalized GLCM [29]

213 Gray-Level Run Lengths (GLRL) GLRL is a texturedescription method proposed by Galloway [30] ismethod is highly effective in discriminating textures fea-turing different fineness and has been successfully applied invarious fields of study [31 32] It is because GLRL is con-structed based on the fact that relatively long gray-level runsare observed more frequently in a coarse texture and a finetexture typically has more short runs [33] A run-lengthmatrix p( i middot j ) in a certain direction is defined as the numberof times that a run length j of gray level i is observed [30]

Using this matrix the short run emphasis (SRE) longrun emphasis (LRE) gray-level nonuniformity (GLN) runlength nonuniformity (RLN) and run percentage (RP)[30 33] can be computed Additionally Chu et al [34] putforward the indices of low gray-level run emphasis (LGRE)and high gray-level run emphasis (HGRE) Dasarathy andHolder [35] proposed to compute the short run low gray-level emphasis (SRLGE) short run high gray-level em-phasis (SRHGE) long run low gray-level emphasis(LRLGE) and long run high gray-level emphasis (LRHGE)e above indices are summarized in Table 1 It is notedthat one run length matrix is computed for each of di-rection in the set of 0deg 45deg 90deg 135deg and each matrixresults in 11 GLRL-based features erefore the totalnumber of features obtained from GLRL matrices is11times 4 44

22 Computational Intelligence Methods

221 Support Vector Machine (SVM) SVM described in[21] is a robust pattern recognition method established onthe theory of statistical learning Given the task at hand is toclassify a set of input feature xk into two categories of yk minus1(noncorrosion) and yk +1 (corrosion) a SVM modelconstructs a decision surface that separates the input spaceinto two distinctive regions characterizing the two differenttwo categories e SVM algorithm aims at identifying adecision boundary so that the gap between classes is as largeas possible [36] In addition SVM employs the kernel trickto convert a nonlinear classification task into a linear one ASVMmodel first maps the input data from the original spaceto a high-dimensional feature space within which the datacan be separated by a hyperplane (see Figure 2)

e SVM training process can be formulated as thefollowing constrained optimization problem [36]

minimize Jp(w e) 12w

Tw + c

12

1113944

N

k1e2k

subjected to yk wTφ xk( 1113857 + b( 1113857ge 1minus ek k 1 N ek ge 0

(5)

where w isinRn is a normal vector to the classification hy-perplane and b isinR is the model bias ekge 0 is called a slack

variable c denotes a penalty constant and φ(x) is a non-linear mapping from the input space to the high-di-mensional feature space

During the construction of a SVM model it is not re-quired to obtain the explicit form of φ(x) Instead of thatonly the dot product of φ(x) in the input space is requiredand expressed via a kernel function shown as follows

K xk xl( 1113857 φ xk( 1113857Tφ xl( 1113857 (6)

e radial basis function (RBF) kernel function [37] isoften employed for data classification its functional form isgiven below

K xk xl( 1113857 exp minusxk minusxl

2

2σ2⎛⎝ ⎞⎠ (7)

where σ is a free parameterAccordingly a SVMmodel used for data classification is

given compactly as follows

y xl( 1113857 sign 1113944

SV

k1αkykK xk xl( 1113857 + b⎛⎝ ⎞⎠ (8)

where αk denotes the solution of the dual form of theaforementioned constrained optimization SV is the numberof support vectors which is the number of αk gt 0

222 Differential Flower Pollination (DFP) As shown in theprevious section the model training and prediction phasesof a SVM model depend on a proper selection of itshyperparameters including the penalty coefficient (c) andthe kernel function parameter (σ) e first hyperparameteraffects the penalty imposed on data samples deviating fromthe established decision surface the later hyperparameterspecifies the smoothness of the decision surface Since thisproblem of hyperparameter selection can be formulated asan optimization problem [38ndash40] this study employs theDFP metaheuristic to optimize the model training phase ofSVM

DFP proposed in [41] is a population-based meta-heuristic that combines the advantages of the standard al-gorithms of differential evolution (DE) [42] and flowerpollination algorithm (FPA) [43] e employed hybridmetaheuristic consists of three main steps initialization ofpopulation members alteration of member locations andcost function evaluation Each member of the DFP meta-heuristic is presented as a numerical vector consisting of thetwo SVM hyperparameters In the first step all populationmembers are randomly generated within the feasible do-main In the second step the location of populationmembers is altered by local and global search phases In thenext step the cost function of each member is computed anda greedy selection operator is performed to update the lo-cation of the DFPrsquos population

e second step of the DFP includes the FPA-basedglobal pollination operator and the DE-based local polli-nation operator A switching probability p 08 is used togovern the frequencies of these two operators [43]e FPA-


based global pollination and the DE-based local pollinationoperators are presented as follows

(i) e FPA-based global pollination

Xtriali X

gDFPi + L middot X

gDFPi minusXbest( 1113857 (9)

where g is the index of the current generation Xtriali

is a trial solution XgDFPi denotes a solution of the

current population Xbest represents the best solu-tion and L denotes a random number generatedfrom the Levy distribution [43]

(ii) e DE-based local pollination modifies the currentmember by creating a mutated flower and a crossedflower according to the following equations

(a) Creating a mutated flower

XmutatedigDFP

xr1gDFP + F middot xr2gDFP minusxr3gDFP1113872 1113873 (10)

where r1 r2 and r3 are three random integers and Fdenotes a mutation scale factor which is drawn froma Gaussian distribution with the mean 05 and thestandard deviation 015 [41]

(b) Creating a crossed flower

XcrossedjigDFP+1

Xmutatedji if r and jleCr or j rnb(i)

XjigDFP if r and jgtCr and jne rnb(i)

⎧⎨

⎩

(11)

where Cr 08 is the crossover probability [44]

23 Collected Image Samples Because SVM is a supervisedmachine learning algorithm a dataset consisting of 2000image samples of pipe surface with the ground truth labelhas been collected to construct the SVM-based corrosiondetection model It is proper to note that the numbers ofimage samples in the two labels of noncorrosion (negativeclass) and corrosion (positive class) are both 1000 edigital image samples have been collected during surveys ofseveral high-rise buildings in Danang city (Vietnam) eused digital camera is the 18-megapixel resolution CanonEOSM10 and the images were manually acquired by humaninspectors

Accordingly image samples of the two labels of non-corrosion (labelminus1) and corrosion (label+1) have been

Texture descriptor xi

Kernel mappingΦ(x)

Φ(xu)

Φ(xv)

Input space Feature space

Text

ure d

escr

ipto

r xj

Φ(xl)

CorrosionNoncorrosion

Figure 2 e data classification process of a SVM model

Table 1 Texture descriptors using GLRL

Descriptor Notation EquationShort run emphasis SRE SRE 1Nr1113936

Mi11113936

Nj1p(i j)j2

Long run emphasis LRE LRE 1Nr1113936Mi11113936

Nj1p(i j) times j2

Gray-level nonuniformity GLN GLN 1Nr1113936Mi1(1113936

Nj1p(i j)2)

Run length nonuniformity RLN RLN 1Nr1113936Nj1(1113936

Mi1p(i j)2)

Run percentage RP RP NrNp

Low gray-level run emphasis LGRE LGRE 1Nr1113936Nj11113936

Mi1p(i j)i2

High gray-level run emphasis HGRE HGRE 1Nr1113936Nj11113936

Mi1p(i j) times i2

Short run low gray-level emphasis SRLGE SRLGE 1Nr1113936Nj11113936

Mi1p(i j)(i2 times j2)

Short run high gray-level emphasis SRHGE SRHGE 1Nr1113936Nj11113936

Mi1(p(i j) times i2)j2

Long run low gray-level emphasis LRLGE LRLGE 1Nr1113936Nj11113936

Mi1(p(i j) times j2)i2

Long run high gray-level emphasis LRHGE LRHGE 1Nr1113936Nj11113936

Mi1p(i j) times i2 times j2

NoteM andN are the number of gray levels and the maximum run length respectively LetNr andNp be the total number of runs and the number of pixels inthe image respectively


prepared for SVM-based classification process In order toaccelerate the texture computation process the size of imagesamples has been set to be 50times 50 pixels Hence imagecropping operation is performed to generate the imagesamples used to train the SVM model e collected imageset is illustrated in Figure 3

24 Proposed Hybridization of Image Processing and Meta-heuristic-Optimized SVM for Pipe Corrosion Detectionis section of the study describes the structure of the newlydeveloped hybrid model of image processing and meta-heuristic-optimized SVM for pipe corrosion detection eproposed model named as MO-SVM-PCD is a combinationof image texture analysis and a metaheuristic-optimizedmachine learning approach As mentioned earlier the sta-tistical measurements of color channels GLCM and GLRLare used to extract texture-based features from image samplese hybrid model relies on SVM to classify data samples intothe categories of noncorrosion and corrosion In addition theDFP metaheuristic is employed to optimize the SVM-basedtraining and prediction phases e overall structure of theMO-SVM-PCD model is shown in Figure 4 e modelstructure can be divided into two separated modules com-putation of image texture and data classification based onSVM e first module is constructed in Visual CNET thesecond module is developed in MATLAB

Within the first module the image texture descriptorsbased on statistical analysis of color channels GLCM andGLRL compute numerical features from image samples Foreach of the three color channels (red green and blue) sixstatistical measurements of mean standard deviationskewness kurtosis entropy and range are calculated Hencethe total number of numerical features extracted from theaforementioned statistical indices of an image sample is6times 318 Subsequently the group of features extractedfrom the four co-occurrence matrices corresponding to thedirections of 0deg 45deg 90deg and 135deg is computed Because fourindices of the angular second moment contrast correlationentropy are acquired from one co-occurrence matrix thetotal number of GLCM-based features is 4times 416

In addition each GLRL matrix yields 11 properties ofSRE LRE GLN RLN RP LGRE HGRE SRLGE SRHGELRLGE and LRHGE us as stated earlier the number ofGLRL-based features is 4times11 44 Accordingly each imagesample is characterized by a feature vector having18 + 16 + 44 78 elementsis module can compute textureof one image for illustration purpose and can extract featuresfrom a batch of image samples to construct the training andtesting numerical datasets

When the module of feature computation is accom-plished a dataset consisting of 2000 data samples and 78input features is ready for further analysis is dataset hastwo class outputs minus1 meaning noncorrosion (negative class)and +1 meaning corrosion (positive class) In addition forstandardizing the data ranges and enhancing the datamodeling process the numerical dataset is preprocessed bythe Z-score data normalization [45] e equation of theZ-score data normalization is given as follows

XZN Xo minusmX

sX

(12)

where Xo and XZN represent an original and a normalizedinput variable respectively and mX and sX denote the meanand the standard deviation of the original input variablerespectively

Subsequently the normalized dataset is randomly di-vided into two subsets a training set (70) and a testing set(30) e first data subset is employed for model trainingthe later data subset is reserved for model testing etraining dataset is employed by the SVM-based data clas-sification module to generalize a corrosion recognitionmodel In addition DFP is utilized to finetune the SVMmodel hyperparameters including the penalty coefficientand the RBF kernel parameter It is worth mentioning thatthe SVM model operates via the help of the MATLABrsquosStatistics and Machine Learning Toolbox [46] in additionthe DFP and the hybridization of DFP and SVMmodel havebeen constructed in MATLAB by the authors

As shown in Figure 4 the two SVM hyperparameters arerandomly initialized at the first generation (g 1) Using thelocal and global pollination operators the DFP algorithmgradually guides the population of SVM hyperparameters toexplore the search space and identify better solutions Basedon the guidance of parameter setting in previous studies[44 47] the population size and the number of DFPsearching generations are selected to be 12 and 100 efeasible domain of the SVMrsquos penalty coefficient and kernelparameter is [1 100] and [01 100] respectively In the phaseof solution evaluation the quality of each member in thepopulation is appraised via the following cost function

fCF 1113936

Kk12 PPVk + NPVk( 1113857

K (13)

where K 5 denotes the number of data folds and PPV andNPV are the positive predictive value and the negativepredictive value PPV and NPV are employed to express themodel performance associated with a set of SVMhyperparameters

PPV and NPV are computed according to the followingequations [48]

PPV TP

TP + FP

NPV TN

TN + FN

(14)

where TP TN FP and FN are the true positive truenegative false positive and false negative valuesrespectively

It is noted that to compute the modelrsquos cost function aK-fold cross validation process with K 5 is employedUsing this cross fold validation the original dataset isseparated into 5 mutually exclusive subsets Accordinglythe SVM model training and evaluation is repeated 5times In each time 4 subsets are utilized for modeltraining and one subset is used for model validation eoverall model performance is obtained via averaging


predictive outcomes of the 5 data folds is process hasbeen proved to be a robust method for model hyper-parameter selection [49] Notably in each generationbased on the computed cost function the location ofpopulation members is updated and the stopping criterionis checked to verify whether the current generationnumber exceeds the allowable value If the stopping cri-terion is met the DFP-based optimization process ter-minates and the optimized SVM model is ready to predictcorrosion status for novel image samples

3 Experimental Results and Discussion

As stated earlier the dataset featuring 2000 samples and 78image texture variables has been separated into the training

and testing subset e training and testing subsets occupy70 and 30 of the original dataset respectively e firstsubset is used for model training e second subset isemployed for testing the model predictive capability when itpredicts corrosion status of novel image samples which hasnot been encountered in the training subset Moreover toreliably assess the model performance and to diminish therandomness caused by the data sampling process this re-search work has conducted a random subsampling of theoriginal dataset consisting of 20 runs In each run 30 of thedata is randomly extracted to constitute the testing subsetthe rest of the data is used for model training Accordinglythe overall model performance is reliably evaluated by av-eraging prediction results obtained from the repeated datasampling

(a)

(b)

Figure 3 e collected image samples (a) noncorrosion class and (b) corrosion class

Training dataset

The optimized SVM model

SVM model training

Corrosion detection result

The collected image samples

Texture descriptors

GLCM-basedfeatures

Statistical properties of color channels

GLRL-based features

DFP optimization

Testing dataset

Hyper parameter initialization

Start

Cost function evaluation

Stopping criterion verification

SVM model prediction

g = 1

g = g + 1

Figure 4 e proposed MO-SVM-PCD model used for pipe corrosion detection


In addition to the aforementioned PPV and NPV theclassification accuracy rate (CAR) recall and F1 score arealso used for expressing the modelrsquos predictive accuracyese indices are computed as follows [50]

classification accuracy rate CAR TP + TN

TP + TN + FP + FN

times 100

recall TP

TP + FN

F1 score 2TP

2TP + FP + FN

(15)

where TP TN FP and FN are the true positive truenegative false positive and false negative values

Demonstration of the feature extraction phase whichcomputes image sample texture is provided in Figure 5Herein for each image sample 78 features representing thestatistical measurements of image colors GLCM and GLRLare attained and used for data classification purpose Inaddition the evolutionary process of the DFPmetaheuristic-based SVM model optimization is illustrated in Figure 6which shows the best and the average cost function values ineach generatione optimal values of the penalty coefficientand the RBF kernel function parameter are found to be 430and 886 with the best cost function 108

e performance of the MO-SVM-PCD in the trainingand testing phases is reported in Table 2 As shown in thistable the proposed model has attained good predictiveaccuracy in both phases with CAR gt90 In detail theMO-SVM-PCD has achieved CAR 9117 PPV 091recall 092 NPV 092 and F1 score 091 in the testingphase ere is a focus on the MO-SVM-PCD performancein the testing phase because this reflects the generalizationcapability of the model

In addition corrosion detection based on the MO-SVM-PCD for a large-sized image samples can be achieved via ablockwise image separation process is image separationprocess is illustrated in Figure 7(a) In this figure each blockcorresponds to a sample having the size of 50times 50 eclassification result for the entire image is carried out bycombining the MO-SVM-PCD-based corrosion detectionfor each image block (see Figure 7(b)) e computationaltime required to classify one image block is about 4 secondstherefore the corrosion detection of the whole large-sizedimage (800times 600 pixels) requires about 768 seconds It isnoted that the detected positive class (corrosion class)samples are highlighted by red squares As can be seen fromthis figure the proposed approach can achieved relativelygood classification result Nevertheless several positivesamples located in the boundary of the corroded area havenot been identified correctly

Furthermore to better demonstrate the prediction ca-pability of the newly constructed MO-SVM-PCD employedfor detecting metal pipe corrosion its performance has beencompared to that of the least squares support vector machine

(LSSVM) [51] classification tree (CTree) [52] back-propagation artificial neural network (BPANN) [53] andconvolutional neural network (CNN) [54] e reason for

Image texture-basedfeatures

GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

20356 19016 19687 3270771 3430 947721X1 X2 X3 X76 X77 X78

(a)


GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

128762 101986 929316 116259 317479 59422X1 X2 X3 X76 X77 X78

(b)

Figure 5 Illustration of the feature extraction process (a) anoncorrosion case and (b) a corrosion case

0 20 40 60 80 100Iteration

108

109

11

111

112

113

114

115

Cost

func

tion

valu

e

Best found cost function valueAverage cost function value

Figure 6 e DFP optimization process

Table 2 Average prediction performance of the MO-SVM-PCD

Indices Training phase Testing phaseCAR () 98382 92808PPV 0982 0922Recall 0986 0936NPV 0986 0935F1 score 0984 0929


the selection of these benchmark models is that they havebeen confirmed to be capable methods for pattern classifi-cation by previous studies [5 40 55ndash57]

e LSSVM model is programmed in MATLAB by theauthors its tuning parameters including the regularizationcoefficient and kernel function parameter are also auto-matically identified by the DFP metaheuristic e CTree isdeveloped by the built-in functions provided in the MAT-LAB Statistics and Machine Learning Toolbox [46] eBPANNmodel is programmed inMATLAB environment bythe authors Via experiment the suitable parameter ofminimum leaf size of the employed CTree model has beenfound to be 2 Based on the suggestions of Heaton [58] andseveral trial-and-error runs the number of neuron in thehidden layer of the BPANN model is selected to be 2timesNI3 +NO 54 where NI 78 is the number of input featuresand NO 2 is the number of class labels Moreover thelearning rate and the number of training epochs of theneural network are set to be 01 and 1000 respectively

In addition the CNN model employed for corrosiondetection is constructed by the MATLAB image processingtoolbox [59] the stochastic gradient descent with mo-mentum (SGDM) and mini-batch mode are used in themodel training phase Via experimental runs the appro-priate configuration of the deep learning method is as fol-lows input image size is 50times 50 pixels e number ofconvolution layers is 4 e sizes of the filters are 20times 2016times16 8times 8 and 4times 4 in the 1st 2nd 3rd and 4th convo-lution layer respectively e number of filters in each layer

is 36 e batch size is 20 of the training data In additionthe CNN model has been trained in 1000 epochs In CNNthe feature extraction phase is automatically performed byconvolution layers therefore the CNN model does notrequires the feature computation done by the threeemployed image texture descriptors

e prediction results of all the models obtained fromthe repeated data sampling with 20 runs are summarized inTable 3 which reports the mean and the standard deviation(Std) of the model performance Observably the MO-SVM-PCD has attained the most desired predictive accuracy interms of CAR followed by BPANN LSSVM CNN andCTree e proposed pipe corrosion approach also achievesthe highest values of PPV recall NPV and F1 score ecomparison of model performance is graphically displayedin Figure 8

In addition the Wilcoxon signed-rank test [60] is uti-lized in this section to better confirm the statistical signif-icance of the differences in the model performancesis is anonparametric statistical test commonly employed formodel comparison [61] With the significance level of thetest 005 if the p value computed from the Wilcoxonsigned-rank test is lower than this significance level it is ableto reject the null hypothesis of insignificant difference inprediction outcomes of the two predictors Hence it isconfident to conclude that the predictive results of the twopipe corrosion detection models are statistically differentUsing the CAR values the outcome of the Wilcoxon signed-rank tests is reported in Table 4 is test points out that the

Original image

Blockwise separation

e MO-SVM-PCD corrosion

detection


(a)

Image with detected corrosion

(b)

Figure 7 e MO-SVM-PCD-based corrosion detection result (a) blockwise image separation process and (b) detection outcome with alarge-sized image sample


MO-SVM-PCD is statistically better than the LSSVMCTree BPANN and CNNwith p valueslt 005 Based on thisstatistical test it is able to state that the proposed method isthe most suited method for the task of interest

4 Conclusion

Corrosion is a commonly observed type of pipe defectsTimely detection of corrosion is very crucial to ensure theintegrity of the water supply system and avoid water

contamination In addition information regarding corrodedpipe sections obtained during periodic building surveys cansignificantly help to establish cost-effective maintenancestrategies for building owners is study puts forward anautomatic method based on image processing and machinelearning for pipe corrosion recognition Image processingtechniques have been employed to extract useful featuresfrom images of pipe surface to characterize the corrosionstatus In total 78 features are extracted using three texturedescriptors of the statistical properties of image colorGLCM and GLRL

e SVM machine learning method integrated with theDFP metaheuristic is utilized to construct a decisionboundary used for classifying pipe surface images into twocategories of noncorrosion and corrosion A dataset con-sisting of 2000 image samples has been used to train andvalidate the proposed hybrid model of the MO-SVM-PCDExperimental results supported by the Wilcoxon signed-rank test point out that the newly developed method issuperior to other benchmark approaches with an averageCAR 9281erefore the newly developed model can be

Table 3 Corrosion detection result of the machine learning models

Phase IndicesMO-SVM-PCD LSSVM CTree BPANN CNNMean Std Mean Std Mean Std Mean Std Mean Std

Train

CAR () 98382 0236 96432 0435 97018 0532 94593 1674 90890 1649PPV 0982 0004 0937 0008 0970 0006 0937 0020 0891 0024Recall 0986 0004 0996 0002 0971 0008 0956 0016 0933 0020NPV 0986 0004 0995 0002 0971 0007 0956 0016 0930 0019F1 0984 0002 0965 0004 0970 0005 0947 0016 0911 0016

Test


MO-SVM-PCD LSSVM CTree BPANN CNN080

085

090

095

Perfo

rman

ce in

dice

s

CAR

PPV

Recall

NPV

F1 score

Figure 8 Result comparison

Table 4 p values obtained from the Wilcoxon signed-rank test

Models MO-SVM-PCD LSSVM CTree BPANN CNN

MO-SVM-PCD mdash 000009 000009 000009 000009

LSSVM 000009 mdash 000427 001407 048933CTree 000009 000427 mdash 000022 001185BPANN 000009 001407 000022 mdash 001954CNN 000009 048933 001185 001954 mdash


a useful tool for building maintenance agents to quicklyevaluate the status of pipe systems Further extensions of thecurrent study may include the utilization of other advancedmachine learning for data classification employment ofother metaheuristic for model optimization employment ofhigher-order statistical features as input to machine learningbased classifiers enhancement of the detection accuracy forimage samples located in the boundary of the corroded areaimprovement of the computational efficiency of the currentmethod by employing advanced image segmentation tech-niques and collection of more image samples to enhance thegeneralization of the current MO-SVM-PCD model

Data Availability

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

Conflicts of Interest

e authors declare that there are no conflicts of interestregarding the publication of this paper

Supplementary Materials

e supplementary material of this study contains thedataset used to construct the machine learning-based modele first 78 columns of the dataset are the extracted texturefeatures e last column is the class label (0 for non-corrosion and 1 for corrosion) (Supplementary Materials)

References

[1] E Fleming Construction Technology An Illustrated In-troduction Blackwell Publishing Ltd Hoboken NJ USA2005

[2] F Bonnin-Pascual and A Ortiz ldquoCorrosion detection forautomated visual inspection developments in corrosionprotectionrdquo in Developments in Corrosion Protectionpp 620ndash632 IntechOpen London UK 2014

[3] V Bondada D K Pratihar and C S Kumar ldquoDetection andquantitative assessment of corrosion on pipelines throughimage analysisrdquo Procedia Computer Science vol 133pp 804ndash811 2018

[4] L Liu E Tan Y Zhen X J Yin and Z Q Cai ldquoAI-facilitatedcoating corrosion assessment system for productivity en-hancementrdquo in Proceedings of the 2018 13th IEEE Conferenceon Industrial Electronics and Applications (ICIEA) pp 606ndash610 Munich Germany May 2018

[5] D J Atha and M R Jahanshahi ldquoEvaluation of deep learningapproaches based on convolutional neural networks forcorrosion detectionrdquo Structural Health Monitoring vol 17no 5 pp 1110ndash1128 2018

[6] 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

[7] Y Jung H Oh and M M Jeong ldquoAn approach to automateddetection of structural failure using chronological imageanalysis in temporary structuresrdquo International Journal ofConstruction Management vol 19 no 2 pp 178ndash185 2019

[8] J-M Maatta J Vanne T Hamalainen and J NikkanenldquoGeneric software framework for a line-buffer-based imageprocessing pipelinerdquo IEEE Transactions on Consumer Elec-tronics vol 57 no 3 pp 1442ndash1449 2011

[9] D Itzhak I Dinstein and T Zilberberg ldquoPitting corrosionevaluation by computer image processingrdquo Corrosion Sciencevol 21 no 1 pp 17ndash22 1981

[10] K Y Choi and S S Kim ldquoMorphological analysis andclassification of types of surface corrosion damage by digitalimage processingrdquo Corrosion Science vol 47 no 1 pp 1ndash152005

[11] F N S Medeiros G L B Ramalho M P Bento andL C L Medeiros ldquoOn the evaluation of texture and colorfeatures for nondestructive corrosion detectionrdquo EURASIPJournal on Advances in Signal Processing article 817473 2010

[12] G Ji Y Zhu and Y Zhang ldquoe corroded defect ratingsystem of coating material based on computer visionrdquo inTransactions on Edutainment VIII Z Pan Ed pp 210ndash220Springer Berlin Germany 2012

[13] S A Idris and F A Jafar ldquoImage enhancement based onsoftware filter optimization for corrosion inspectionrdquo inProceedings of the 2014 5th International Conference on In-telligent Systems Modelling and Simulation vol 27ndash29pp 345ndash350 Langkawi Malaysia January 2014

[14] H Son N Hwang C Kim and C Kim ldquoRapid and auto-mated determination of rusted surface areas of a steel bridgefor robotic maintenance systemsrdquo Automation in Construc-tion vol 42 pp 13ndash24 2014

[15] K-W Liao and Y-T Lee ldquoDetection of rust defects on steelbridge coatings via digital image recognitionrdquo Automation inConstruction vol 71 pp 294ndash306 2016

[16] L Petricca T Moss G Figueroa and S Broen ldquoCorrosiondetection using AI a comparison of standard computervision techniques and deep learning modelrdquo in Proceedings ofthe Sixth International Conference on Computer ScienceEngineering and Information Technology vol 91ndash99 IEEEPiscataway NJ USA December 2016

[17] S Safari and M A Shoorehdeli ldquoDetection and isolation ofinterior defects based on image processing and neural net-worksrdquo HDPE Pipeline Case Study Journal of Pipeline SystemsEngineering and Practice vol 9 no 2 article 05018001 2018

[18] N Cheriet N E Bacha and A Skender ldquoKnowledge basesystem (KBS) applied on corrosion damage assessment onmetallic structure pipesrdquoHeliyon vol 4 no 10 article e008652018

[19] T Gibbons G Pierce K Worden and I Antoniadou ldquoAGaussian mixture model for automated corrosion detection inremanufacturingrdquo in Advances in Manufacturing TechnologyXXXII vol 8 pp 63ndash68 IOS Press Amsterdam Netherlands2018

[20] S K Ahuja and M K Shukla ldquoA survey of computer visionbased corrosion detection approachesrdquo in Proceedings of theInformation and Communication Technology for IntelligentSystems (ICTIS 2017) vol 2 pp 55ndash63 Springer InternationalPublishing Cham Switzerland August 2018

[21] V N Vapnik Statistical Learning Keory John Wiley amp SonsHoboken NJ USA 1998 ISBN-10 0471030031

[22] 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

[23] M Wang Y Wan Z Ye and X Lai ldquoRemote sensing imageclassification based on the optimal support vector machine


and modified binary coded ant colony optimization algo-rithmrdquo Information Sciences vol 402 pp 50ndash68 2017

[24] 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

[25] S eodoridis and K Koutroumbas Pattern RecognitionAcademic Press Cambridge MA USA 2009 ISBN 978-1-59749-272-0

[26] R M Haralick and L G Shapiro Computer and Robot VisionAddison-Wesley Longman Publishing Co Inc Boston MAUSA 1992 ISBN 0201569434

[27] M Sonka V Hlavac and R Boyle Image Processing Analysisand Machine Vision Cengage Learning Boston MA USA2013

[28] F Tomita and S Tsuji Computer Analysis of Visual TexturesSpringer Berlin Germany 1990

[29] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on Sys-tems Man and Cybernetics vol SMC-3 no 6 pp 610ndash6211973

[30] M M Galloway ldquoTexture analysis using gray level runlengthsrdquo Computer Graphics and Image Processing vol 4no 2 pp 172ndash179 1975

[31] B Abraham andM S Nair ldquoComputer-aided classification ofprostate cancer grade groups from MRI images using texturefeatures and stacked sparse autoencoderrdquo ComputerizedMedical Imaging and Graphics vol 69 pp 60ndash68 2018

[32] M R K Mookiah T Baum K Mei et al ldquoEffect of radiationdose reduction on texture measures of trabecular bone mi-crostructure an in vitro studyrdquo Journal of Bone and MineralMetabolism vol 36 no 3 pp 323ndash335 2018

[33] T Xiaoou ldquoTexture information in run-length matricesrdquoIEEE Transactions on Image Processing vol 7 no 11pp 1602ndash1609 1998

[34] A Chu C M Sehgal and J F Greenleaf ldquoUse of gray valuedistribution of run lengths for texture analysisrdquo PatternRecognition Letters vol 11 no 6 pp 415ndash419 1990

[35] B V Dasarathy and E B Holder ldquoImage characterizationsbased on joint gray levelmdashrun length distributionsrdquo PatternRecognition Letters vol 12 no 8 pp 497ndash502 1991

[36] L HamelKnowledge Discovery with Support Vector MachinesJohn Wiley amp Sons Hoboken NJ USA 2009 ISBN 978-0-470-37192-3

[37] M-Y Cheng and N-D Hoang ldquoTyphoon-induced slopecollapse assessment using a novel bee colony optimizedsupport vector classifierrdquo Natural Hazards vol 78 no 3pp 1961ndash1978 2015

[38] D Niu and S Dai ldquoA short-term load forecasting model witha modified particle swarm optimization algorithm and leastsquares support vector machine based on the denoisingmethod of empirical mode decomposition and grey relationalanalysisrdquo Energies vol 10 no 3 p 408 2017

[39] D Prayogo and Y T T Susanto ldquoOptimizing the predictionaccuracy of friction capacity of driven piles in cohesive soilusing a novel self-tuning least squares support vector ma-chinerdquo Advances in Civil Engineering vol 2018 Article ID6490169 9 pages 2018

[40] T Yi H Zheng Y Tian and J-p Liu ldquoIntelligent predictionof transmission line project cost based on least squaressupport vector machine optimized by particle swarm opti-mizationrdquo Mathematical Problems in Engineering vol 2018Article ID 5458696 11 pages 2018

[41] N-D Hoang D Tien Bui and K-W Liao ldquoGroutabilityestimation of grouting processes with cement grouts usingdifferential flower pollination optimized support vector ma-chinerdquo Applied Soft Computing vol 45 pp 173ndash186 2016

[42] R Storn and K Price ldquoDifferential evolutionmdasha simple andefficient heuristic for global optimization over continuousspacesrdquo Journal of Global Optimization vol 11 no 4pp 341ndash359 1997

[43] X-S Yang ldquoFlower pollination algorithm for global opti-mizationrdquo in Unconventional Computation and NaturalComputation pp 240ndash249 Springer Berlin Germany 2012

[44] K Price R M Storn and J A Lampinen DifferentialEvolutionmdashA Practical Approach to Global OptimizationSpringer-Verlag Berlin Germany 2005

[45] V-H Nhu N-D Hoang V-B Duong H-D Vu and D TienBui ldquoA hybrid computational intelligence approach forpredicting soil shear strength for urban housing constructiona case study at Vinhomes Imperia project Hai Phong City(Vietnam)rdquo Engineering with Computers 2019

[46] MathWork Statistics and Machine Learning Toolbox UserrsquosGuide MathWork Inc Natick MA USA 2017 httpswwwmathworkscomhelppdf_docstatsstatspdf

[47] Z Guo X Shao Y Xu H Miyazaki W Ohira andR Shibasaki ldquoIdentification of village building via Googleearth images and supervised machine learning methodsrdquoRemote Sensing vol 8 no 4 p 271 2016

[48] D G Altman and J M Bland ldquoStatistics notes diagnostictests 2 predictive valuesrdquo BMJ vol 309 no 6947 p 1021994

[49] T Kawasaki T Iwamoto M Matsumoto et al ldquoA method fordetecting damage of traffic marks by half celestial cameraattached to carsrdquo in Proceedings of the 12th EAI InternationalConference on Mobile and Ubiquitous Systems ComputingNetworking and Services Coimbra Portugal July2015

[50] N-D Hoang and Q-L Nguyen ldquoMetaheuristic optimizededge detection for recognition of concrete wall cracks acomparative study on the performances of roberts prewittcanny and sobel algorithmsrdquo Advances in Civil Engineeringvol 2018 Article ID 7163580 16 pages 2018

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

[52] L Breiman J H Friedman R A Olshen and C J StoneClassifcation and Regression Trees Wadsworth and BrooksMontery CA USA 1984 ISBN-13 978-0412048418

[53] D E Rumelhart G E Hinton and R J Williams ldquoLearningrepresentations by back-propagating errorsrdquoNature vol 323no 6088 pp 533ndash536 1986

[54] Y LeCun Y Bengio and G Hinton ldquoDeep learningrdquoNaturevol 521 no 7553 pp 436ndash444 2015

[55] K Khosravi B T Pham K Chapi et al ldquoA comparativeassessment of decision trees algorithms for flash flood sus-ceptibility modeling at Haraz watershed Northern IranrdquoScience of the Total Environment vol 627 pp 744ndash755 2018

[56] P-T Ngo N-D Hoang B Pradhan et al ldquoA novel hybridswarm optimized multilayer neural network for spatial pre-diction of flash floods in tropical areas using sentinel-1 SARimagery and geospatial datardquo Sensors vol 18 no 11 p 37042018

[57] Z Tong J Gao A Sha L Hu and S Li ldquoConvolutional neuralnetwork for asphalt pavement surface texture analysisrdquoComputer-Aided Civil and Infrastructure Engineering vol 33no 12 pp 1056ndash1072 2018


[58] J Heaton Introduction to Neural Networks for C HeatonResearch Inc St Louis MO USA 2008

[59] MathWorks Image Processing Toolbox Userrsquos Guide MathWorkInc Natick MA USA 2016 httpswwwmathworkscomhelppdf_docimagesimages_tbpdf

[60] S Sidney Non-Parametric Statistics for the Behavioral Sci-ences McGraw-Hill New York NY USA 1988 ISBN0070573573

[61] 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


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

technicians erefore there is a practical need to come upwith a more productive and accurate method of pipe con-dition survey

Although there is a wide range of existing pipe in-spection approaches (such as magnetic flux leakage ul-trasonic testing and external corrosion direct assessment)all of these methods have limitations including highequipment cost restricted range of inspection and in-capability of detecting small pitting regions [3] Consid-ering the large amount of pipe systems needed to besurveyed and the limited access to sophisticated equipmentin developing countries there is an urgent need for aproductive and low-cost solution for periodic surveys ofpipe system condition Recently digital image processinghas gained a great attention within the field of structuralheath monitoring [6 7]

Particularly image processing techniques can be effec-tively employed to investigate the outer surface for detectingdefects on pipes or other metal structures including cor-rosion and cracks [8] Itzhak et al [9] relied on statisticmeasurement of image pixels to quantify pitting corrosionChoi and Kim [10] identified corrosion based on themorphology of the corroded surface features of image colortexture and shape are employed for corrosion recognitionA model for classifying corroded and noncorroded surfacesusing texture descriptors obtained from gray level co-oc-currence matrix and image color has been proposed inMedeiros et al [11]

A method based on watershed segmentation has beenemployed in [12] for rating of corrosion defects the per-centage area of corroded region was used for determiningthe grade of defects Idris and Jafar [13] used image filter-based image enhancement and neural network for corrosioninspection Son et al [14] proposed a model based on de-cision tree algorithm for identifying rusted surface area of

steel bridge A model based on image color analysis andK-means clustering for bridge rust identification has beenconstructed and verified by Liao and Lee [15]

Petricca et al [16] compared standard computer visiontechniques and deep neural network for rust and nonrustdetection Deep neural networks have also been employedfor corrosion detection by Liu et al [4] and Atha andJahanshahi [5] Safari and Shoorehdeli [17] applied artificialneural network Gabor filter and entropy filter for pipedefect detection Cheriet et al [18] incorporated expertknowledge and field data to construct a knowledge-basedsystem for assessing corrosive damage on metallic pipeconduits Gibbons et al [19] relied on a Gaussian mixturemodel for probabilistic classification of corroded andnoncorroded areas Bondada et al [3] detected and quan-titatively assessed corrosion damages on pipelines bycomputing the mean of saturation value of image pixels byimage analysis the corroded areas on pipelines can besegmented

From the above literature it can be seen that imageprocessing and machine learning have been a feasible al-ternative for replacing the tedious process of manual surveyBased on a recent review work of Ahuja and Shukla [20]there is an increasing trend of applying computer visiontechniques for corrosion detection Moreover due to theimportance of the research theme exploring other imageprocessing and machine learning methods used for pipecorrosion detection can be highly meaningful in both aca-demic and practical aspects

As reported in the literature although image textureanalysis has been applied few previous studies haveemployed a combination of image texture descriptors forpipe corrosion recognition Hence this study is an attemptto fill this gap in the literature by proposing a method usedfor analyzing texture of water pipe surface that integrates

(a) (b)

(c) (d)

Figure 1 Corroded areas on pipe surface







Pc(I) NIc

W times H (1)



μc 1113944NLminus1

i0Iic times Pc(I)

σc

1113944

NLminus1

i0Iic minus μc1113872 1113873

2times Pc(I)

11139741113972

δc 1113936


3times Pc(I)

σ3c

ηc 1113936


4times Pc(I)

σ4c




(2)



PNδ (i j)

Pδ(i j)

SP (3)




AM 1113944

Ng

i11113944

Ng

j1P

Nδ (i j)

2

CO 1113944

Ngminus1

k0k2

1113944

Ng

i1|iminusj|k

1113944

Ng

j1P

Nδ (i j)

CR 1113936

Ngi11113936



σXσY

ET minus1113944

Ng

i11113944

Ng

j1P

Nδ (i j)log P

Nδ (i j)1113872 1113873

(4)









Tw + c

12

1113944

N

k1e2k


(5)




K xk xl( 1113857 φ xk( 1113857Tφ xl( 1113857 (6)



2

2σ2⎛⎝ ⎞⎠ (7)



y xl( 1113857 sign 1113944

SV

k1αkykK xk xl( 1113857 + b⎛⎝ ⎞⎠ (8)








Xtriali X

gDFPi + L middot X







XmutatedigDFP




XcrossedjigDFP+1



⎧⎨

⎩

(11)





Kernel mappingΦ(x)

Φ(xu)

Φ(xv)


Text

ure d

escr

ipto

r xj

Φ(xl)





Mi11113936

Nj1p(i j)j2


Nj1p(i j) times j2


Nj1p(i j)2)


Mi1p(i j)2)



Mi1p(i j)i2


Mi1p(i j) times i2
















XZN Xo minusmX

sX

(12)




fCF 1113936


K (13)



PPV TP

TP + FP

NPV TN

TN + FN

(14)








(a)

(b)


Training dataset


SVM model training



Texture descriptors

GLCM-basedfeatures


GLRL-based features

DFP optimization

Testing dataset


Start




g = 1

g = g + 1





TP + TN + FP + FN

times 100

recall TP

TP + FN

F1 score 2TP

2TP + FP + FN

(15)








GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

20356 19016 19687 3270771 3430 947721X1 X2 X3 X76 X77 X78

(a)


GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

128762 101986 929316 116259 317479 59422X1 X2 X3 X76 X77 X78

(b)


0 20 40 60 80 100Iteration

108

109

11

111

112

113

114

115

Cost

func

tion

valu

e












Original image



detection


(a)


(b)




4 Conclusion






Train


Test



085

090

095

Perfo

rman

ce in

dice

s

CAR

PPV

Recall

NPV

F1 score




MO-SVM-PCD mdash 000009 000009 000009 000009




Data Availability






References







































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in









Pc(I) NIc

W times H (1)



μc 1113944NLminus1

i0Iic times Pc(I)

σc

1113944

NLminus1

i0Iic minus μc1113872 1113873

2times Pc(I)

11139741113972

δc 1113936


3times Pc(I)

σ3c

ηc 1113936


4times Pc(I)

σ4c




(2)



PNδ (i j)

Pδ(i j)

SP (3)




AM 1113944

Ng

i11113944

Ng

j1P

Nδ (i j)

2

CO 1113944

Ngminus1

k0k2

1113944

Ng

i1|iminusj|k

1113944

Ng

j1P

Nδ (i j)

CR 1113936

Ngi11113936



σXσY

ET minus1113944

Ng

i11113944

Ng

j1P

Nδ (i j)log P

Nδ (i j)1113872 1113873

(4)









Tw + c

12

1113944

N

k1e2k


(5)




K xk xl( 1113857 φ xk( 1113857Tφ xl( 1113857 (6)



2

2σ2⎛⎝ ⎞⎠ (7)



y xl( 1113857 sign 1113944

SV

k1αkykK xk xl( 1113857 + b⎛⎝ ⎞⎠ (8)








Xtriali X

gDFPi + L middot X







XmutatedigDFP




XcrossedjigDFP+1



⎧⎨

⎩

(11)





Kernel mappingΦ(x)

Φ(xu)

Φ(xv)


Text

ure d

escr

ipto

r xj

Φ(xl)





Mi11113936

Nj1p(i j)j2


Nj1p(i j) times j2


Nj1p(i j)2)


Mi1p(i j)2)



Mi1p(i j)i2


Mi1p(i j) times i2
















XZN Xo minusmX

sX

(12)




fCF 1113936


K (13)



PPV TP

TP + FP

NPV TN

TN + FN

(14)








(a)

(b)


Training dataset


SVM model training



Texture descriptors

GLCM-basedfeatures


GLRL-based features

DFP optimization

Testing dataset


Start




g = 1

g = g + 1





TP + TN + FP + FN

times 100

recall TP

TP + FN

F1 score 2TP

2TP + FP + FN

(15)








GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

20356 19016 19687 3270771 3430 947721X1 X2 X3 X76 X77 X78

(a)


GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

128762 101986 929316 116259 317479 59422X1 X2 X3 X76 X77 X78

(b)


0 20 40 60 80 100Iteration

108

109

11

111

112

113

114

115

Cost

func

tion

valu

e












Original image



detection


(a)


(b)




4 Conclusion






Train


Test



085

090

095

Perfo

rman

ce in

dice

s

CAR

PPV

Recall

NPV

F1 score




MO-SVM-PCD mdash 000009 000009 000009 000009




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


(5)




K xk xl( 1113857 φ xk( 1113857Tφ xl( 1113857 (6)



2

2σ2⎛⎝ ⎞⎠ (7)



y xl( 1113857 sign 1113944

SV

k1αkykK xk xl( 1113857 + b⎛⎝ ⎞⎠ (8)








Xtriali X

gDFPi + L middot X







XmutatedigDFP




XcrossedjigDFP+1



⎧⎨

⎩

(11)





Kernel mappingΦ(x)

Φ(xu)

Φ(xv)


Text

ure d

escr

ipto

r xj

Φ(xl)





Mi11113936

Nj1p(i j)j2


Nj1p(i j) times j2


Nj1p(i j)2)


Mi1p(i j)2)



Mi1p(i j)i2


Mi1p(i j) times i2
















XZN Xo minusmX

sX

(12)




fCF 1113936


K (13)



PPV TP

TP + FP

NPV TN

TN + FN

(14)








(a)

(b)


Training dataset


SVM model training



Texture descriptors

GLCM-basedfeatures


GLRL-based features

DFP optimization

Testing dataset


Start




g = 1

g = g + 1





TP + TN + FP + FN

times 100

recall TP

TP + FN

F1 score 2TP

2TP + FP + FN

(15)








GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

20356 19016 19687 3270771 3430 947721X1 X2 X3 X76 X77 X78

(a)


GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

128762 101986 929316 116259 317479 59422X1 X2 X3 X76 X77 X78

(b)


0 20 40 60 80 100Iteration

108

109

11

111

112

113

114

115

Cost

func

tion

valu

e












Original image



detection


(a)


(b)




4 Conclusion






Train


Test



085

090

095

Perfo

rman

ce in

dice

s

CAR

PPV

Recall

NPV

F1 score




MO-SVM-PCD mdash 000009 000009 000009 000009




Data Availability






References







































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in






Xtriali X

gDFPi + L middot X







XmutatedigDFP




XcrossedjigDFP+1



⎧⎨

⎩

(11)





Kernel mappingΦ(x)

Φ(xu)

Φ(xv)


Text

ure d

escr

ipto

r xj

Φ(xl)





Mi11113936

Nj1p(i j)j2


Nj1p(i j) times j2


Nj1p(i j)2)


Mi1p(i j)2)



Mi1p(i j)i2


Mi1p(i j) times i2
















XZN Xo minusmX

sX

(12)




fCF 1113936


K (13)



PPV TP

TP + FP

NPV TN

TN + FN

(14)








(a)

(b)


Training dataset


SVM model training



Texture descriptors

GLCM-basedfeatures


GLRL-based features

DFP optimization

Testing dataset


Start




g = 1

g = g + 1





TP + TN + FP + FN

times 100

recall TP

TP + FN

F1 score 2TP

2TP + FP + FN

(15)








GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

20356 19016 19687 3270771 3430 947721X1 X2 X3 X76 X77 X78

(a)


GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

128762 101986 929316 116259 317479 59422X1 X2 X3 X76 X77 X78

(b)


0 20 40 60 80 100Iteration

108

109

11

111

112

113

114

115

Cost

func

tion

valu

e












Original image



detection


(a)


(b)




4 Conclusion






Train


Test



085

090

095

Perfo

rman

ce in

dice

s

CAR

PPV

Recall

NPV

F1 score




MO-SVM-PCD mdash 000009 000009 000009 000009




Data Availability






References







































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in









XZN Xo minusmX

sX

(12)




fCF 1113936


K (13)



PPV TP

TP + FP

NPV TN

TN + FN

(14)








(a)

(b)


Training dataset


SVM model training



Texture descriptors

GLCM-basedfeatures


GLRL-based features

DFP optimization

Testing dataset


Start




g = 1

g = g + 1





TP + TN + FP + FN

times 100

recall TP

TP + FN

F1 score 2TP

2TP + FP + FN

(15)








GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

20356 19016 19687 3270771 3430 947721X1 X2 X3 X76 X77 X78

(a)


GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

128762 101986 929316 116259 317479 59422X1 X2 X3 X76 X77 X78

(b)


0 20 40 60 80 100Iteration

108

109

11

111

112

113

114

115

Cost

func

tion

valu

e












Original image



detection


(a)


(b)




4 Conclusion






Train


Test



085

090

095

Perfo

rman

ce in

dice

s

CAR

PPV

Recall

NPV

F1 score




MO-SVM-PCD mdash 000009 000009 000009 000009




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)


Training dataset


SVM model training



Texture descriptors

GLCM-basedfeatures


GLRL-based features

DFP optimization

Testing dataset


Start




g = 1

g = g + 1





TP + TN + FP + FN

times 100

recall TP

TP + FN

F1 score 2TP

2TP + FP + FN

(15)








GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

20356 19016 19687 3270771 3430 947721X1 X2 X3 X76 X77 X78

(a)


GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

128762 101986 929316 116259 317479 59422X1 X2 X3 X76 X77 X78

(b)


0 20 40 60 80 100Iteration

108

109

11

111

112

113

114

115

Cost

func

tion

valu

e












Original image



detection


(a)


(b)




4 Conclusion






Train


Test



085

090

095

Perfo

rman

ce in

dice

s

CAR

PPV

Recall

NPV

F1 score




MO-SVM-PCD mdash 000009 000009 000009 000009




Data Availability






References







































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in






TP + TN + FP + FN

times 100

recall TP

TP + FN

F1 score 2TP

2TP + FP + FN

(15)








GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

20356 19016 19687 3270771 3430 947721X1 X2 X3 X76 X77 X78

(a)


GLCM-basedfeatures


GLRL-basedfeatures

Texture descriptors

128762 101986 929316 116259 317479 59422X1 X2 X3 X76 X77 X78

(b)


0 20 40 60 80 100Iteration

108

109

11

111

112

113

114

115

Cost

func

tion

valu

e












Original image



detection


(a)


(b)




4 Conclusion






Train


Test



085

090

095

Perfo

rman

ce in

dice

s

CAR

PPV

Recall

NPV

F1 score




MO-SVM-PCD mdash 000009 000009 000009 000009




Data Availability






References







































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in










Original image



detection


(a)


(b)




4 Conclusion






Train


Test



085

090

095

Perfo

rman

ce in

dice

s

CAR

PPV

Recall

NPV

F1 score




MO-SVM-PCD mdash 000009 000009 000009 000009




Data Availability






References







































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in





4 Conclusion






Train


Test



085

090

095

Perfo

rman

ce in

dice

s

CAR

PPV

Recall

NPV

F1 score




MO-SVM-PCD mdash 000009 000009 000009 000009




Data Availability






References







































































Advances in

FuzzySystems


Volume 2018













Journal of

Journal of



Hindawi


Advances in

Multimedia


Biomedical Imaging





RoboticsJournal of









Volume 2018



Advances in





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




ImageProcessing-BasedDetectionofPipeCorrosionUsing...

Documents

Transcript of ImageProcessing-BasedDetectionofPipeCorrosionUsing...