analysis of edge-detection techniques for crack...

Analysis of Edge-Detection Techniques for CrackIdentification in Bridges

Ikhlas Abdel-Qader, P.E.1; Osama Abudayyeh, P.E., M.ASCE2; and Michael E. Kelly3

struc-bridgend cans paper

bridgetion

Abstract: Bridge monitoring and maintenance is an expensive yet essential task in maintaining a safe national transportation infrature. Traditional monitoring methods use visual inspection of bridges on a regular basis and often require inspectors to travel to theof concern and determine the deterioration level of the bridge. Automation of this process may result in great monetary savings alead to more frequent inspection cycles. One aspect of this automation is the detection of cracks and deterioration of a bridge. Thiprovides a comparison of the effectiveness of four crack-detection techniques: fast Haar transform~FHT!, fast Fourier transform, Sobel,and Canny. These imaging edge-detection algorithms were implemented in MatLab and simulated using a sample of 50 concreteimages~25 with cracks and 25 without!. The results show that the FHT was significantly more reliable than the other three edge-detectechniques in identifying cracks.

DOI: 10.1061/~ASCE!0887-3801~2003!17:4~255!

CE Database subject headings: Cracking; Bridge maintenance; Monitoring.

ettfima

n

inr

h

c

h

s

in-

-d

onntes

tff-n

-

-

-s-at

iot8

nil:

a0motoiba

.

Introduction

The transportation infrastructure is key to economic developmin the United States. Transportation-based industries accounabout 20% of the U.S. gross national product and approxima800,000 jobs~Brecher 1995!. Therefore, providing a high level oserviceability through periodic inspection and maintenance isportant in keeping the transportation system operationalavoiding major replacement efforts. However, for a numberyears many state departments of transportation~DOTs! have spentmost of their planning and budgeting efforts and allocated monon new construction, while maintenance and rehabilitation geally were managed with less formal methods~Hass et al. 1994;PCA 1995!. In several cases, crises were the driving force behinitiating maintenance and/or rehabilitation actions, particulawhen funds were limited. Such an approach, however, waslonger sufficient or appropriate, particularly since most of ttransportation infrastructure had reached its design service~PCA 1995!.

As a result of the shortcomings in dealing with maintenanand rehabilitation needs, pavement and bridge managementtems evolved to help plan maintenance and rehabilitation of paments and bridges and to avoid the crisis reaction approac

1Assistant Professor and Director of the Center for InformatTechnology and Imaging Analysis, Dept. of Electrical and CompuEngineering, Western Michigan Univ., Kalamazoo, MI 4900E-mail: [email protected]

2Associate Professor, Dept. of Civil and Construction EngineeriWestern Michigan Univ., Kalamazoo, MI 49008. [email protected]

3Undergraduate REU Research Assistant, Dept. of ElectricalComputer Engineering, Western Michigan Univ., Kalamazoo, MI 490

Note. Discussion open until March 1, 2004. Separate discussionsbe submitted for individual papers. To extend the closing date bymonth, a written request must be filed with the ASCE Managing EdiThe manuscript for this paper was submitted for review and posspublication on October 1, 2002; approved on January 2, 2003. This pis part of theJournal of Computing in Civil Engineering, Vol. 17, No. 4,October 1, 2003. ©ASCE, ISSN 0887-3801/2003/4-255–263/$18.00

JOURNAL OF

ntfor

ely

-ndof

ieser-

dlynoelife

esys-ve-

to

maintaining the transportation infrastructure~Gole 1985; Hasset al. 1994!. However, the costs associated with developing andmaintaining the transportation infrastructure in the United Statecontinue to increase, and funding continues to shrink~AASHTO1988!.

The Federal Highway Administration~FHWA! estimates thatan annual $50 billion would be necessary to maintain the roadstheir present condition, and $215 billion would be needed to rehabilitate all deficient roads and bridges~Brecher 1995; Robertsand Shepard 2000!. Furthermore, a recent FHWA study indicatedthat one-third of existing bridges in the United States are structurally deficient, 10% of pavements require immediate repair, an60% of pavements need rehabilitation~Brecher 1995; Roberts andShepard 2000!. Therefore, a good transportation infrastructuremanagement system is key to the success of the transportatisystem. The inventory of bridges in the national transportatioinfrastructure is of particular importance due to their high cosand direct impact on public safety; thus bridge maintenance issuare the focus of the project described in this paper.

Bridge Inspection

The total cost of maintenance, rehabilitation, and replacemen~MR&R! is related to the average age of the entire network obridges. Many factors tend to accelerate the deterioration obridges; for instance, truck weights and traffic counts have increased dramatically, causing premature physical weathering oall networks, while available funds for MR&R have not been ableto keep up with the declining situation. A major component ofbridge maintenance and rehabilitation functions is their data contents~Saito et al. 1990; AASHTO 1993; Hass et al. 1994!. Dataon the condition of bridges are periodically collected and analyzed to determine the optimum allocation of funds among newconstruction, maintenance, and rehabilitation programs~PCA1995!. Since more time and funds are now being spent on maintenance and rehabilitation of the existing system, there is a presing need to develop effective bridge management systems th

ner.

g,

nd8.ustner.le

per

COMPUTING IN CIVIL ENGINEERING © ASCE / OCTOBER 2003 / 255

fan

ntprtiofw

and

pe

anluacate.mecor t

sizsntioe

d ttegfre

t

rl

hoawalpoes

er

toodnsoinaa

ntatnem

suem

-

dlu-

ge

-e

n

i-essetst

-e-

ita-sm

arehet-ed

,x,ise-

on

lt-

-

s

gate

enhance the method of collecting, organizing, and using dataproper planning and appropriate performance of the maintenand rehabilitation processes.

Bridge monitoring and inspection are expensive yet essetasks in maintaining a safe infrastructure. Traditionally, themary method used to monitor bridges has been visual inspecDuring a typical bridge inspection, the various components ostructure are examined at close range by trained inspectorsevaluate the condition of the components and give them a ring. This ranking is a subjective evaluation of the current contion based on a set of guidelines and on the inspector’s exence.

For many situations, this type of evaluation is appropriateeffective. However, due to the subjective nature of this evation, rankings of the conditions of similar bridge componentsvary widely from inspector to inspector and from state to sta

Moreover, inspections are not necessarily always perforon each bridge at the appropriate time. Several factors cantribute to the selection of inspection procedures and may altetiming of inspections~Silano 1993; AASHTO 2000!. These fac-tors include condition of structure as related to design, age,and complexity of a bridge; traffic density and consequencetraffic disruption; availability of personnel and equipmeweather pattern; environmental conditions; geographic locatand methods and records of construction procedures. Any onthese factors can affect the deterioration rate of the bridge anneed for MR&R. Therefore an agency should develop a strafor systematic inspection and documentation that includesquency of inspection, the nature of observations, and equipmfor measurements. The goals of an inspection program musclude• Assurance of safety and serviceability;• Identifying actual and potential sources of trouble at ea

stages;• Systematically recording state of structure; and• Studying effect of changes in load.

In the past two decades the need for more advanced metof bridge inspection to ensure safety and early detection of flhas been recognized~Rens and Greimann 1997; Abudayyeh et2000!. Many evaluation methods are designed to operate uexisting bridges without damaging their usability. Some of thmethods, known as nondestructive evaluation~NDE! methods,may be very broad and versatile and can be used in a numbapplications, while others are very specialized.

More recently, bridge inspection has included remote moniing devices that record stress and vibration in real time to maccurately determine the wear on a particular bridge and aiearly detection of problems. Remote monitoring used in conjution with visual inspection has the potential to reduce the costbridge inspection and maintenance. A fully automated remmonitoring system could reduce the number of trips a bridgespector must make over the lifetime of a bridge, should increthe frequency of inspection cycles, and would serve as an ewarning system when a bridge begins to deteriorate significaOne component of a remote monitoring system is an automcrack-detection technique. This paper compares the effectiveof several traditional and newly developed imaging algoriththat can be used in crack detection in concrete bridges.

Imaging Algorithms for Bridge Crack Detection

Accurate and timely information on the status of a pavementface is crucial to the effective management of pavement syst

256 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING © ASCE / OC

orce

iali-n.ahok-i-ri-

d-

n

dn-

he

e,of;n;ofhey

e-nt

in-

y

dss

.n

e

of

r-reinc-ofte-

serlyly.edss

s

r-s.

The importance of timely data acquisition and analysis of pavement surfaces is widely acknowledged~El-Korchi 1990; Koutso-poulos and Downey 1993; Lee 1993!. To achieve timeliness andaccuracy in bridge inspection, automation of data acquisition ananalysis is necessary. A number of automated systems for evaating images of pavement surfaces have been developed~Lee1990; Longenecker 1990; Klassen and Swindall 1993!. This paperfocuses on automated crack-detection techniques for the briddeck inspection process~these techniques are based on imagingedge-detection algorithms!. In particular, this paper proposes anddemonstrates a framework for evaluating the effectiveness of imaging crack-detection algorithms in inspecting concrete bridgdeck surfaces~concrete pavements!, resulting in the selection ofthe best algorithm. This framework should be used before aautomated pavement evaluation system is developed.

Edges in digital images are defined as sharp intensity transtions. Edge-detection algorithms seek to detect and localize edgwithout any input or interference from humans. These algorithmare typically application based and may not produce the samresults for a given image. Therefore, investigating the differenalgorithms as proposed in this paper can help in choosing the beone for bridge images. Additionally, once an algorithm is identified as the most suitable for the problem at hand, further enhancments can be researched and developed to overcome any limtions that may exist. Four imaging edge-detection algorithmwere chosen for comparison in this paper: fast Haar transfor~FHT!, fast Fourier transform~FFT!, Sobel, and Canny. Thesealgorithms are used in numerous edge-detection problems andconsidered here as possible crack-detection techniques for tbridge inspection problem. This section provides a brief theoreical background for each of these techniques, which are evaluatusing the proposed framework in the next section.

Sobel Edge Detector

The Sobel edge detector is known for its simplicity and speedcompared to other algorithms that are computationally compleand is based on the spatial gradient algorithm. However, Sobelvery susceptible to image noise, which may result in false positivdetections when attempting to identify edges in real-world situations. This is inherent in all gradient-based algorithms. Filteringto smooth the noise out can reduce the false edge identificatibut does not eliminate the problem.

The Sobel algorithm detects gradient~changes! in image in-tensity. The image data are convolved with a Sobel mask, resuing in first-order partial derivatives for the pixel in the middle ofthe mask~Parker 1997!. The Sobel edge detector uses two convolution masks (Cx and Cy) of size 333 each to compute thederivatives in both thex and y directions. The values of thesemasks are

F 1 2 1

0 0 0

21 22 21G5Cx; F 21 0 1

22 0 2

21 0 1G5Cy

Canny Edge Detector

The Canny edge detector is also a convolution filter, but islightly more complicated and powerful than the Sobel~Canny1986!. Canny first smoothes an image by convolving with aGaussian mask to eliminate most of the noise, thus overcominSobel’s limitation. Next, the algorithm attempts to detect edgesthe maxima of the gradient modulus taken in the direction of th

TOBER 2003

Fig. 1. Example Haar image:~a! four quadrants;~b! combination of quadrants 2, 3, and 4

ios

iors

-

g

t

nss

-

gh-ings of

wsis

then

he

gradient. This algorithm produces an edge strength and directat each pixel in the smoothed image. The computations are baon determining the gradient magnitude as the root of the sumthe squares of the derivatives in both directionsx,y. The directionof the edge is computed using the arctan of the derivative rat~Canny 1986!. Thus the algorithm is based on three parametes, which is the standard deviation of the Gaussian mask, andt low

and thigh, which are used for thresholding to determine if a pixebelongs to an edge or not.

Fourier Transform

The Fourier transform~FT! is a frequency-based, discrete transform that is optimized for machine calculation~Cooley and Tukey1965!. It is an extremely popular transform in the engineerinworld, with a wide range of applications. The FT is used to determine the frequency content of a signal in digital signal procesing applications. In imaging applications, the FT allows interconversion between the spatial domain representation andfrequency domain representation of images, causing the FTbecome one of the most powerful tools in image processing afiltering. To improve the computational speed, fast Fourier tranforms ~FFTs! emerged. The FFT mathematical formulation ishown below:

F~u,v !51

MN (x50

M21

(y50

N21

f ~x,y!e22p j ~xu/M1yv/N! (1)

JOURNAL OF

nedof

s:

l

-s--hetod-

f ~x,y!51

MN (x50

M21

(y50

N21

F~u,v !e22p j ~xu/M1yv/N! (2)

where f (x,y)5image of sizeM3N that is transformed into thecoefficient spectrum imageF(u,v) by means of the FFT;x andy5spatial samples~pixels!; and u and v5frequency samples.After performing the transformation, and based on a threshold~tobe discussed in the experimental results section!, a decision ofcrack or no crack can be made.

Fast Haar Transform

The fast Haar transform~FHT! is relatively new and shows promise in its ability to detect edges~Bachman and Beckenstein 2000!.Haar decomposes the image into low-frequency and hifrequency components. This process is followed by isolatthose high-frequency coefficients from which the edge featurean image are identified~Alageel and Abdel-Qader 2002!. To per-form operations on images, the 2D FHT is used, which follothe same principles as the 1D FHT. Each row of the imagepassed through a 1D FHT, the resulting image is saved, andeach column of the new image is passed through a 1D FHT.

The Haar is a simple form of wavelets introduced in 1910. Tmother wavelet of Haarc(t) and the scaling functionf(t) aredefined as

Fig. 2. Crack image

C

Fig. 3. No-crack image

OMPUTING IN CIVIL ENGINEERING © ASCE / OCTOBER 2003 / 257

aaringlyks,rstise

a

sisetee-

h a

a-to
Fig. 4. Crack-detection imaging algorithm evaluation process
ion

sso-

50

c~ t !551 0<t,

1

2

211

2<t,1

0 elsewhere

(3)

f~ t !5H 1 0<t,1

0 elsewhere(4)

The general forms for the bases off(t) andc(t) are given in theequations below:

f j ,k~ t !52 j /2f~2 j t2k! (5)

c j ,k~ t !52 j /2c~2 j t2k! (6)

where 2j /25normalization factor;j 50,1,...; andk50,1,...,2j21.The Haar wavelet relates to the scaling function by the relat

c~ t !5f~2t !2f~2t21! (7)

and the scaling functionsf relate to each other by the relation

258 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING © ASCE / OCT

f~ t !5f~2t !1f~2t21! (8)

Using the basis ensemble for the Haar wavelets, the fast Htransform is created and used to implement the transform, leadto a simple 2D implementation of the Haar, which is essentialaveraging, differencing, and scaling processes. To detect cracthe magnitude of quadrants 2, 3, and 4 produced by Haar is fitaken and passed through a high-pass filter to eliminate no@Fig. 1~a!#. Then the three quadrants are combined to producemagnitude image@Fig. 1~b!#, the intensity of which is then com-pared to a threshold level to determine if a crack exists.

Discussion of Experimental Results

A sample of 50 concrete bridge images was used in the analyand comparison study. Half the images were of healthy concrbridge components, and the other half were of cracking and dteriorating concrete. The images were all gray-scale images witresolution of 6403480 pixels. All images were of bridge decksurfaces~concrete pavements! and do not include backgroundsuch as grass or motor vehicle traffic~Figs. 2 and 3 show twosample images!. Ideally, a stationary, bridge-mounted camerwould focus directly on the concrete, but if background is included in the image, it may be necessary to include a prefiltereliminate such anomalies.

A MatLab code was developed for each of the four algorithmto read the images, perform the transforms, and output the ilated cracks~usually referred to as the edge image!. The edgeimage was recorded for each technique and for each of the

Fig. 5. Haar for crack image of Fig. 2

Fig. 6. Haar for no-crack image of Fig. 3

OBER 2003

Fig. 7. Intensity data for fast Haar transform

thti

helu

. I

g

e

images. Based on the threshold value for the algorithm andintensity in the edge image, an output image was determinedhave a crack or no crack; this process is schematically shownFig. 4. A threshold value is a parameter that is crucial to tperformance of any edge-detection algorithm. Based on the vaof this parameter, it is decided that a crack exists or does notthis project, the threshold is determined as the average value

JOURNAL OF C

eon

enof

the intensity of all pixels in the crack images. The following is alist of the experimental results obtained from the four imagintechniques.

Fast Haar TransformThe fast Haar transform performed significantly better than thtraditional gradient-based algorithms~Sobel and Canny!. The

Fig. 8. Fourier for crack image of Fig. 2
Fig. 9. Fourier for no-crack image of Fig. 3

Fig. 10. Intensity data for fast Fourier transform

sftiofw

hic

ndres-

s

FHT eliminated much of the noise caused by normal patternconcrete~highly textures images!, resulting in a high level oaccuracy determined by the number of correct crack detecproduced by the technique. The overall accuracy was 43 out o~86%! correct detections. The mean value of the edge images2.4 times greater than that of images without a crack, w


in

ns50ash

shows a large distinction between crack and no-crack images awas much better than other algorithms tested. Figs. 5 and 6 athe isolated cracks for the two bridge images of Figs. 2 and 3 adetermined by the FHT technique, and Fig. 7 shows the imageintensity data calculated by the FHT algorithm for the 50 imageused in the analysis.

Fig. 11. Sobel for crack image of Fig. 2

T

Fig. 12. Sobel for no-crack image of Fig. 3

OBER 2003

Fig. 13. Intensity data for Sobel

histo

larg50n-ck

ogesageible

om-ta-

lyter

Fast Fourier TransformThe fast Fourier transform performed poorly in these tests. Twas slightly surprising because of the FFT’s excellent abilitydetect edges. However, the texture of the images caused aamount of misclassifications. The overall accuracy of 32 out of~64%! was the lowest of the four algorithms. The average intesity of crack images was only 6% higher than that of no-craimages, making this transform nearly useless in the detectioncracks. Figs. 8 and 9 are the isolated cracks for two bridge imaas determined by the FFT technique, and Fig. 10 is the imintensity data for the 50 images used in the FFT analysis. Poss

JOURNAL OF C

e

f

improvement of this filter could be achieved by using a prefilter treduce some of the noise and texture in the images. A large iprovement was not expected, however, and so this implemention was not tested.

SobelThe Sobel edge-detection technique also performed relativepoorly. The average intensity of crack images was 25% grea

Table 1. Summary of Analysis Resultsa

AnalysisFast Haartransform

Fast Fouriertransform Sobel Canny

Crackdetection~%!

24/2596

20/2580

23/2592

25/25100

No crackdetection~%!

19/2576

12/2548

11/2544

13/2552

Combinedaccuracy~%!

43/5086

32/5064

34/5068

38/5076

aData shown asx/y, wherex5number of correct detections out of totalyimages; %5accuracy.

Fig. 14. Canny for crack image of Fig. 2


beltee

s.er-ty

teeyhstd

Fig. 15. Canny for no-crack image of Fig. 3

tneio

i

.nc-henoce

than that of no-crack images. The overall accuracy was 34 ou50 ~68%!. Again, the texture of the concrete is the reason behithe large number of misclassifications. This algorithm would rquire significant changes to be of use in concrete crack detectFigs. 11 and 12 are the isolated cracks for two bridge imagesdetermined by the Sobel technique, and Fig. 13 is the imagetensity data for the 50 images used in the Sobel analysis.


ofd-n.asn-

CannyThe Canny edge-detection technique performed better than Soand FFT. The Canny uses a prefilter blur that seems to eliminanoise and aids in the correct classification of most images. Thoverall accuracy was 38 out of 50~76%!. The average intensity ofcrack images was 18% greater than that of no-crack images. Fig14 and 15 are the isolated cracks for two bridge images as detmined by the Canny technique, and Fig. 16 is the image intensidata for the 50 images used in the Canny analysis.

In summary, the fast Haar transform had the most accuracrack detection. The overall accuracy of the FHT was 86%. ThCanny edge detector performed relatively well with an accuracof 76%. The Sobel edge detector performed relatively poorly witan accuracy of 68% due to the noise in concrete images. The faFourier transform also responded poorly due to the texture anpatterns in the concrete images with an overall accuracy of 64%As shown in Table 1, there were far more false positives thafalse negatives in the detection of cracks, where a positive detetion indicates the presence of a crack regardless of whether timage has a crack or not, while a negative detection suggestscracks in an image. For example, the FHT indicated the presenof cracks in images that have actual cracks~correct positive! for24 out of the 25 images used, and 1 false positive~Table 1!.

Fig. 16. Intensity data for Canny

TOBER 2003

helyasn bee-oreece

teonthein-e

lo-ol-t

efnd-

ecieonandeinrd

aseortork

ob

on

e

n.’’-

i-

enti-

t

-

ea

o-inw

.’’

vi-

-ia-

t

e

Optimizing the criteria for threshold selection can enhance tresults. By raising or lowering the threshold value, it is generalpossible to alter the distribution of false positives and false negtives without a major impact on the ranking of the algorithmused in the paper. Once an algorithm is selected, the results cafurther improved to maximize the performance by focusing thresearch efforts on developing optimum criteria for threshold slection that are specific to concrete bridge images. The perfmance of the Haar wavelet was not a surprise since, like othmembers of the wavelet family, it has the ability to capture imagdiscontinuities. This is due to the fact that wavelets are spalimited, which is not true for FFT, and thus have the ability toshow local information in space~e.g., cracks!.

Concluding Remarks

The traditional visual inspection method used to monitor concrebridges is costly and time-consuming. Automated crack-detectitechniques that limit the necessity of human inspection havepotential to lower the cost and time associated with surfacespection of concrete bridges. However, engineers have behighly skeptical of using automated NDE systems and technogies such as imaging techniques. Much of this has been due tlack of standardization and to the difficulty in using the technoogy ~equipment! and/or interpreting the results. The developmenof commercial systems, widespread collection of data on thefectiveness of methods, drafting and revision of standards acodes, and training in NDE methods will hopefully work to rectify these difficulties.

Highway departments must begin to include automated insption techniques in their bridge evaluation programs and strategto enhance the quality of bridge condition data used in decisimaking. Crack-detection techniques are only one module inautomated bridge monitoring system. Other modules may inclua decision system that reads the data produced by the imagtechniques and formulates a plan for maintenance, real-time haware components~cameras and microprocessors!, and possiblywireless communication technologies for transmitting data tocentral bridge management system location. This paper focuon the crack-detection module and provided a framework fevaluating image-processing techniques in support of an aumated bridge monitoring system. Such an evaluation framewois necessary for the selection of the best techniques for the prlem at hand.

Acknowledgments

This project was funded in part by a National Science Foundati~NSF! Research Experiences for Undergraduates~REU! GrantNo. 9820310 and in part by an NSF MRI Grant No. 0215356. Ththird writer is an REU participant on this project.

JOURNAL OF C

-

e

-r

n

a

-

-s

g-

d

-

-

References

AASHTO. ~1988!. Keeping America moving—The bottom line, Washing-ton, D.C.

AASHTO. ~1993!. Guidelines for bridge management systems, Washing-ton, D.C.

AASHTO. ~2000!. Manual for condition evaluation of bridges, Washing-ton, D.C.

Abudayyeh, O., Weber, J., and Abdel-Qader, I.~2000!. ‘‘A survey ofnon-destructive testing methods for bridge condition evaluatioTechnical Rep. No. CEM-00-05, Dept. of Civil and Construction Engineering, Western Michigan Univ., Kalamazoo, Mich.

Alageel, K., and Abdel-Qader, I.~2002!. ‘‘Haar transform use in imageprocessing.’’Technical Rep., Dept. of Electrical and Computer Engneering, Western Michigan Univ., Kalamazoo, Mich.

Bachman, N., and Beckenstein, F.~2000!. Wavelet analysis, Springer,New York.

Brecher, A.~1995!. ‘‘Infrastructure: A national priority.’’SWE,4~6!, 14–16.

Canny, J. F.~1986!. ‘‘A computational approach to edge detection.’’IEEETrans. Pattern Anal. Mach. Intell.,8~6!.

Cooley, J. W., and Tukey, J. W.~1965!. ‘‘An algorithm for the machinecalculation of complex Fourier series.’’Math. Comput.,19~90!.

El-Korchi, T. ~1990!. ‘‘An engineering approach to automated pavemsurface distress evaluation.’’Proc., Pavement Distress Data Recogntion Seminar.

Gole, B. ~1985!. ‘‘Management vs. crisis reaction.’’APWA Reporter.Hass, R., Hudson, W. R., and Zaniewski, J.~1994!. Modern pavemen

management, Krieger Publishing Co., Malabar, Fla.Klassen, G., and Swindall, B.~1993!. ‘‘Automated crack detection system

implementation in ARAN.’’Proc., Digital Image Processing: Techniques and Applications in Civil Engineering Conf., ASCE, NewYork.

Koutsopoulos, H. N., and Downey, A. B.~1993!. ‘‘Primitive-based clas-sification of pavement cracking images.’’J. Transp. Eng.,119~3!,402–418.

Lee, H.~1990!. ‘‘Evaluation of PAVEDEX computerized pavement imagprocessing system in Washington.’’Proc., Pavement Distress DatRecognition Seminar.

Lee, H. ~1993!. ‘‘Survey: Fundamental pavement crack imaging algrithms.’’ Digital Image Processing: Techniques and ApplicationsCivil Engineering, J. D. Frost and J. R. Wright, eds., ASCE, NeYork, 195–202.

Longenecker, K.~1990!. ‘‘Pavement surface video image work in IdahoProc., Pavement Distress Data Recognition Seminar.

Parker, J. R.~1997!. Algorithms for image processing and computersion, Wiley, New York.

Portland Cement Association~PCA!. ~1995!. ‘‘Concrete pavement engineering.’’Seminar Handouts, American Concrete Pavement Assoction and the Portland Cement Association, Skokie, Ill.

Rens, K. L., and Greimann, L. F.~1997!. ‘‘Ultrasonic approach for non-destructive testing of civil infrastructure.’’J. Perform. Constr. Facil.,11~3!, 97–104.

Roberts, E., and Shepard, R.~2000!. ‘‘Bridge management for the 21scentury.’’ Transportation Research Record 1696, Transportation Re-search Board, Washington, D.C., 5B0138.

Saito, M., and Sinha, C.~1990!. ‘‘Data collection and analysis of bridgrehabilitation and maintenance costs.’’Transportation ResearchRecord 1276, National Research Council, Washington, D.C.

Silano, L.~1993!. Bridge inspection and rehabilitation: A practical guide,Wiley, New York.


analysis of edge-detection techniques for crack...

Documents