Research ArticleWeld Inspection Based on Radiography Image Segmentationwith Level Set Active Contour Guided Off-Center Saliency Map

Mohamed Ben Gharsallah and Ezzeddine Ben Braiek

University of Tunis Tunis National Higher School of Engineering (ENSIT) Research CEREP Unit5 Avenue Taha Hussein 1008 Tunis Tunisia

Correspondence should be addressed to Mohamed Ben Gharsallah medgharsallahyahoofr

Received 7 September 2015 Revised 28 November 2015 Accepted 30 November 2015

Academic Editor Ying Li

Copyright copy 2015 M Ben Gharsallah and E Ben Braiek This is an open access article distributed under the Creative CommonsAttribution License which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

Radiography is one of the most used techniques in weld defect inspection Weld defect detection becomes a complex task whenuneven illumination and low contrast characterize radiographic images In this paper we propose a new active contour based levelset method for weld defect detection in radiography images An off-center saliency map exploited as a feature to represent imagepixels is embedded into a region energy minimization function to guide the level set active contour to defects boundaries The aimbehind using salient feature is that a small defect can frequently attract attention of human eyes which permits enhancing defectsin low contrasted image Experiment results on different weld radiographic images with various kinds of defects show robustnessand good performance of the proposed approach comparing with other segmentation methods

1 Introduction

Nondestructive testing (NDT) is widely used in many fieldsprincipally for serious applications where malfunction canbe catastrophic such as welds of pressure vessels aircraftand power plants One of the most famous techniques usedin weld inspection is radiography which is based on thetransmission of X-rays or gamma rays through an objectto generate a radiological image on a photographic plate(Figure 1) Unfortunately the traditional interpretation ofradiography images by artificial methods is subjective time-consuming and easy to cause fatigue in order to improve theautomation level and avoid drawbacks of manual interpreta-tion it is desirable to develop some forms of computer-aidedsystems to assist the human interpreter in evaluating thequality of welded joints In general this system of automaticinspection should have the following stages [1 2] after digitalimage acquisition only a region of interest (ROI) is furtherprocessed some preprocessing may take place like noisereduction and contrast enhancement and then segmentationof regions that may represent defects is done as soon as thedefects are segmented features can be extracted and thengiven as input to classifiers to detect possible defects and

eventually to identify the exact defect type Moreover thedefect dimensions are compared to some acceptance criteriadefined by experts or international standards and a decisionis taken on the acceptability of the monitored weld

As shown in Figure 2 weld radiography image con-tains two main parts the base metal part and the weldseam part The weld region is brighter than the weld areaDefects are randomly found at the weld area with differentsmall shapes circular and rectangular Weld flaws can becategorized in various types like incomplete penetrationslag line slag inclusion cracks undercuts porosity andwormholes Porosity or gas cavity has rounded contours anddark shadows cracks are fine line straight or wandering indirection slag are line more or less interrupted parallel tothe edges of weld Radiography images are characterized bya low contrast between defects and background (weld) andsmall defects with blurred and unsharpened edgesMoreoveruneven illumination is frequently found in radiographyimages which is a nonuniform light distributed generally atthe middle of weld area For weld inspectors these factorsmake defect localisation and segmentation with conventionalsegmentation methods a complicated mission To overcomethese difficulties and to facilitate human weld inspection we

Hindawi Publishing CorporationAdvances in Materials Science and EngineeringVolume 2015 Article ID 871602 10 pageshttpdxdoiorg1011552015871602

2 Advances in Materials Science and Engineering

Radiation

Welding seam

Plate

Digital radiographicimage

Figure 1 Weld radiography image acquisition [6]

Met

alW

eld se

am

Defects

Figure 2 Example of weld radiography image with defects porosi-ties (circular) and a longitudinal crack

propose in this paper an image segmentation method forradiography weld defect extraction

In literature many radiography weld defect detectionmethods are proposed [3] We find local and global thresh-olding approaches [4] texture analysis based methods [5]watershed [6] artificial neural network [7 8] and activecontours [9] Segmentation with deformablemodels or activecontours seems to be quite suitable for radiographic imagesto extract defects because of many reasons principally theability of integration of various image properties such asedge and region informationwithminimisationmethods andcurve theory Several research works have been explored andmany active contour models are proposed In general activecontour models can be categorized into two different classesedge and region based models Edge based models [10 11]uses the edge information like image gradient to drive theactive contour toward the object boundaries and to stop itthere these kinds of models are sensitive to noise and toinitial active contour position which should be initializednear object boundaries Moreover the boundary leakageproblem at weak edges is a major drawback of edge basedmodels since they relymuch on the gradient value Comparedwith edge based models region based models depend onstatistical information inside and outside of regions delimitedby the contour thus they are less sensitive to the noise andto poor edges Moreover they are less dependent to theinitialization since they exploit the global region informationof the image statistics

One of the most popular region based approaches isthe Chan and Vese model well known as C-V [12] C-V model energy function is a simplification of MumfordShah formulation [13] It has been successfully applied inmany applications for images with two regions this modelis less sensitive to image noise and contour initializationwhich can be located everywhere in the image Howeverthe major weakness of C-V model is the segmentationof images with intensity inhomogeneities where pixels ofthe same object have a nonuniform gray level intensitydistribution To resolve this problem active contours withlocalized energy functions have been proposed The idea isto compute statistical information only in a local windowin the pixel neighbourhood One of the famous local regionbased active contours is proposed by Wang et al [14 15]called the local binary fitting (LBF) model In this methodtwo fitting energies are used to calculate LBF energy functionthat approximates the local image intensities means insideand outside of the contour using a Gaussian kernel The LBFmodel solved the problemcaused by intensity inhomogeneityHowever this model is sensitive to initial contour locationand it increases greatly the computational complexity Zhanget al [16] proposed a selective local global level set activecontour known as IVCmodel which introduces a new regionsigned pressure force (SPF) function Zhang model uses aGaussian smoothing kernel to regularize the level set functionwhich decreases significantly the computation time Themain advantage of this model is the ability to select local orglobal segmentation so the user can choose to segment oneobject or the whole image Nevertheless experiments showthat this model is very sensitive to nonuniform illuminationand to low contrasted images like radiography images Zhanget al [17] proposed also a region local level set active contourwhere a local image fitting (LIF) energy function is usedto guide the level set active contour to object boundariesbased on local statistical information of the image LIF energyfunction determines differences between the fitted image inthe LBF model and the original image In addition a filteringmethod with a Gaussian kernel is applied to regularize levelset function iteratively LIF active contour model is wellconsuming computation time besides it is very sensitive tothe local window sizes which should be adjusted carefullychoosing a small window size around pixel permits detectionof small objects but increases sensitivity to noise Song andYan [18] proposed a local level set active contour to segmentsmall defects found in hot rolled steel coloured images Themain idea is to fuse an image feature called the saliencymap in the active contour energy formulationThis approachshows high performance in segmentation of defects withsmall regions particularly with rounded shapes in colouredimages Nevertheless the saliency map used is adapted tocoloured images so three information channels are neededMuch information might be missed when using images withone channel like radiography images Moreover Song modeluses a Gaussian filter in the computation of the saliency mapwhich is not suitable for images with low contrast images Inthe other side this approach is a level set local based modelwhich increases the computation time considerably besides

Advances in Materials Science and Engineering 3

segmentation accuracy depends greatly to the level set localwindow sizes

In this work we propose a new global level set activecontour to segment weld defects in radiography images Animage feature called the off-center saliency map computedwith integral images is embedded in the energy formulationof a global region level set active contour The remainderof this paper is organized as follows the proposed methodis presented in Section 2 tests and experimental results aredepicted in Section 3 and we finish the paper by a conclusionin Section 4

2 Proposed Level Set Active Contour Model

In this section we present the modified level set activecontour We show the off-center saliency map extrac-tion method in the first part and the new level setactive contour energy formulation after fusion with theextracted saliency map feature is illustrated in the secondpart

21 Off-Center Saliency Map Computation Systems thatmimic the biological attention system arewidely developed toextract the saliency map which aims to select the interestingparts of the sensory input data in order to reduce thevast amount of information that a computer vision systemnormally needs to process In this area two types of retinalganglion cells are defined ldquoon-centerrdquo and ldquooff-centerrdquo Anon-center cell is stimulated when the center of its receptivefield is exposed to light and is inhibited when the surroundis exposed to light off-center cells have just the opposite

reaction Two image features that imitate the behavior of thetwo cells are computed and known as on-center and off-center saliency maps In weld radiography images defectscorrespond to dark regions surrounded by brighten back-ground therefore we decide to use the off-center saliencymap as an image feature to guide the level set active contourto defect boundaries Off-center saliency map is calculatedusing center-surround difference Many approaches are usedto achieve computation we find Gabor filters [19] differenceof Gaussian filters [20] and spectral residual [21] RecentlyMontabone and Soto [22] present a fast method to computecenter-surround differences with rectangular filters basedon the concept of integral images (or summed area tables)introduced first by Viola and Jones [23] this approach allowsspeeding up the calculations considerably and preservingobjects borders This approach is used in our work The off-center saliency submaps are calculated by a difference centerand surround with this relation

119878119888(119909 119910) = max surround (119909 119910 119888) minus center (119909 119910) 0 (1)

where ldquo119888rdquo represents the surround size chosen empirically forour application as 119888 = 2 4 8 16 and center(119909 119910) representsthe gray level pixel at the surround middle 119868(119909 119910)

The surround is calculated in a fast time using integralimage 119868

119892(119909 119910) of the image 119868(119909 119910)

119868119892(119909 119910) =

119909

sum

119894=0

119910

sum

119895=0

119868 (119894 119895) (2)

The surround is the local average in neighbourhoodwindow size ldquo119888rdquo

surround (119909 119910 119888) =119868119892(119909 + 119888 119910 + 119888) minus 119868

119892(119909 minus 119888 119910 + 119888) minus 119868

119892(119909 + 119888 119910 minus 119888) + 119868

119892(119909 minus 119888 119910 minus 119888) minus 119868 (119909 119910)

(2119888 + 1)2

minus 1 (3)

The computed off-center saliency submaps are summedpixel by pixel in a one map 119878(119909 119910) as follows

119878 (119909 119910) = sum

119888

119878119888(119909 119910) (4)

119878(119909 119910) is the off-center saliencymapwhichwill be used ina further step with the level set active contourThe advantageof the feature 119878(119909 119910) in defect enhancement is demonstratedby an example on radiographic weld image in Figure 3 Asshown the weld radiography image has a low level contrastand a background affected by uneven illumination defectsand background are not easily distinguishable thereforedefect extraction is very difficult Figure 3(b) shows the off-center saliency map 119878(119909 119910) as we can see the differencebetween the defects and the background is highly enhancedthe surface plot of 119878(119909 119910) shown in Figure 3(c) shows alsohow defects gray level are well amplified

22 Level Set Active Contour Energy Formulation The off-center saliency map 119878(119909 119910) obtained in the previous sectionis used as statistical information representing pixels in theimage The feature image 119878(119909 119910) is embedded in the formu-lation of a level set active contour using a global Gaussiandistribution fitting energy The aim is to segment image intotwo regions defect and background Ω

1 Ω2 with a contour

ldquo119862rdquo separating these two regions depending on statisticalinformation inside and outside of the off-center saliencymap 119878(119909 119910) The segmentation is achieved with energy min-imization [19] We define an energy function to the contourldquo119862rdquo composed of three terms the salient energy 119864

1(119862) the

regularization energy 1198642(119862) and 119864

3(119862) the reinitialization

energy

119864 (119862) = 1198641(119862) + 119864

2(119862) + 119864

3(119862) (5)

4 Advances in Materials Science and Engineering

(a) (b)

300

200

00

50

100

150

200

1

08

06

04

02

0

100

(c)

Figure 3 Example of off-center saliency map (a) ROI radiography weld image (b) off-center saliency map 119878(119909 119910) and (c) surface plot ofoff-center saliency map

The first term 1198641(119862) attracts the contour ldquo119862rdquo to defect

edges We choose to minimize the Bayes error [24] definedas follows

1198641(119878 (119909 119910) 119862) = minusint

Ω1

log (1198751(119878 (119909 119910) Ω

1)) 119889119909 119889119910

minus intΩ2

log (1198752(119878 (119909 119910) Ω

2)) 119889119909 119889119910

(6)

where Ω1 Ω2are the regions inside and outside the contour

119862 1198751 1198752are two probability density functions (PDF) 119878(119909 119910)

is the off-center saliency mapBased on the level set theory [16] the contour 119862 is

embedded as the zero level of a level set function 120601(119909 119910)

119862 = (119909 119910) isin Ω 120601 (119909 119910) = 0 (7)

Thus (6) becomes

1198641(119878 (119909 119910) 120601 (119909 119910))

= minusintΩ

log (1198751(119878 (119909 119910))119867 (120601 (119909 119910))) 119889119909 119889119910

minus intΩ

log (1198752(119878 (119909 119910)) (1 minus 119867 (120601 (119909 119910)))) 119889119909 119889119910

(8)

where Ω is the image domain and 119867(120601) is the Heavisidefunction

1198751 1198752are two probability density functions assumed as a

Gaussian distribution with means 1198981 1198982and variances 120590

1

1205902 Consider

119875119894(119878 (119909 119910)) =

1

radic2120587120590119894

exp(minus(119878 (119909 119910) minus 119898

119894)2

21205902

119894

)

119894 = 1 2

(9)

Therefore (8) can be written as

1198641(120601 (119909 119910) 119898

1 1198982 1205901 1205902)

= intΩ

(log (radic2120587) + log (1205901) +

(119878 (119909 119910) minus 1198981)2

21205902

1

)

sdot 119867 (120601 (119909 119910)) 119889119909 119889119910

+ intΩ

(log (radic2120587) + log (1205902) +

(119878 (119909 119910) minus 1198982)2

21205902

2

)

sdot (1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

(10)

Advances in Materials Science and Engineering 5

Means11989811198982and variances1205902

112059022of 119878(119909 119910) respectively

inside and outside the contour ldquo119862rdquo can be calculated asfollows

1198981=

intΩ

119878 (119909 119910) sdot 119867 (120601 (119909 119910)) 119889119909 119889119910

intΩ

119867(120601 (119909 119910)) 119889119909 1198891199101205902

1

=

intΩ

(119878 (119909 119910) minus 1198981)2

sdot 119867 (120601 (119909 119910)) 119889119909 119889119910

intΩ

119867(120601 (119909 119910)) 119889119909 119889119910

1198982=

intΩ

119878 (119909 119910) sdot (1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

intΩ

(1 minus 119867 (120601 (119909 119910))) 119889119909 1198891199101205902

2

=

intΩ

(119878 (119909 119910) minus 1198982)2

sdot (1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

intΩ

(1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

(11)

1198642(120601) is a regularization energy added to keep a smooth

contour of the segmented region so we should minimize thelength of the contourwhich can be calculated by the followingrelation

1198642(120601 (119909 119910)) = int

Ω

1003816100381610038161003816nabla119867 (120601 (119909 119910))1003816100381610038161003816 119889119909 119889119910 (12)

When evolving level set function usually it may be toosteep or flat so we add a third energy term119864

3(120601) to reinitialize

the level set function and to maintain the signed distanceproperty We use a distance regularized level set evolution(DRLSE) proposed by Wang et al [24]

1198643(120601 (119909 119910)) = int

Ω

119875 (1003816100381610038161003816nabla120601 (119909 119910)

1003816100381610038161003816) 119889119909 119889119910 (13)

where 119875 is called a double well potential function defined by

119875 (119911) =

1

21205872(1 minus cos (2120587119911)) if 119911 lt 1

1

2(119911 minus 1) if 119911 gt 1

(14)

As a result the total energy function 119864(120601) can be writtenas

119864 (1206011198981 1198982 1205901 1205902) = 120572119864

1(1206011198981 1198982 1205901 1205902)

+ 1205731198642(120601) + 120574119864

3(120601)

(15)

We have added 120572 120573 120574 as controlling parameters We notethat the parameter 120572 controls the signed distance propertyof the level set function and the parameter 120573 governs itssmoothness The parameter 120574 permits attracting the level setfunction to defect regions

Now we keep all the variables fixed except for 120601 min-imization of the total energy 119864(120601) with respect to 120601 isequivalent to solving the gradient descent flow equation

120597120601

120597119905= minus

120597119864 (120601)

120597120601 (16)

We obtain the following evolution equation

120597120601

120597119905= 120572 div (119889119901 (1003816100381610038161003816nabla120601

1003816100381610038161003816) nabla120601) 120573120575 (120601) div(nabla120601

1003816100381610038161003816nabla1206011003816100381610038161003816

)

+ 120574120575 (120601) (1198651minus 1198652)

(17)

120575(120601) is the Dirac function 1198651 1198652 119889119901 are defined as

follows

119865119894= log (120590

119894) +

(119878 (119909 119910) minus 119898119894)2

21205902

119894

119894 = 1 2

119889119901 (119911) =1198751015840

(119911)

119911

(18)

The level set evolution equation (17) is composed of 3terms the first term in the right side allows keeping theregularity of the level set function the second term preservesa smooth contour and the third term segments the image intodefect and background

Using a simple finite difference explicit scheme to (17) weobtain the discrete level set formulation as follows

120601119899+1

= 120601119899

+ Δ119905 [120572 div (119889119901 (1003816100381610038161003816nabla1206011198991003816100381610038161003816) nabla120601

119899

)

+ 120573120575 (120601119899

) div(nabla120601119899

1003816100381610038161003816nabla1206011198991003816100381610038161003816

) minus 120574120575 (120601119899

) (1198651minus 1198652)]

(19)

where 119899 Δ119905 are the index iteration number and time steprespectively

The main stages of the proposed scheme for segmentingdefects can be summarized as follows

Step 1 Input image

Step 2 Select a region of interest (ROI)

Step 3 Initialize parameters Δ119905 119899 120572 120573 120574

Step 4 Compute off-center saliency map 119878(119909 119910)

Step 5 Compute the initial level set mask 1206010

Step 6

For a fixed number of iterations 119899Compute average and variance119898

11198982 1205901 1205902

Update level set function 120601 with (19)

Step 7 Keep the zero level of the function 120601

3 Experimental Results

In this section the proposed level set active contour model istested and evaluated on a set of radiography images of weldjoints obtained from Federal institute of Material Researchand Testing (Bam) [25] The radiograph films have beenscanned with the scanner LS85 SDR from Lumisys mostly

6 Advances in Materials Science and Engineering

ROI

(a)

RD 0

1

(b)

RD 0

1

(c)

RD 0

1

(d)

Figure 4 Weld defect detection with the proposed active contour (a) the whole weld radiogram (b) ROI selected image (c) off-centersaliency map and (d) final segmentation with proposed level set active contour (blue color)

ROI

(a)

RD 0

2

(b)

RD 0

2

(c)

RD 0

2

(d)

Figure 5 Weld defect detection with the proposed active contour (a) the whole weld radiogram (b) ROI selected image (c) off-centersaliency map and (d) final segmentation with proposed level set active contour (red color)

Advances in Materials Science and Engineering 7RD

03

RD 0

4RD

05

(a)

RD 0

3RD

04

RD 0

5

(b)RD

03

RD 0

4RD

05

(c)

Figure 6 Comparative defect detection tests (a) ROI selected image (b) result of LIF active contour and (c) result of the proposed method

in high density mode The original 12-bit data depth wasrescaled to 8 bits with a linear LUT proportional to opticalfilm density by visual adjustment to the image content Thepixel size is 403 micron (630 dpi) and the images are 8-bit gray values Due to the big size of radiography weldimages the nonuniform illumination and the small defectshape it is difficult to detect the presence of small defectsand determine accurately their sizes during the radiogramvisualization Consequently for the seeking of simplifyingthe task one could begin by selecting the region of interestROI which can be considered as the parts of the imagewhere the radiograph interpreters suspect the presence ofimperfectionsThe selection of the ROI prevents the operatorto make treatments on the irrelevant regions of the imageMoreover it allows reducing the computing time for real-time applications noting that the technique of ROI local-ization is commonly used by researchers in several worksAfter the selection of an ROI region the obtained images areresized to 250 times 350

We present in Figures 4 and 5 the main steps of theproposed segmentation algorithm on two weld radiographyexamples First row corresponds to thewhole weld radiogramthen a region of interest (ROI) is selected Next images inFigures 4(b) and 5(b) show the selected region to processincluding various defects like porosities and slags distributedat the weld middle The off-center saliency map is shownin the next Figures 4(c) and 5(c) as we can see defects are

enhanced and the background is suppressed Figures 4(d)and 5(d) depict the defect contour detection with a blue andred colour obtained with the proposed off-center saliencylevel set active contour Even though the low contrast andthe nonuniform illumination we can see that the most partsof defects are segmented with low false detections We notethat parameters of the proposed method are set as Δ119905 = 01120572 = 004 120573 = 0001 times 255 times 255 120574 = 15 and 119899 = 30

31 Comparative Test 1 We show a comparison between theproposed model and a level set active contour dependingonly on gray level image intensity proposed by Zhang et al[17] and called the local image fitting active contour LIFLIF model Matlab implementation is given in authorrsquos web-site (httpkaihuazhangnetJ papersPR 10rar) Figure 6(b)presents results of the local fittingmodel LIF and the results ofthe proposed active contour model are shown in Figure 6(c)Parameters of the proposed algorithm are fixed as Δ119905 = 01120572 = 002 120573 = 0001 times 255 times 255 120574 = 10 and 119899 = 30

Through a visual evaluation the superiority of the pro-posed level set active contour can be verified Local fittingactive contour LIF [17] fails to obtain satisfactory resultsdue to the influence of the clutter in the background andthe low contrast of defects Much false detection and a lotof defects are not detected The best results are obtainedwith the proposed level set active contour The most parts ofdefects are identified with low false detections A quantitative

8 Advances in Materials Science and EngineeringRD

06

RD 0

7RD

08

RD 0

9

(a)

RD 0

6RD

07

RD 0

8RD

09

(b)

RD 0

6RD

07

RD 0

8RD

09

(c)

Figure 7 Comparative defect segmentation tests (a) ROI selected image (b) results of Songmethod and (c) results of the proposedmethod

comparison between the proposed active contour and LIFmethod is shown in the Table 1 We compute a segmentationevaluation measure used by many researchers called the 119865-measure [26] To compute this evaluation criterion we needideal image segmentation and the proposed segmentationresults the two results should be given in binary masks Notethat ground truth or ideal segmentations used are obtained

from expert visual interpretation We note also that a highervalue (max 1) denotes accurate segmentation The detectionrates are shown in Table 1 which demonstrates the highperformance of our method comparing with LIF approachMoreover the computation time (cpu-time) is computed fortwo methods The step time algorithm is fixed to 01 forthe two methods Algorithms are implemented on Matlab

Advances in Materials Science and Engineering 9

Table 1 Evaluation of segmentation accuracy and computationtime

119865-measure Cpu-Time (sec)LIF Prop method LIF Prop method

RD02 052 076 354 41RD 03 042 081 254 38RD04 062 087 186 35

Table 2 Evaluation of cpu-time convergence

Cpu-time (sec)Song Prop method

RD06 826 385RD07 941 481RD 08 967 523RD09 966 477

R2010a (on a PCCore i5 29GHz) Obtained values show thatour proposed model is very fast compared with local activecontour LIF

32 Comparative Test 2 We show in this paragraph a com-parative test between the proposed method and saliencyconvex active contour model proposed by Song and Yan[18] known as SCACM Images in Figure 7(a) correspondsto regions of interest (ROI) selected from different weldradiogram films containing various kinds of defects likeporosities and horizontal and vertical thin cracks Figure 7(b)presents SCACM method detection results and Figure 7(c)shows the segmentation results with the proposed method

Obtained results in Figure 7 show that Song methoddetects only defects with small and round shapes due to theinfluence of the local window size used in this method In theother side our method does not need to use a local windowas is explained before we use a global Gaussian distributionapproximation of the off-center saliency intensities Theproposed method detects defects with various forms andsizes with low false detections Most parts of defects areextracted despite their low contrasts Using Song methodmany low contrasted defects aremissed and not identified Asin previous experiment the convergence time of twomethodsevaluated in Table 2 shows that proposed model achievesdefect segmentation in a fast time comparing with Songmethod We choose parameters of the proposed algorithm as120572 = 001 120573 = 0001 times 255 times 255 120574 = 13 and 119899 = 30 We notealso that we have used the Matlab implementation given bySCACM authors in their website (httpfacultyneueducnyunhyanWebpage20for20articleSCACMDemoSCACMrar)

4 Conclusion

Detection of small defects in low contrasted radiographyimages corrupted with uneven illumination is very compli-cated Usually proposed methods in this area have limitedresults In this work our aim is to improve robustness of

weld defect segmentation in radiography images in order toobtain satisfactory results We have developed a new methodbased on level set active contour guided with an off-centersaliency map The segmentation is achieved when an energyfunction is minimized Different tests on weld radiographyimages with various kinds of defects prove efficiency androbustness of the proposed method A comparison test isshown with two kinds of level set active contour models Thefirst depends only on gray level intensity and the second usesa saliencymap to derive the segmentation active contourThecomparison tests reveal that the proposed method permitsovercoming the problem of nonuniform illumination andthe low contrast level in radiography weld images Moreoverour method shows a fast convergence time comparing withother methods However some weakness of the proposedmethod must be studied in a future work To reduce humaninteraction it is recommended to develop a method to selectautomatically the region of the interest It is also interestingto validate this segmentation method to another radiographyimage application like medical area

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] I Valavanis and D Kosmopoulos ldquoMulticlass defect detectionand classification in weld radiographic images using geometricand texture featuresrdquo Expert Systems with Applications vol 37no 12 pp 7606ndash7614 2010

[2] J Zapata R Vilar and R Ruiz ldquoPerformance evaluation of anautomatic inspection system of weld defects in radiographicimages based on neuro-classifiersrdquo Expert Systems with Appli-cations vol 38 no 7 pp 8812ndash8824 2011

[3] C Stolojescu-Crisan and S Holban ldquoA comparison of X-rayimage segmentation techniquesrdquo Advances in Electrical andComputer Engineering vol 13 no 3 pp 85ndash92 2013

[4] A Mahmoudi and F Regragui ldquoWelding defect detection bysegmentation of radiographic imagesrdquo in Proceedings of theWorld Congress on Computer Science and Information Engineer-ing (WRI rsquo09) vol 7 pp 111ndash115 Los Angeles Calif USAMarch2009

[5] D Mery and M A Berti ldquoAutomatic detection of weldingdefects using texture featuresrdquo Insight Non-Destructive Testingand Condition Monitoring vol 45 no 10 pp 676ndash681 2003

[6] M A Carrasco and D Mery ldquoSegmentation of welding defectsusing a robust algorithmrdquo Materials Evaluation vol 62 no 11pp 1142ndash1147 2004

[7] E S Amin ldquoApplication of artificial neural networks to evaluateweld defects of nuclear componentsrdquo Journal of Nuclear andRadiation Physics vol 3 no 2 pp 83ndash92 2008

[8] S-B Zhou A-Q Shen and G-F Li ldquoConcrete image segmen-tation based on multiscale mathematic morphology operatorsand Otsu methodrdquo Advances in Materials Science and Engineer-ing vol 2015 Article ID 208473 11 pages 2015

[9] S Osher and N Paragios Geometric Level Set Methods inImaging Vision and Graphics Springer 2003

10 Advances in Materials Science and Engineering

[10] X-F Wang D-S Huang and H Xu ldquoAn efficient local Chan-Vese model for image segmentationrdquo Pattern Recognition vol43 no 3 pp 603ndash618 2010

[11] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998

[12] T F Chan and L A Vese ldquoActive contours without edgesrdquo IEEETransactions on Image Processing vol 10 no 2 pp 266ndash2772001

[13] DMumford and J Shah ldquoOptimal approximations by piecewisesmooth functions and associated variational problemsrdquo Com-munications on Pure and Applied Mathematics vol 42 no 5pp 577ndash685 1989

[14] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009

[15] X Liu S-J Peng Y-MCheung Y Y Tang and J-XDu ldquoActivecontours with a joint and region-scalable distributionmetric forinteractive natural image segmentationrdquo IET Image Processingvol 8 no 12 pp 824ndash832 2014

[16] K Zhang L Zhang H Song and W Zhou ldquoActive contourswith selective local or global segmentation a new formulationand level set methodrdquo Image and Vision Computing vol 28 no4 pp 668ndash676 2010

[17] K Zhang H Song and L Zhang ldquoActive contours driven bylocal image fitting energyrdquo Pattern Recognition vol 43 no 4pp 1199ndash1206 2010

[18] K Song and Y Yan ldquoMicro surface defect detection methodfor silicon steel strip based on saliency convex active contourmodelrdquoMathematical Problems in Engineering vol 2013 ArticleID 429094 13 pages 2013

[19] L Itti and C Koch ldquoComputational modelling of visual atten-tionrdquo Nature Reviews Neuroscience vol 2 no 3 pp 194ndash2032001

[20] R Achantay S Hemamiz F Estraday and S SusstrunkyldquoFrequency-tuned salient region detectionrdquo in Proceedings ofthe IEEE Computer Society Conference on Computer Vision andPatternRecognition (CVPR rsquo09) pp 1597ndash1604 IEEE June 2009

[21] X Hou and L Zhang ldquoSaliency detection a spectral residualapproachrdquo in Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition (CVPR rsquo07) pp 1ndash8 Minneapo-lis Minn USA June 2007

[22] S Montabone and A Soto ldquoHuman detection using a mobileplatform and novel features derived from a visual saliencymechanismrdquo Image and Vision Computing vol 28 no 3 pp391ndash402 2010

[23] P Viola and M J Jones ldquoRobust real-time face detectionrdquoInternational Journal of Computer Vision vol 57 no 2 pp 137ndash154 2004

[24] L Wang L He A Mishra and C Li ldquoActive contours drivenby local Gaussian distribution fitting energyrdquo Signal Processingvol 89 no 12 pp 2435ndash2447 2009

[25] httpwwwbamdeenindexhtm[26] httpsenwikipediaorgwikiPrecision and recall

Submit your manuscripts athttpwwwhindawicom

ScientificaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CorrosionInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Polymer ScienceInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CeramicsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CompositesJournal of

NanoparticlesJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Biomaterials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

NanoscienceJournal of

TextilesHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Journal of

NanotechnologyHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

CrystallographyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CoatingsJournal of

Advances in

Materials Science and EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Smart Materials Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MetallurgyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

MaterialsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Nano

materials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal ofNanomaterials

2 Advances in Materials Science and Engineering

Radiation

Welding seam

Plate

Digital radiographicimage

Figure 1 Weld radiography image acquisition [6]

Met

alW

eld se

am

Defects

Figure 2 Example of weld radiography image with defects porosi-ties (circular) and a longitudinal crack

propose in this paper an image segmentation method forradiography weld defect extraction

In literature many radiography weld defect detectionmethods are proposed [3] We find local and global thresh-olding approaches [4] texture analysis based methods [5]watershed [6] artificial neural network [7 8] and activecontours [9] Segmentation with deformablemodels or activecontours seems to be quite suitable for radiographic imagesto extract defects because of many reasons principally theability of integration of various image properties such asedge and region informationwithminimisationmethods andcurve theory Several research works have been explored andmany active contour models are proposed In general activecontour models can be categorized into two different classesedge and region based models Edge based models [10 11]uses the edge information like image gradient to drive theactive contour toward the object boundaries and to stop itthere these kinds of models are sensitive to noise and toinitial active contour position which should be initializednear object boundaries Moreover the boundary leakageproblem at weak edges is a major drawback of edge basedmodels since they relymuch on the gradient value Comparedwith edge based models region based models depend onstatistical information inside and outside of regions delimitedby the contour thus they are less sensitive to the noise andto poor edges Moreover they are less dependent to theinitialization since they exploit the global region informationof the image statistics

One of the most popular region based approaches isthe Chan and Vese model well known as C-V [12] C-V model energy function is a simplification of MumfordShah formulation [13] It has been successfully applied inmany applications for images with two regions this modelis less sensitive to image noise and contour initializationwhich can be located everywhere in the image Howeverthe major weakness of C-V model is the segmentationof images with intensity inhomogeneities where pixels ofthe same object have a nonuniform gray level intensitydistribution To resolve this problem active contours withlocalized energy functions have been proposed The idea isto compute statistical information only in a local windowin the pixel neighbourhood One of the famous local regionbased active contours is proposed by Wang et al [14 15]called the local binary fitting (LBF) model In this methodtwo fitting energies are used to calculate LBF energy functionthat approximates the local image intensities means insideand outside of the contour using a Gaussian kernel The LBFmodel solved the problemcaused by intensity inhomogeneityHowever this model is sensitive to initial contour locationand it increases greatly the computational complexity Zhanget al [16] proposed a selective local global level set activecontour known as IVCmodel which introduces a new regionsigned pressure force (SPF) function Zhang model uses aGaussian smoothing kernel to regularize the level set functionwhich decreases significantly the computation time Themain advantage of this model is the ability to select local orglobal segmentation so the user can choose to segment oneobject or the whole image Nevertheless experiments showthat this model is very sensitive to nonuniform illuminationand to low contrasted images like radiography images Zhanget al [17] proposed also a region local level set active contourwhere a local image fitting (LIF) energy function is usedto guide the level set active contour to object boundariesbased on local statistical information of the image LIF energyfunction determines differences between the fitted image inthe LBF model and the original image In addition a filteringmethod with a Gaussian kernel is applied to regularize levelset function iteratively LIF active contour model is wellconsuming computation time besides it is very sensitive tothe local window sizes which should be adjusted carefullychoosing a small window size around pixel permits detectionof small objects but increases sensitivity to noise Song andYan [18] proposed a local level set active contour to segmentsmall defects found in hot rolled steel coloured images Themain idea is to fuse an image feature called the saliencymap in the active contour energy formulationThis approachshows high performance in segmentation of defects withsmall regions particularly with rounded shapes in colouredimages Nevertheless the saliency map used is adapted tocoloured images so three information channels are neededMuch information might be missed when using images withone channel like radiography images Moreover Song modeluses a Gaussian filter in the computation of the saliency mapwhich is not suitable for images with low contrast images Inthe other side this approach is a level set local based modelwhich increases the computation time considerably besides

Advances in Materials Science and Engineering 3

segmentation accuracy depends greatly to the level set localwindow sizes

In this work we propose a new global level set activecontour to segment weld defects in radiography images Animage feature called the off-center saliency map computedwith integral images is embedded in the energy formulationof a global region level set active contour The remainderof this paper is organized as follows the proposed methodis presented in Section 2 tests and experimental results aredepicted in Section 3 and we finish the paper by a conclusionin Section 4

2 Proposed Level Set Active Contour Model

In this section we present the modified level set activecontour We show the off-center saliency map extrac-tion method in the first part and the new level setactive contour energy formulation after fusion with theextracted saliency map feature is illustrated in the secondpart

21 Off-Center Saliency Map Computation Systems thatmimic the biological attention system arewidely developed toextract the saliency map which aims to select the interestingparts of the sensory input data in order to reduce thevast amount of information that a computer vision systemnormally needs to process In this area two types of retinalganglion cells are defined ldquoon-centerrdquo and ldquooff-centerrdquo Anon-center cell is stimulated when the center of its receptivefield is exposed to light and is inhibited when the surroundis exposed to light off-center cells have just the opposite

reaction Two image features that imitate the behavior of thetwo cells are computed and known as on-center and off-center saliency maps In weld radiography images defectscorrespond to dark regions surrounded by brighten back-ground therefore we decide to use the off-center saliencymap as an image feature to guide the level set active contourto defect boundaries Off-center saliency map is calculatedusing center-surround difference Many approaches are usedto achieve computation we find Gabor filters [19] differenceof Gaussian filters [20] and spectral residual [21] RecentlyMontabone and Soto [22] present a fast method to computecenter-surround differences with rectangular filters basedon the concept of integral images (or summed area tables)introduced first by Viola and Jones [23] this approach allowsspeeding up the calculations considerably and preservingobjects borders This approach is used in our work The off-center saliency submaps are calculated by a difference centerand surround with this relation

119878119888(119909 119910) = max surround (119909 119910 119888) minus center (119909 119910) 0 (1)

where ldquo119888rdquo represents the surround size chosen empirically forour application as 119888 = 2 4 8 16 and center(119909 119910) representsthe gray level pixel at the surround middle 119868(119909 119910)

The surround is calculated in a fast time using integralimage 119868

119892(119909 119910) of the image 119868(119909 119910)

119868119892(119909 119910) =

119909

sum

119894=0

119910

sum

119895=0

119868 (119894 119895) (2)

The surround is the local average in neighbourhoodwindow size ldquo119888rdquo

surround (119909 119910 119888) =119868119892(119909 + 119888 119910 + 119888) minus 119868

119892(119909 minus 119888 119910 + 119888) minus 119868

119892(119909 + 119888 119910 minus 119888) + 119868

119892(119909 minus 119888 119910 minus 119888) minus 119868 (119909 119910)

(2119888 + 1)2

minus 1 (3)

The computed off-center saliency submaps are summedpixel by pixel in a one map 119878(119909 119910) as follows

119878 (119909 119910) = sum

119888

119878119888(119909 119910) (4)

119878(119909 119910) is the off-center saliencymapwhichwill be used ina further step with the level set active contourThe advantageof the feature 119878(119909 119910) in defect enhancement is demonstratedby an example on radiographic weld image in Figure 3 Asshown the weld radiography image has a low level contrastand a background affected by uneven illumination defectsand background are not easily distinguishable thereforedefect extraction is very difficult Figure 3(b) shows the off-center saliency map 119878(119909 119910) as we can see the differencebetween the defects and the background is highly enhancedthe surface plot of 119878(119909 119910) shown in Figure 3(c) shows alsohow defects gray level are well amplified

22 Level Set Active Contour Energy Formulation The off-center saliency map 119878(119909 119910) obtained in the previous sectionis used as statistical information representing pixels in theimage The feature image 119878(119909 119910) is embedded in the formu-lation of a level set active contour using a global Gaussiandistribution fitting energy The aim is to segment image intotwo regions defect and background Ω

1 Ω2 with a contour

ldquo119862rdquo separating these two regions depending on statisticalinformation inside and outside of the off-center saliencymap 119878(119909 119910) The segmentation is achieved with energy min-imization [19] We define an energy function to the contourldquo119862rdquo composed of three terms the salient energy 119864

1(119862) the

regularization energy 1198642(119862) and 119864

3(119862) the reinitialization

energy

119864 (119862) = 1198641(119862) + 119864

2(119862) + 119864

3(119862) (5)

4 Advances in Materials Science and Engineering

(a) (b)

300

200

00

50

100

150

200

1

08

06

04

02

0

100

(c)

Figure 3 Example of off-center saliency map (a) ROI radiography weld image (b) off-center saliency map 119878(119909 119910) and (c) surface plot ofoff-center saliency map

The first term 1198641(119862) attracts the contour ldquo119862rdquo to defect

edges We choose to minimize the Bayes error [24] definedas follows

1198641(119878 (119909 119910) 119862) = minusint

Ω1

log (1198751(119878 (119909 119910) Ω

1)) 119889119909 119889119910

minus intΩ2

log (1198752(119878 (119909 119910) Ω

2)) 119889119909 119889119910

(6)

where Ω1 Ω2are the regions inside and outside the contour

119862 1198751 1198752are two probability density functions (PDF) 119878(119909 119910)

is the off-center saliency mapBased on the level set theory [16] the contour 119862 is

embedded as the zero level of a level set function 120601(119909 119910)

119862 = (119909 119910) isin Ω 120601 (119909 119910) = 0 (7)

Thus (6) becomes

1198641(119878 (119909 119910) 120601 (119909 119910))

= minusintΩ

log (1198751(119878 (119909 119910))119867 (120601 (119909 119910))) 119889119909 119889119910

minus intΩ

log (1198752(119878 (119909 119910)) (1 minus 119867 (120601 (119909 119910)))) 119889119909 119889119910

(8)

where Ω is the image domain and 119867(120601) is the Heavisidefunction

1198751 1198752are two probability density functions assumed as a

Gaussian distribution with means 1198981 1198982and variances 120590

1

1205902 Consider

119875119894(119878 (119909 119910)) =

1

radic2120587120590119894

exp(minus(119878 (119909 119910) minus 119898

119894)2

21205902

119894

)

119894 = 1 2

(9)

Therefore (8) can be written as

1198641(120601 (119909 119910) 119898

1 1198982 1205901 1205902)

= intΩ

(log (radic2120587) + log (1205901) +

(119878 (119909 119910) minus 1198981)2

21205902

1

)

sdot 119867 (120601 (119909 119910)) 119889119909 119889119910

+ intΩ

(log (radic2120587) + log (1205902) +

(119878 (119909 119910) minus 1198982)2

21205902

2

)

sdot (1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

(10)

Advances in Materials Science and Engineering 5

Means11989811198982and variances1205902

112059022of 119878(119909 119910) respectively

inside and outside the contour ldquo119862rdquo can be calculated asfollows

1198981=

intΩ

119878 (119909 119910) sdot 119867 (120601 (119909 119910)) 119889119909 119889119910

intΩ

119867(120601 (119909 119910)) 119889119909 1198891199101205902

1

=

intΩ

(119878 (119909 119910) minus 1198981)2

sdot 119867 (120601 (119909 119910)) 119889119909 119889119910

intΩ

119867(120601 (119909 119910)) 119889119909 119889119910

1198982=

intΩ

119878 (119909 119910) sdot (1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

intΩ

(1 minus 119867 (120601 (119909 119910))) 119889119909 1198891199101205902

2

=

intΩ

(119878 (119909 119910) minus 1198982)2

sdot (1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

intΩ

(1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

(11)

1198642(120601) is a regularization energy added to keep a smooth

contour of the segmented region so we should minimize thelength of the contourwhich can be calculated by the followingrelation

1198642(120601 (119909 119910)) = int

Ω

1003816100381610038161003816nabla119867 (120601 (119909 119910))1003816100381610038161003816 119889119909 119889119910 (12)

When evolving level set function usually it may be toosteep or flat so we add a third energy term119864

3(120601) to reinitialize

the level set function and to maintain the signed distanceproperty We use a distance regularized level set evolution(DRLSE) proposed by Wang et al [24]

1198643(120601 (119909 119910)) = int

Ω

119875 (1003816100381610038161003816nabla120601 (119909 119910)

1003816100381610038161003816) 119889119909 119889119910 (13)

where 119875 is called a double well potential function defined by

119875 (119911) =

1

21205872(1 minus cos (2120587119911)) if 119911 lt 1

1

2(119911 minus 1) if 119911 gt 1

(14)

As a result the total energy function 119864(120601) can be writtenas

119864 (1206011198981 1198982 1205901 1205902) = 120572119864

1(1206011198981 1198982 1205901 1205902)

+ 1205731198642(120601) + 120574119864

3(120601)

(15)

We have added 120572 120573 120574 as controlling parameters We notethat the parameter 120572 controls the signed distance propertyof the level set function and the parameter 120573 governs itssmoothness The parameter 120574 permits attracting the level setfunction to defect regions

Now we keep all the variables fixed except for 120601 min-imization of the total energy 119864(120601) with respect to 120601 isequivalent to solving the gradient descent flow equation

120597120601

120597119905= minus

120597119864 (120601)

120597120601 (16)

We obtain the following evolution equation

120597120601

120597119905= 120572 div (119889119901 (1003816100381610038161003816nabla120601

1003816100381610038161003816) nabla120601) 120573120575 (120601) div(nabla120601

1003816100381610038161003816nabla1206011003816100381610038161003816

)

+ 120574120575 (120601) (1198651minus 1198652)

(17)

120575(120601) is the Dirac function 1198651 1198652 119889119901 are defined as

follows

119865119894= log (120590

119894) +

(119878 (119909 119910) minus 119898119894)2

21205902

119894

119894 = 1 2

119889119901 (119911) =1198751015840

(119911)

119911

(18)

The level set evolution equation (17) is composed of 3terms the first term in the right side allows keeping theregularity of the level set function the second term preservesa smooth contour and the third term segments the image intodefect and background

Using a simple finite difference explicit scheme to (17) weobtain the discrete level set formulation as follows

120601119899+1

= 120601119899

+ Δ119905 [120572 div (119889119901 (1003816100381610038161003816nabla1206011198991003816100381610038161003816) nabla120601

119899

)

+ 120573120575 (120601119899

) div(nabla120601119899

1003816100381610038161003816nabla1206011198991003816100381610038161003816

) minus 120574120575 (120601119899

) (1198651minus 1198652)]

(19)

where 119899 Δ119905 are the index iteration number and time steprespectively

The main stages of the proposed scheme for segmentingdefects can be summarized as follows

Step 1 Input image

Step 2 Select a region of interest (ROI)

Step 3 Initialize parameters Δ119905 119899 120572 120573 120574

Step 4 Compute off-center saliency map 119878(119909 119910)

Step 5 Compute the initial level set mask 1206010

Step 6

For a fixed number of iterations 119899Compute average and variance119898

11198982 1205901 1205902

Update level set function 120601 with (19)

Step 7 Keep the zero level of the function 120601

3 Experimental Results

In this section the proposed level set active contour model istested and evaluated on a set of radiography images of weldjoints obtained from Federal institute of Material Researchand Testing (Bam) [25] The radiograph films have beenscanned with the scanner LS85 SDR from Lumisys mostly

6 Advances in Materials Science and Engineering

ROI

(a)

RD 0

1

(b)

RD 0

1

(c)

RD 0

1

(d)

Figure 4 Weld defect detection with the proposed active contour (a) the whole weld radiogram (b) ROI selected image (c) off-centersaliency map and (d) final segmentation with proposed level set active contour (blue color)

ROI

(a)

RD 0

2

(b)

RD 0

2

(c)

RD 0

2

(d)

Figure 5 Weld defect detection with the proposed active contour (a) the whole weld radiogram (b) ROI selected image (c) off-centersaliency map and (d) final segmentation with proposed level set active contour (red color)

Advances in Materials Science and Engineering 7RD

03

RD 0

4RD

05

(a)

RD 0

3RD

04

RD 0

5

(b)RD

03

RD 0

4RD

05

(c)

Figure 6 Comparative defect detection tests (a) ROI selected image (b) result of LIF active contour and (c) result of the proposed method

in high density mode The original 12-bit data depth wasrescaled to 8 bits with a linear LUT proportional to opticalfilm density by visual adjustment to the image content Thepixel size is 403 micron (630 dpi) and the images are 8-bit gray values Due to the big size of radiography weldimages the nonuniform illumination and the small defectshape it is difficult to detect the presence of small defectsand determine accurately their sizes during the radiogramvisualization Consequently for the seeking of simplifyingthe task one could begin by selecting the region of interestROI which can be considered as the parts of the imagewhere the radiograph interpreters suspect the presence ofimperfectionsThe selection of the ROI prevents the operatorto make treatments on the irrelevant regions of the imageMoreover it allows reducing the computing time for real-time applications noting that the technique of ROI local-ization is commonly used by researchers in several worksAfter the selection of an ROI region the obtained images areresized to 250 times 350

We present in Figures 4 and 5 the main steps of theproposed segmentation algorithm on two weld radiographyexamples First row corresponds to thewhole weld radiogramthen a region of interest (ROI) is selected Next images inFigures 4(b) and 5(b) show the selected region to processincluding various defects like porosities and slags distributedat the weld middle The off-center saliency map is shownin the next Figures 4(c) and 5(c) as we can see defects are

enhanced and the background is suppressed Figures 4(d)and 5(d) depict the defect contour detection with a blue andred colour obtained with the proposed off-center saliencylevel set active contour Even though the low contrast andthe nonuniform illumination we can see that the most partsof defects are segmented with low false detections We notethat parameters of the proposed method are set as Δ119905 = 01120572 = 004 120573 = 0001 times 255 times 255 120574 = 15 and 119899 = 30

31 Comparative Test 1 We show a comparison between theproposed model and a level set active contour dependingonly on gray level image intensity proposed by Zhang et al[17] and called the local image fitting active contour LIFLIF model Matlab implementation is given in authorrsquos web-site (httpkaihuazhangnetJ papersPR 10rar) Figure 6(b)presents results of the local fittingmodel LIF and the results ofthe proposed active contour model are shown in Figure 6(c)Parameters of the proposed algorithm are fixed as Δ119905 = 01120572 = 002 120573 = 0001 times 255 times 255 120574 = 10 and 119899 = 30

Through a visual evaluation the superiority of the pro-posed level set active contour can be verified Local fittingactive contour LIF [17] fails to obtain satisfactory resultsdue to the influence of the clutter in the background andthe low contrast of defects Much false detection and a lotof defects are not detected The best results are obtainedwith the proposed level set active contour The most parts ofdefects are identified with low false detections A quantitative

8 Advances in Materials Science and EngineeringRD

06

RD 0

7RD

08

RD 0

9

(a)

RD 0

6RD

07

RD 0

8RD

09

(b)

RD 0

6RD

07

RD 0

8RD

09

(c)

Figure 7 Comparative defect segmentation tests (a) ROI selected image (b) results of Songmethod and (c) results of the proposedmethod

comparison between the proposed active contour and LIFmethod is shown in the Table 1 We compute a segmentationevaluation measure used by many researchers called the 119865-measure [26] To compute this evaluation criterion we needideal image segmentation and the proposed segmentationresults the two results should be given in binary masks Notethat ground truth or ideal segmentations used are obtained

from expert visual interpretation We note also that a highervalue (max 1) denotes accurate segmentation The detectionrates are shown in Table 1 which demonstrates the highperformance of our method comparing with LIF approachMoreover the computation time (cpu-time) is computed fortwo methods The step time algorithm is fixed to 01 forthe two methods Algorithms are implemented on Matlab

Advances in Materials Science and Engineering 9

Table 1 Evaluation of segmentation accuracy and computationtime

119865-measure Cpu-Time (sec)LIF Prop method LIF Prop method

RD02 052 076 354 41RD 03 042 081 254 38RD04 062 087 186 35

Table 2 Evaluation of cpu-time convergence

Cpu-time (sec)Song Prop method

RD06 826 385RD07 941 481RD 08 967 523RD09 966 477

R2010a (on a PCCore i5 29GHz) Obtained values show thatour proposed model is very fast compared with local activecontour LIF

32 Comparative Test 2 We show in this paragraph a com-parative test between the proposed method and saliencyconvex active contour model proposed by Song and Yan[18] known as SCACM Images in Figure 7(a) correspondsto regions of interest (ROI) selected from different weldradiogram films containing various kinds of defects likeporosities and horizontal and vertical thin cracks Figure 7(b)presents SCACM method detection results and Figure 7(c)shows the segmentation results with the proposed method

Obtained results in Figure 7 show that Song methoddetects only defects with small and round shapes due to theinfluence of the local window size used in this method In theother side our method does not need to use a local windowas is explained before we use a global Gaussian distributionapproximation of the off-center saliency intensities Theproposed method detects defects with various forms andsizes with low false detections Most parts of defects areextracted despite their low contrasts Using Song methodmany low contrasted defects aremissed and not identified Asin previous experiment the convergence time of twomethodsevaluated in Table 2 shows that proposed model achievesdefect segmentation in a fast time comparing with Songmethod We choose parameters of the proposed algorithm as120572 = 001 120573 = 0001 times 255 times 255 120574 = 13 and 119899 = 30 We notealso that we have used the Matlab implementation given bySCACM authors in their website (httpfacultyneueducnyunhyanWebpage20for20articleSCACMDemoSCACMrar)

4 Conclusion

Detection of small defects in low contrasted radiographyimages corrupted with uneven illumination is very compli-cated Usually proposed methods in this area have limitedresults In this work our aim is to improve robustness of

weld defect segmentation in radiography images in order toobtain satisfactory results We have developed a new methodbased on level set active contour guided with an off-centersaliency map The segmentation is achieved when an energyfunction is minimized Different tests on weld radiographyimages with various kinds of defects prove efficiency androbustness of the proposed method A comparison test isshown with two kinds of level set active contour models Thefirst depends only on gray level intensity and the second usesa saliencymap to derive the segmentation active contourThecomparison tests reveal that the proposed method permitsovercoming the problem of nonuniform illumination andthe low contrast level in radiography weld images Moreoverour method shows a fast convergence time comparing withother methods However some weakness of the proposedmethod must be studied in a future work To reduce humaninteraction it is recommended to develop a method to selectautomatically the region of the interest It is also interestingto validate this segmentation method to another radiographyimage application like medical area

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] I Valavanis and D Kosmopoulos ldquoMulticlass defect detectionand classification in weld radiographic images using geometricand texture featuresrdquo Expert Systems with Applications vol 37no 12 pp 7606ndash7614 2010

[2] J Zapata R Vilar and R Ruiz ldquoPerformance evaluation of anautomatic inspection system of weld defects in radiographicimages based on neuro-classifiersrdquo Expert Systems with Appli-cations vol 38 no 7 pp 8812ndash8824 2011

[3] C Stolojescu-Crisan and S Holban ldquoA comparison of X-rayimage segmentation techniquesrdquo Advances in Electrical andComputer Engineering vol 13 no 3 pp 85ndash92 2013

[4] A Mahmoudi and F Regragui ldquoWelding defect detection bysegmentation of radiographic imagesrdquo in Proceedings of theWorld Congress on Computer Science and Information Engineer-ing (WRI rsquo09) vol 7 pp 111ndash115 Los Angeles Calif USAMarch2009

[5] D Mery and M A Berti ldquoAutomatic detection of weldingdefects using texture featuresrdquo Insight Non-Destructive Testingand Condition Monitoring vol 45 no 10 pp 676ndash681 2003

[6] M A Carrasco and D Mery ldquoSegmentation of welding defectsusing a robust algorithmrdquo Materials Evaluation vol 62 no 11pp 1142ndash1147 2004

[7] E S Amin ldquoApplication of artificial neural networks to evaluateweld defects of nuclear componentsrdquo Journal of Nuclear andRadiation Physics vol 3 no 2 pp 83ndash92 2008

[8] S-B Zhou A-Q Shen and G-F Li ldquoConcrete image segmen-tation based on multiscale mathematic morphology operatorsand Otsu methodrdquo Advances in Materials Science and Engineer-ing vol 2015 Article ID 208473 11 pages 2015

[9] S Osher and N Paragios Geometric Level Set Methods inImaging Vision and Graphics Springer 2003

10 Advances in Materials Science and Engineering

[10] X-F Wang D-S Huang and H Xu ldquoAn efficient local Chan-Vese model for image segmentationrdquo Pattern Recognition vol43 no 3 pp 603ndash618 2010

[11] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998

[12] T F Chan and L A Vese ldquoActive contours without edgesrdquo IEEETransactions on Image Processing vol 10 no 2 pp 266ndash2772001

[13] DMumford and J Shah ldquoOptimal approximations by piecewisesmooth functions and associated variational problemsrdquo Com-munications on Pure and Applied Mathematics vol 42 no 5pp 577ndash685 1989

[14] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009

[15] X Liu S-J Peng Y-MCheung Y Y Tang and J-XDu ldquoActivecontours with a joint and region-scalable distributionmetric forinteractive natural image segmentationrdquo IET Image Processingvol 8 no 12 pp 824ndash832 2014

[16] K Zhang L Zhang H Song and W Zhou ldquoActive contourswith selective local or global segmentation a new formulationand level set methodrdquo Image and Vision Computing vol 28 no4 pp 668ndash676 2010

[17] K Zhang H Song and L Zhang ldquoActive contours driven bylocal image fitting energyrdquo Pattern Recognition vol 43 no 4pp 1199ndash1206 2010

[18] K Song and Y Yan ldquoMicro surface defect detection methodfor silicon steel strip based on saliency convex active contourmodelrdquoMathematical Problems in Engineering vol 2013 ArticleID 429094 13 pages 2013

[19] L Itti and C Koch ldquoComputational modelling of visual atten-tionrdquo Nature Reviews Neuroscience vol 2 no 3 pp 194ndash2032001

[20] R Achantay S Hemamiz F Estraday and S SusstrunkyldquoFrequency-tuned salient region detectionrdquo in Proceedings ofthe IEEE Computer Society Conference on Computer Vision andPatternRecognition (CVPR rsquo09) pp 1597ndash1604 IEEE June 2009

[21] X Hou and L Zhang ldquoSaliency detection a spectral residualapproachrdquo in Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition (CVPR rsquo07) pp 1ndash8 Minneapo-lis Minn USA June 2007

[22] S Montabone and A Soto ldquoHuman detection using a mobileplatform and novel features derived from a visual saliencymechanismrdquo Image and Vision Computing vol 28 no 3 pp391ndash402 2010

[23] P Viola and M J Jones ldquoRobust real-time face detectionrdquoInternational Journal of Computer Vision vol 57 no 2 pp 137ndash154 2004

[24] L Wang L He A Mishra and C Li ldquoActive contours drivenby local Gaussian distribution fitting energyrdquo Signal Processingvol 89 no 12 pp 2435ndash2447 2009

[25] httpwwwbamdeenindexhtm[26] httpsenwikipediaorgwikiPrecision and recall

Submit your manuscripts athttpwwwhindawicom

ScientificaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CorrosionInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Polymer ScienceInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CeramicsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CompositesJournal of

NanoparticlesJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Biomaterials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

NanoscienceJournal of

TextilesHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Journal of

NanotechnologyHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

CrystallographyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CoatingsJournal of

Advances in

Materials Science and EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Smart Materials Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MetallurgyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

MaterialsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Nano

materials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal ofNanomaterials

Advances in Materials Science and Engineering 3

segmentation accuracy depends greatly to the level set localwindow sizes

In this work we propose a new global level set activecontour to segment weld defects in radiography images Animage feature called the off-center saliency map computedwith integral images is embedded in the energy formulationof a global region level set active contour The remainderof this paper is organized as follows the proposed methodis presented in Section 2 tests and experimental results aredepicted in Section 3 and we finish the paper by a conclusionin Section 4

2 Proposed Level Set Active Contour Model

In this section we present the modified level set activecontour We show the off-center saliency map extrac-tion method in the first part and the new level setactive contour energy formulation after fusion with theextracted saliency map feature is illustrated in the secondpart

21 Off-Center Saliency Map Computation Systems thatmimic the biological attention system arewidely developed toextract the saliency map which aims to select the interestingparts of the sensory input data in order to reduce thevast amount of information that a computer vision systemnormally needs to process In this area two types of retinalganglion cells are defined ldquoon-centerrdquo and ldquooff-centerrdquo Anon-center cell is stimulated when the center of its receptivefield is exposed to light and is inhibited when the surroundis exposed to light off-center cells have just the opposite

reaction Two image features that imitate the behavior of thetwo cells are computed and known as on-center and off-center saliency maps In weld radiography images defectscorrespond to dark regions surrounded by brighten back-ground therefore we decide to use the off-center saliencymap as an image feature to guide the level set active contourto defect boundaries Off-center saliency map is calculatedusing center-surround difference Many approaches are usedto achieve computation we find Gabor filters [19] differenceof Gaussian filters [20] and spectral residual [21] RecentlyMontabone and Soto [22] present a fast method to computecenter-surround differences with rectangular filters basedon the concept of integral images (or summed area tables)introduced first by Viola and Jones [23] this approach allowsspeeding up the calculations considerably and preservingobjects borders This approach is used in our work The off-center saliency submaps are calculated by a difference centerand surround with this relation

119878119888(119909 119910) = max surround (119909 119910 119888) minus center (119909 119910) 0 (1)

where ldquo119888rdquo represents the surround size chosen empirically forour application as 119888 = 2 4 8 16 and center(119909 119910) representsthe gray level pixel at the surround middle 119868(119909 119910)

The surround is calculated in a fast time using integralimage 119868

119892(119909 119910) of the image 119868(119909 119910)

119868119892(119909 119910) =

119909

sum

119894=0

119910

sum

119895=0

119868 (119894 119895) (2)

The surround is the local average in neighbourhoodwindow size ldquo119888rdquo

surround (119909 119910 119888) =119868119892(119909 + 119888 119910 + 119888) minus 119868

119892(119909 minus 119888 119910 + 119888) minus 119868

119892(119909 + 119888 119910 minus 119888) + 119868

119892(119909 minus 119888 119910 minus 119888) minus 119868 (119909 119910)

(2119888 + 1)2

minus 1 (3)

The computed off-center saliency submaps are summedpixel by pixel in a one map 119878(119909 119910) as follows

119878 (119909 119910) = sum

119888

119878119888(119909 119910) (4)

119878(119909 119910) is the off-center saliencymapwhichwill be used ina further step with the level set active contourThe advantageof the feature 119878(119909 119910) in defect enhancement is demonstratedby an example on radiographic weld image in Figure 3 Asshown the weld radiography image has a low level contrastand a background affected by uneven illumination defectsand background are not easily distinguishable thereforedefect extraction is very difficult Figure 3(b) shows the off-center saliency map 119878(119909 119910) as we can see the differencebetween the defects and the background is highly enhancedthe surface plot of 119878(119909 119910) shown in Figure 3(c) shows alsohow defects gray level are well amplified

22 Level Set Active Contour Energy Formulation The off-center saliency map 119878(119909 119910) obtained in the previous sectionis used as statistical information representing pixels in theimage The feature image 119878(119909 119910) is embedded in the formu-lation of a level set active contour using a global Gaussiandistribution fitting energy The aim is to segment image intotwo regions defect and background Ω

1 Ω2 with a contour

ldquo119862rdquo separating these two regions depending on statisticalinformation inside and outside of the off-center saliencymap 119878(119909 119910) The segmentation is achieved with energy min-imization [19] We define an energy function to the contourldquo119862rdquo composed of three terms the salient energy 119864

1(119862) the

regularization energy 1198642(119862) and 119864

3(119862) the reinitialization

energy

119864 (119862) = 1198641(119862) + 119864

2(119862) + 119864

3(119862) (5)

4 Advances in Materials Science and Engineering

(a) (b)

300

200

00

50

100

150

200

1

08

06

04

02

0

100

(c)

Figure 3 Example of off-center saliency map (a) ROI radiography weld image (b) off-center saliency map 119878(119909 119910) and (c) surface plot ofoff-center saliency map

The first term 1198641(119862) attracts the contour ldquo119862rdquo to defect

edges We choose to minimize the Bayes error [24] definedas follows

1198641(119878 (119909 119910) 119862) = minusint

Ω1

log (1198751(119878 (119909 119910) Ω

1)) 119889119909 119889119910

minus intΩ2

log (1198752(119878 (119909 119910) Ω

2)) 119889119909 119889119910

(6)

where Ω1 Ω2are the regions inside and outside the contour

119862 1198751 1198752are two probability density functions (PDF) 119878(119909 119910)

is the off-center saliency mapBased on the level set theory [16] the contour 119862 is

embedded as the zero level of a level set function 120601(119909 119910)

119862 = (119909 119910) isin Ω 120601 (119909 119910) = 0 (7)

Thus (6) becomes

1198641(119878 (119909 119910) 120601 (119909 119910))

= minusintΩ

log (1198751(119878 (119909 119910))119867 (120601 (119909 119910))) 119889119909 119889119910

minus intΩ

log (1198752(119878 (119909 119910)) (1 minus 119867 (120601 (119909 119910)))) 119889119909 119889119910

(8)

where Ω is the image domain and 119867(120601) is the Heavisidefunction

1198751 1198752are two probability density functions assumed as a

Gaussian distribution with means 1198981 1198982and variances 120590

1

1205902 Consider

119875119894(119878 (119909 119910)) =

1

radic2120587120590119894

exp(minus(119878 (119909 119910) minus 119898

119894)2

21205902

119894

)

119894 = 1 2

(9)

Therefore (8) can be written as

1198641(120601 (119909 119910) 119898

1 1198982 1205901 1205902)

= intΩ

(log (radic2120587) + log (1205901) +

(119878 (119909 119910) minus 1198981)2

21205902

1

)

sdot 119867 (120601 (119909 119910)) 119889119909 119889119910

+ intΩ

(log (radic2120587) + log (1205902) +

(119878 (119909 119910) minus 1198982)2

21205902

2

)

sdot (1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

(10)

Advances in Materials Science and Engineering 5

Means11989811198982and variances1205902

112059022of 119878(119909 119910) respectively

inside and outside the contour ldquo119862rdquo can be calculated asfollows

1198981=

intΩ

119878 (119909 119910) sdot 119867 (120601 (119909 119910)) 119889119909 119889119910

intΩ

119867(120601 (119909 119910)) 119889119909 1198891199101205902

1

=

intΩ

(119878 (119909 119910) minus 1198981)2

sdot 119867 (120601 (119909 119910)) 119889119909 119889119910

intΩ

119867(120601 (119909 119910)) 119889119909 119889119910

1198982=

intΩ

119878 (119909 119910) sdot (1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

intΩ

(1 minus 119867 (120601 (119909 119910))) 119889119909 1198891199101205902

2

=

intΩ

(119878 (119909 119910) minus 1198982)2

sdot (1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

intΩ

(1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

(11)

1198642(120601) is a regularization energy added to keep a smooth

contour of the segmented region so we should minimize thelength of the contourwhich can be calculated by the followingrelation

1198642(120601 (119909 119910)) = int

Ω

1003816100381610038161003816nabla119867 (120601 (119909 119910))1003816100381610038161003816 119889119909 119889119910 (12)

When evolving level set function usually it may be toosteep or flat so we add a third energy term119864

3(120601) to reinitialize

the level set function and to maintain the signed distanceproperty We use a distance regularized level set evolution(DRLSE) proposed by Wang et al [24]

1198643(120601 (119909 119910)) = int

Ω

119875 (1003816100381610038161003816nabla120601 (119909 119910)

1003816100381610038161003816) 119889119909 119889119910 (13)

where 119875 is called a double well potential function defined by

119875 (119911) =

1

21205872(1 minus cos (2120587119911)) if 119911 lt 1

1

2(119911 minus 1) if 119911 gt 1

(14)

As a result the total energy function 119864(120601) can be writtenas

119864 (1206011198981 1198982 1205901 1205902) = 120572119864

1(1206011198981 1198982 1205901 1205902)

+ 1205731198642(120601) + 120574119864

3(120601)

(15)

We have added 120572 120573 120574 as controlling parameters We notethat the parameter 120572 controls the signed distance propertyof the level set function and the parameter 120573 governs itssmoothness The parameter 120574 permits attracting the level setfunction to defect regions

Now we keep all the variables fixed except for 120601 min-imization of the total energy 119864(120601) with respect to 120601 isequivalent to solving the gradient descent flow equation

120597120601

120597119905= minus

120597119864 (120601)

120597120601 (16)

We obtain the following evolution equation

120597120601

120597119905= 120572 div (119889119901 (1003816100381610038161003816nabla120601

1003816100381610038161003816) nabla120601) 120573120575 (120601) div(nabla120601

1003816100381610038161003816nabla1206011003816100381610038161003816

)

+ 120574120575 (120601) (1198651minus 1198652)

(17)

120575(120601) is the Dirac function 1198651 1198652 119889119901 are defined as

follows

119865119894= log (120590

119894) +

(119878 (119909 119910) minus 119898119894)2

21205902

119894

119894 = 1 2

119889119901 (119911) =1198751015840

(119911)

119911

(18)

The level set evolution equation (17) is composed of 3terms the first term in the right side allows keeping theregularity of the level set function the second term preservesa smooth contour and the third term segments the image intodefect and background

Using a simple finite difference explicit scheme to (17) weobtain the discrete level set formulation as follows

120601119899+1

= 120601119899

+ Δ119905 [120572 div (119889119901 (1003816100381610038161003816nabla1206011198991003816100381610038161003816) nabla120601

119899

)

+ 120573120575 (120601119899

) div(nabla120601119899

1003816100381610038161003816nabla1206011198991003816100381610038161003816

) minus 120574120575 (120601119899

) (1198651minus 1198652)]

(19)

where 119899 Δ119905 are the index iteration number and time steprespectively

The main stages of the proposed scheme for segmentingdefects can be summarized as follows

Step 1 Input image

Step 2 Select a region of interest (ROI)

Step 3 Initialize parameters Δ119905 119899 120572 120573 120574

Step 4 Compute off-center saliency map 119878(119909 119910)

Step 5 Compute the initial level set mask 1206010

Step 6

For a fixed number of iterations 119899Compute average and variance119898

11198982 1205901 1205902

Update level set function 120601 with (19)

Step 7 Keep the zero level of the function 120601

3 Experimental Results

In this section the proposed level set active contour model istested and evaluated on a set of radiography images of weldjoints obtained from Federal institute of Material Researchand Testing (Bam) [25] The radiograph films have beenscanned with the scanner LS85 SDR from Lumisys mostly

6 Advances in Materials Science and Engineering

ROI

(a)

RD 0

1

(b)

RD 0

1

(c)

RD 0

1

(d)

Figure 4 Weld defect detection with the proposed active contour (a) the whole weld radiogram (b) ROI selected image (c) off-centersaliency map and (d) final segmentation with proposed level set active contour (blue color)

ROI

(a)

RD 0

2

(b)

RD 0

2

(c)

RD 0

2

(d)

Figure 5 Weld defect detection with the proposed active contour (a) the whole weld radiogram (b) ROI selected image (c) off-centersaliency map and (d) final segmentation with proposed level set active contour (red color)

Advances in Materials Science and Engineering 7RD

03

RD 0

4RD

05

(a)

RD 0

3RD

04

RD 0

5

(b)RD

03

RD 0

4RD

05

(c)

Figure 6 Comparative defect detection tests (a) ROI selected image (b) result of LIF active contour and (c) result of the proposed method

in high density mode The original 12-bit data depth wasrescaled to 8 bits with a linear LUT proportional to opticalfilm density by visual adjustment to the image content Thepixel size is 403 micron (630 dpi) and the images are 8-bit gray values Due to the big size of radiography weldimages the nonuniform illumination and the small defectshape it is difficult to detect the presence of small defectsand determine accurately their sizes during the radiogramvisualization Consequently for the seeking of simplifyingthe task one could begin by selecting the region of interestROI which can be considered as the parts of the imagewhere the radiograph interpreters suspect the presence ofimperfectionsThe selection of the ROI prevents the operatorto make treatments on the irrelevant regions of the imageMoreover it allows reducing the computing time for real-time applications noting that the technique of ROI local-ization is commonly used by researchers in several worksAfter the selection of an ROI region the obtained images areresized to 250 times 350

We present in Figures 4 and 5 the main steps of theproposed segmentation algorithm on two weld radiographyexamples First row corresponds to thewhole weld radiogramthen a region of interest (ROI) is selected Next images inFigures 4(b) and 5(b) show the selected region to processincluding various defects like porosities and slags distributedat the weld middle The off-center saliency map is shownin the next Figures 4(c) and 5(c) as we can see defects are

enhanced and the background is suppressed Figures 4(d)and 5(d) depict the defect contour detection with a blue andred colour obtained with the proposed off-center saliencylevel set active contour Even though the low contrast andthe nonuniform illumination we can see that the most partsof defects are segmented with low false detections We notethat parameters of the proposed method are set as Δ119905 = 01120572 = 004 120573 = 0001 times 255 times 255 120574 = 15 and 119899 = 30

31 Comparative Test 1 We show a comparison between theproposed model and a level set active contour dependingonly on gray level image intensity proposed by Zhang et al[17] and called the local image fitting active contour LIFLIF model Matlab implementation is given in authorrsquos web-site (httpkaihuazhangnetJ papersPR 10rar) Figure 6(b)presents results of the local fittingmodel LIF and the results ofthe proposed active contour model are shown in Figure 6(c)Parameters of the proposed algorithm are fixed as Δ119905 = 01120572 = 002 120573 = 0001 times 255 times 255 120574 = 10 and 119899 = 30

Through a visual evaluation the superiority of the pro-posed level set active contour can be verified Local fittingactive contour LIF [17] fails to obtain satisfactory resultsdue to the influence of the clutter in the background andthe low contrast of defects Much false detection and a lotof defects are not detected The best results are obtainedwith the proposed level set active contour The most parts ofdefects are identified with low false detections A quantitative

8 Advances in Materials Science and EngineeringRD

06

RD 0

7RD

08

RD 0

9

(a)

RD 0

6RD

07

RD 0

8RD

09

(b)

RD 0

6RD

07

RD 0

8RD

09

(c)

Figure 7 Comparative defect segmentation tests (a) ROI selected image (b) results of Songmethod and (c) results of the proposedmethod

comparison between the proposed active contour and LIFmethod is shown in the Table 1 We compute a segmentationevaluation measure used by many researchers called the 119865-measure [26] To compute this evaluation criterion we needideal image segmentation and the proposed segmentationresults the two results should be given in binary masks Notethat ground truth or ideal segmentations used are obtained

from expert visual interpretation We note also that a highervalue (max 1) denotes accurate segmentation The detectionrates are shown in Table 1 which demonstrates the highperformance of our method comparing with LIF approachMoreover the computation time (cpu-time) is computed fortwo methods The step time algorithm is fixed to 01 forthe two methods Algorithms are implemented on Matlab

Advances in Materials Science and Engineering 9

Table 1 Evaluation of segmentation accuracy and computationtime

119865-measure Cpu-Time (sec)LIF Prop method LIF Prop method

RD02 052 076 354 41RD 03 042 081 254 38RD04 062 087 186 35

Table 2 Evaluation of cpu-time convergence

Cpu-time (sec)Song Prop method

RD06 826 385RD07 941 481RD 08 967 523RD09 966 477

R2010a (on a PCCore i5 29GHz) Obtained values show thatour proposed model is very fast compared with local activecontour LIF

32 Comparative Test 2 We show in this paragraph a com-parative test between the proposed method and saliencyconvex active contour model proposed by Song and Yan[18] known as SCACM Images in Figure 7(a) correspondsto regions of interest (ROI) selected from different weldradiogram films containing various kinds of defects likeporosities and horizontal and vertical thin cracks Figure 7(b)presents SCACM method detection results and Figure 7(c)shows the segmentation results with the proposed method

Obtained results in Figure 7 show that Song methoddetects only defects with small and round shapes due to theinfluence of the local window size used in this method In theother side our method does not need to use a local windowas is explained before we use a global Gaussian distributionapproximation of the off-center saliency intensities Theproposed method detects defects with various forms andsizes with low false detections Most parts of defects areextracted despite their low contrasts Using Song methodmany low contrasted defects aremissed and not identified Asin previous experiment the convergence time of twomethodsevaluated in Table 2 shows that proposed model achievesdefect segmentation in a fast time comparing with Songmethod We choose parameters of the proposed algorithm as120572 = 001 120573 = 0001 times 255 times 255 120574 = 13 and 119899 = 30 We notealso that we have used the Matlab implementation given bySCACM authors in their website (httpfacultyneueducnyunhyanWebpage20for20articleSCACMDemoSCACMrar)

4 Conclusion

Detection of small defects in low contrasted radiographyimages corrupted with uneven illumination is very compli-cated Usually proposed methods in this area have limitedresults In this work our aim is to improve robustness of

weld defect segmentation in radiography images in order toobtain satisfactory results We have developed a new methodbased on level set active contour guided with an off-centersaliency map The segmentation is achieved when an energyfunction is minimized Different tests on weld radiographyimages with various kinds of defects prove efficiency androbustness of the proposed method A comparison test isshown with two kinds of level set active contour models Thefirst depends only on gray level intensity and the second usesa saliencymap to derive the segmentation active contourThecomparison tests reveal that the proposed method permitsovercoming the problem of nonuniform illumination andthe low contrast level in radiography weld images Moreoverour method shows a fast convergence time comparing withother methods However some weakness of the proposedmethod must be studied in a future work To reduce humaninteraction it is recommended to develop a method to selectautomatically the region of the interest It is also interestingto validate this segmentation method to another radiographyimage application like medical area

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] I Valavanis and D Kosmopoulos ldquoMulticlass defect detectionand classification in weld radiographic images using geometricand texture featuresrdquo Expert Systems with Applications vol 37no 12 pp 7606ndash7614 2010

[2] J Zapata R Vilar and R Ruiz ldquoPerformance evaluation of anautomatic inspection system of weld defects in radiographicimages based on neuro-classifiersrdquo Expert Systems with Appli-cations vol 38 no 7 pp 8812ndash8824 2011

[3] C Stolojescu-Crisan and S Holban ldquoA comparison of X-rayimage segmentation techniquesrdquo Advances in Electrical andComputer Engineering vol 13 no 3 pp 85ndash92 2013

[4] A Mahmoudi and F Regragui ldquoWelding defect detection bysegmentation of radiographic imagesrdquo in Proceedings of theWorld Congress on Computer Science and Information Engineer-ing (WRI rsquo09) vol 7 pp 111ndash115 Los Angeles Calif USAMarch2009

[5] D Mery and M A Berti ldquoAutomatic detection of weldingdefects using texture featuresrdquo Insight Non-Destructive Testingand Condition Monitoring vol 45 no 10 pp 676ndash681 2003

[6] M A Carrasco and D Mery ldquoSegmentation of welding defectsusing a robust algorithmrdquo Materials Evaluation vol 62 no 11pp 1142ndash1147 2004

[7] E S Amin ldquoApplication of artificial neural networks to evaluateweld defects of nuclear componentsrdquo Journal of Nuclear andRadiation Physics vol 3 no 2 pp 83ndash92 2008

[8] S-B Zhou A-Q Shen and G-F Li ldquoConcrete image segmen-tation based on multiscale mathematic morphology operatorsand Otsu methodrdquo Advances in Materials Science and Engineer-ing vol 2015 Article ID 208473 11 pages 2015

[9] S Osher and N Paragios Geometric Level Set Methods inImaging Vision and Graphics Springer 2003

10 Advances in Materials Science and Engineering

[10] X-F Wang D-S Huang and H Xu ldquoAn efficient local Chan-Vese model for image segmentationrdquo Pattern Recognition vol43 no 3 pp 603ndash618 2010

[11] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998

[12] T F Chan and L A Vese ldquoActive contours without edgesrdquo IEEETransactions on Image Processing vol 10 no 2 pp 266ndash2772001

[13] DMumford and J Shah ldquoOptimal approximations by piecewisesmooth functions and associated variational problemsrdquo Com-munications on Pure and Applied Mathematics vol 42 no 5pp 577ndash685 1989

[14] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009

[15] X Liu S-J Peng Y-MCheung Y Y Tang and J-XDu ldquoActivecontours with a joint and region-scalable distributionmetric forinteractive natural image segmentationrdquo IET Image Processingvol 8 no 12 pp 824ndash832 2014

[16] K Zhang L Zhang H Song and W Zhou ldquoActive contourswith selective local or global segmentation a new formulationand level set methodrdquo Image and Vision Computing vol 28 no4 pp 668ndash676 2010

[17] K Zhang H Song and L Zhang ldquoActive contours driven bylocal image fitting energyrdquo Pattern Recognition vol 43 no 4pp 1199ndash1206 2010

[18] K Song and Y Yan ldquoMicro surface defect detection methodfor silicon steel strip based on saliency convex active contourmodelrdquoMathematical Problems in Engineering vol 2013 ArticleID 429094 13 pages 2013

[19] L Itti and C Koch ldquoComputational modelling of visual atten-tionrdquo Nature Reviews Neuroscience vol 2 no 3 pp 194ndash2032001

[20] R Achantay S Hemamiz F Estraday and S SusstrunkyldquoFrequency-tuned salient region detectionrdquo in Proceedings ofthe IEEE Computer Society Conference on Computer Vision andPatternRecognition (CVPR rsquo09) pp 1597ndash1604 IEEE June 2009

[21] X Hou and L Zhang ldquoSaliency detection a spectral residualapproachrdquo in Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition (CVPR rsquo07) pp 1ndash8 Minneapo-lis Minn USA June 2007

[22] S Montabone and A Soto ldquoHuman detection using a mobileplatform and novel features derived from a visual saliencymechanismrdquo Image and Vision Computing vol 28 no 3 pp391ndash402 2010

[23] P Viola and M J Jones ldquoRobust real-time face detectionrdquoInternational Journal of Computer Vision vol 57 no 2 pp 137ndash154 2004

[24] L Wang L He A Mishra and C Li ldquoActive contours drivenby local Gaussian distribution fitting energyrdquo Signal Processingvol 89 no 12 pp 2435ndash2447 2009

[25] httpwwwbamdeenindexhtm[26] httpsenwikipediaorgwikiPrecision and recall

Submit your manuscripts athttpwwwhindawicom

ScientificaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CorrosionInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Polymer ScienceInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CeramicsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CompositesJournal of

NanoparticlesJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Biomaterials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

NanoscienceJournal of

TextilesHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Journal of

NanotechnologyHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

CrystallographyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CoatingsJournal of

Advances in

Materials Science and EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Smart Materials Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MetallurgyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

MaterialsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Nano

materials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal ofNanomaterials

4 Advances in Materials Science and Engineering

(a) (b)

300

200

00

50

100

150

200

1

08

06

04

02

0

100

(c)

Figure 3 Example of off-center saliency map (a) ROI radiography weld image (b) off-center saliency map 119878(119909 119910) and (c) surface plot ofoff-center saliency map

The first term 1198641(119862) attracts the contour ldquo119862rdquo to defect

edges We choose to minimize the Bayes error [24] definedas follows

1198641(119878 (119909 119910) 119862) = minusint

Ω1

log (1198751(119878 (119909 119910) Ω

1)) 119889119909 119889119910

minus intΩ2

log (1198752(119878 (119909 119910) Ω

2)) 119889119909 119889119910

(6)

where Ω1 Ω2are the regions inside and outside the contour

119862 1198751 1198752are two probability density functions (PDF) 119878(119909 119910)

is the off-center saliency mapBased on the level set theory [16] the contour 119862 is

embedded as the zero level of a level set function 120601(119909 119910)

119862 = (119909 119910) isin Ω 120601 (119909 119910) = 0 (7)

Thus (6) becomes

1198641(119878 (119909 119910) 120601 (119909 119910))

= minusintΩ

log (1198751(119878 (119909 119910))119867 (120601 (119909 119910))) 119889119909 119889119910

minus intΩ

log (1198752(119878 (119909 119910)) (1 minus 119867 (120601 (119909 119910)))) 119889119909 119889119910

(8)

where Ω is the image domain and 119867(120601) is the Heavisidefunction

1198751 1198752are two probability density functions assumed as a

Gaussian distribution with means 1198981 1198982and variances 120590

1

1205902 Consider

119875119894(119878 (119909 119910)) =

1

radic2120587120590119894

exp(minus(119878 (119909 119910) minus 119898

119894)2

21205902

119894

)

119894 = 1 2

(9)

Therefore (8) can be written as

1198641(120601 (119909 119910) 119898

1 1198982 1205901 1205902)

= intΩ

(log (radic2120587) + log (1205901) +

(119878 (119909 119910) minus 1198981)2

21205902

1

)

sdot 119867 (120601 (119909 119910)) 119889119909 119889119910

+ intΩ

(log (radic2120587) + log (1205902) +

(119878 (119909 119910) minus 1198982)2

21205902

2

)

sdot (1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

(10)

Advances in Materials Science and Engineering 5

Means11989811198982and variances1205902

112059022of 119878(119909 119910) respectively

inside and outside the contour ldquo119862rdquo can be calculated asfollows

1198981=

intΩ

119878 (119909 119910) sdot 119867 (120601 (119909 119910)) 119889119909 119889119910

intΩ

119867(120601 (119909 119910)) 119889119909 1198891199101205902

1

=

intΩ

(119878 (119909 119910) minus 1198981)2

sdot 119867 (120601 (119909 119910)) 119889119909 119889119910

intΩ

119867(120601 (119909 119910)) 119889119909 119889119910

1198982=

intΩ

119878 (119909 119910) sdot (1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

intΩ

(1 minus 119867 (120601 (119909 119910))) 119889119909 1198891199101205902

2

=

intΩ

(119878 (119909 119910) minus 1198982)2

sdot (1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

intΩ

(1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

(11)

1198642(120601) is a regularization energy added to keep a smooth

contour of the segmented region so we should minimize thelength of the contourwhich can be calculated by the followingrelation

1198642(120601 (119909 119910)) = int

Ω

1003816100381610038161003816nabla119867 (120601 (119909 119910))1003816100381610038161003816 119889119909 119889119910 (12)

When evolving level set function usually it may be toosteep or flat so we add a third energy term119864

3(120601) to reinitialize

the level set function and to maintain the signed distanceproperty We use a distance regularized level set evolution(DRLSE) proposed by Wang et al [24]

1198643(120601 (119909 119910)) = int

Ω

119875 (1003816100381610038161003816nabla120601 (119909 119910)

1003816100381610038161003816) 119889119909 119889119910 (13)

where 119875 is called a double well potential function defined by

119875 (119911) =

1

21205872(1 minus cos (2120587119911)) if 119911 lt 1

1

2(119911 minus 1) if 119911 gt 1

(14)

As a result the total energy function 119864(120601) can be writtenas

119864 (1206011198981 1198982 1205901 1205902) = 120572119864

1(1206011198981 1198982 1205901 1205902)

+ 1205731198642(120601) + 120574119864

3(120601)

(15)

We have added 120572 120573 120574 as controlling parameters We notethat the parameter 120572 controls the signed distance propertyof the level set function and the parameter 120573 governs itssmoothness The parameter 120574 permits attracting the level setfunction to defect regions

Now we keep all the variables fixed except for 120601 min-imization of the total energy 119864(120601) with respect to 120601 isequivalent to solving the gradient descent flow equation

120597120601

120597119905= minus

120597119864 (120601)

120597120601 (16)

We obtain the following evolution equation

120597120601

120597119905= 120572 div (119889119901 (1003816100381610038161003816nabla120601

1003816100381610038161003816) nabla120601) 120573120575 (120601) div(nabla120601

1003816100381610038161003816nabla1206011003816100381610038161003816

)

+ 120574120575 (120601) (1198651minus 1198652)

(17)

120575(120601) is the Dirac function 1198651 1198652 119889119901 are defined as

follows

119865119894= log (120590

119894) +

(119878 (119909 119910) minus 119898119894)2

21205902

119894

119894 = 1 2

119889119901 (119911) =1198751015840

(119911)

119911

(18)

The level set evolution equation (17) is composed of 3terms the first term in the right side allows keeping theregularity of the level set function the second term preservesa smooth contour and the third term segments the image intodefect and background

Using a simple finite difference explicit scheme to (17) weobtain the discrete level set formulation as follows

120601119899+1

= 120601119899

+ Δ119905 [120572 div (119889119901 (1003816100381610038161003816nabla1206011198991003816100381610038161003816) nabla120601

119899

)

+ 120573120575 (120601119899

) div(nabla120601119899

1003816100381610038161003816nabla1206011198991003816100381610038161003816

) minus 120574120575 (120601119899

) (1198651minus 1198652)]

(19)

where 119899 Δ119905 are the index iteration number and time steprespectively

The main stages of the proposed scheme for segmentingdefects can be summarized as follows

Step 1 Input image

Step 2 Select a region of interest (ROI)

Step 3 Initialize parameters Δ119905 119899 120572 120573 120574

Step 4 Compute off-center saliency map 119878(119909 119910)

Step 5 Compute the initial level set mask 1206010

Step 6

For a fixed number of iterations 119899Compute average and variance119898

11198982 1205901 1205902

Update level set function 120601 with (19)

Step 7 Keep the zero level of the function 120601

3 Experimental Results

In this section the proposed level set active contour model istested and evaluated on a set of radiography images of weldjoints obtained from Federal institute of Material Researchand Testing (Bam) [25] The radiograph films have beenscanned with the scanner LS85 SDR from Lumisys mostly

6 Advances in Materials Science and Engineering

ROI

(a)

RD 0

1

(b)

RD 0

1

(c)

RD 0

1

(d)

Figure 4 Weld defect detection with the proposed active contour (a) the whole weld radiogram (b) ROI selected image (c) off-centersaliency map and (d) final segmentation with proposed level set active contour (blue color)

ROI

(a)

RD 0

2

(b)

RD 0

2

(c)

RD 0

2

(d)

Figure 5 Weld defect detection with the proposed active contour (a) the whole weld radiogram (b) ROI selected image (c) off-centersaliency map and (d) final segmentation with proposed level set active contour (red color)

Advances in Materials Science and Engineering 7RD

03

RD 0

4RD

05

(a)

RD 0

3RD

04

RD 0

5

(b)RD

03

RD 0

4RD

05

(c)

Figure 6 Comparative defect detection tests (a) ROI selected image (b) result of LIF active contour and (c) result of the proposed method

in high density mode The original 12-bit data depth wasrescaled to 8 bits with a linear LUT proportional to opticalfilm density by visual adjustment to the image content Thepixel size is 403 micron (630 dpi) and the images are 8-bit gray values Due to the big size of radiography weldimages the nonuniform illumination and the small defectshape it is difficult to detect the presence of small defectsand determine accurately their sizes during the radiogramvisualization Consequently for the seeking of simplifyingthe task one could begin by selecting the region of interestROI which can be considered as the parts of the imagewhere the radiograph interpreters suspect the presence ofimperfectionsThe selection of the ROI prevents the operatorto make treatments on the irrelevant regions of the imageMoreover it allows reducing the computing time for real-time applications noting that the technique of ROI local-ization is commonly used by researchers in several worksAfter the selection of an ROI region the obtained images areresized to 250 times 350

We present in Figures 4 and 5 the main steps of theproposed segmentation algorithm on two weld radiographyexamples First row corresponds to thewhole weld radiogramthen a region of interest (ROI) is selected Next images inFigures 4(b) and 5(b) show the selected region to processincluding various defects like porosities and slags distributedat the weld middle The off-center saliency map is shownin the next Figures 4(c) and 5(c) as we can see defects are

enhanced and the background is suppressed Figures 4(d)and 5(d) depict the defect contour detection with a blue andred colour obtained with the proposed off-center saliencylevel set active contour Even though the low contrast andthe nonuniform illumination we can see that the most partsof defects are segmented with low false detections We notethat parameters of the proposed method are set as Δ119905 = 01120572 = 004 120573 = 0001 times 255 times 255 120574 = 15 and 119899 = 30

31 Comparative Test 1 We show a comparison between theproposed model and a level set active contour dependingonly on gray level image intensity proposed by Zhang et al[17] and called the local image fitting active contour LIFLIF model Matlab implementation is given in authorrsquos web-site (httpkaihuazhangnetJ papersPR 10rar) Figure 6(b)presents results of the local fittingmodel LIF and the results ofthe proposed active contour model are shown in Figure 6(c)Parameters of the proposed algorithm are fixed as Δ119905 = 01120572 = 002 120573 = 0001 times 255 times 255 120574 = 10 and 119899 = 30

Through a visual evaluation the superiority of the pro-posed level set active contour can be verified Local fittingactive contour LIF [17] fails to obtain satisfactory resultsdue to the influence of the clutter in the background andthe low contrast of defects Much false detection and a lotof defects are not detected The best results are obtainedwith the proposed level set active contour The most parts ofdefects are identified with low false detections A quantitative

8 Advances in Materials Science and EngineeringRD

06

RD 0

7RD

08

RD 0

9

(a)

RD 0

6RD

07

RD 0

8RD

09

(b)

RD 0

6RD

07

RD 0

8RD

09

(c)

Figure 7 Comparative defect segmentation tests (a) ROI selected image (b) results of Songmethod and (c) results of the proposedmethod

comparison between the proposed active contour and LIFmethod is shown in the Table 1 We compute a segmentationevaluation measure used by many researchers called the 119865-measure [26] To compute this evaluation criterion we needideal image segmentation and the proposed segmentationresults the two results should be given in binary masks Notethat ground truth or ideal segmentations used are obtained

from expert visual interpretation We note also that a highervalue (max 1) denotes accurate segmentation The detectionrates are shown in Table 1 which demonstrates the highperformance of our method comparing with LIF approachMoreover the computation time (cpu-time) is computed fortwo methods The step time algorithm is fixed to 01 forthe two methods Algorithms are implemented on Matlab

Advances in Materials Science and Engineering 9

Table 1 Evaluation of segmentation accuracy and computationtime

119865-measure Cpu-Time (sec)LIF Prop method LIF Prop method

RD02 052 076 354 41RD 03 042 081 254 38RD04 062 087 186 35

Table 2 Evaluation of cpu-time convergence

Cpu-time (sec)Song Prop method

RD06 826 385RD07 941 481RD 08 967 523RD09 966 477

R2010a (on a PCCore i5 29GHz) Obtained values show thatour proposed model is very fast compared with local activecontour LIF

32 Comparative Test 2 We show in this paragraph a com-parative test between the proposed method and saliencyconvex active contour model proposed by Song and Yan[18] known as SCACM Images in Figure 7(a) correspondsto regions of interest (ROI) selected from different weldradiogram films containing various kinds of defects likeporosities and horizontal and vertical thin cracks Figure 7(b)presents SCACM method detection results and Figure 7(c)shows the segmentation results with the proposed method

Obtained results in Figure 7 show that Song methoddetects only defects with small and round shapes due to theinfluence of the local window size used in this method In theother side our method does not need to use a local windowas is explained before we use a global Gaussian distributionapproximation of the off-center saliency intensities Theproposed method detects defects with various forms andsizes with low false detections Most parts of defects areextracted despite their low contrasts Using Song methodmany low contrasted defects aremissed and not identified Asin previous experiment the convergence time of twomethodsevaluated in Table 2 shows that proposed model achievesdefect segmentation in a fast time comparing with Songmethod We choose parameters of the proposed algorithm as120572 = 001 120573 = 0001 times 255 times 255 120574 = 13 and 119899 = 30 We notealso that we have used the Matlab implementation given bySCACM authors in their website (httpfacultyneueducnyunhyanWebpage20for20articleSCACMDemoSCACMrar)

4 Conclusion

Detection of small defects in low contrasted radiographyimages corrupted with uneven illumination is very compli-cated Usually proposed methods in this area have limitedresults In this work our aim is to improve robustness of

weld defect segmentation in radiography images in order toobtain satisfactory results We have developed a new methodbased on level set active contour guided with an off-centersaliency map The segmentation is achieved when an energyfunction is minimized Different tests on weld radiographyimages with various kinds of defects prove efficiency androbustness of the proposed method A comparison test isshown with two kinds of level set active contour models Thefirst depends only on gray level intensity and the second usesa saliencymap to derive the segmentation active contourThecomparison tests reveal that the proposed method permitsovercoming the problem of nonuniform illumination andthe low contrast level in radiography weld images Moreoverour method shows a fast convergence time comparing withother methods However some weakness of the proposedmethod must be studied in a future work To reduce humaninteraction it is recommended to develop a method to selectautomatically the region of the interest It is also interestingto validate this segmentation method to another radiographyimage application like medical area

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] I Valavanis and D Kosmopoulos ldquoMulticlass defect detectionand classification in weld radiographic images using geometricand texture featuresrdquo Expert Systems with Applications vol 37no 12 pp 7606ndash7614 2010

[2] J Zapata R Vilar and R Ruiz ldquoPerformance evaluation of anautomatic inspection system of weld defects in radiographicimages based on neuro-classifiersrdquo Expert Systems with Appli-cations vol 38 no 7 pp 8812ndash8824 2011

[3] C Stolojescu-Crisan and S Holban ldquoA comparison of X-rayimage segmentation techniquesrdquo Advances in Electrical andComputer Engineering vol 13 no 3 pp 85ndash92 2013

[4] A Mahmoudi and F Regragui ldquoWelding defect detection bysegmentation of radiographic imagesrdquo in Proceedings of theWorld Congress on Computer Science and Information Engineer-ing (WRI rsquo09) vol 7 pp 111ndash115 Los Angeles Calif USAMarch2009

[5] D Mery and M A Berti ldquoAutomatic detection of weldingdefects using texture featuresrdquo Insight Non-Destructive Testingand Condition Monitoring vol 45 no 10 pp 676ndash681 2003

[6] M A Carrasco and D Mery ldquoSegmentation of welding defectsusing a robust algorithmrdquo Materials Evaluation vol 62 no 11pp 1142ndash1147 2004

[7] E S Amin ldquoApplication of artificial neural networks to evaluateweld defects of nuclear componentsrdquo Journal of Nuclear andRadiation Physics vol 3 no 2 pp 83ndash92 2008

[8] S-B Zhou A-Q Shen and G-F Li ldquoConcrete image segmen-tation based on multiscale mathematic morphology operatorsand Otsu methodrdquo Advances in Materials Science and Engineer-ing vol 2015 Article ID 208473 11 pages 2015

[9] S Osher and N Paragios Geometric Level Set Methods inImaging Vision and Graphics Springer 2003

10 Advances in Materials Science and Engineering

[10] X-F Wang D-S Huang and H Xu ldquoAn efficient local Chan-Vese model for image segmentationrdquo Pattern Recognition vol43 no 3 pp 603ndash618 2010

[11] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998

[12] T F Chan and L A Vese ldquoActive contours without edgesrdquo IEEETransactions on Image Processing vol 10 no 2 pp 266ndash2772001

[13] DMumford and J Shah ldquoOptimal approximations by piecewisesmooth functions and associated variational problemsrdquo Com-munications on Pure and Applied Mathematics vol 42 no 5pp 577ndash685 1989

[14] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009

[15] X Liu S-J Peng Y-MCheung Y Y Tang and J-XDu ldquoActivecontours with a joint and region-scalable distributionmetric forinteractive natural image segmentationrdquo IET Image Processingvol 8 no 12 pp 824ndash832 2014

[16] K Zhang L Zhang H Song and W Zhou ldquoActive contourswith selective local or global segmentation a new formulationand level set methodrdquo Image and Vision Computing vol 28 no4 pp 668ndash676 2010

[17] K Zhang H Song and L Zhang ldquoActive contours driven bylocal image fitting energyrdquo Pattern Recognition vol 43 no 4pp 1199ndash1206 2010

[18] K Song and Y Yan ldquoMicro surface defect detection methodfor silicon steel strip based on saliency convex active contourmodelrdquoMathematical Problems in Engineering vol 2013 ArticleID 429094 13 pages 2013

[19] L Itti and C Koch ldquoComputational modelling of visual atten-tionrdquo Nature Reviews Neuroscience vol 2 no 3 pp 194ndash2032001

[20] R Achantay S Hemamiz F Estraday and S SusstrunkyldquoFrequency-tuned salient region detectionrdquo in Proceedings ofthe IEEE Computer Society Conference on Computer Vision andPatternRecognition (CVPR rsquo09) pp 1597ndash1604 IEEE June 2009

[21] X Hou and L Zhang ldquoSaliency detection a spectral residualapproachrdquo in Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition (CVPR rsquo07) pp 1ndash8 Minneapo-lis Minn USA June 2007

[22] S Montabone and A Soto ldquoHuman detection using a mobileplatform and novel features derived from a visual saliencymechanismrdquo Image and Vision Computing vol 28 no 3 pp391ndash402 2010

[23] P Viola and M J Jones ldquoRobust real-time face detectionrdquoInternational Journal of Computer Vision vol 57 no 2 pp 137ndash154 2004

[24] L Wang L He A Mishra and C Li ldquoActive contours drivenby local Gaussian distribution fitting energyrdquo Signal Processingvol 89 no 12 pp 2435ndash2447 2009

[25] httpwwwbamdeenindexhtm[26] httpsenwikipediaorgwikiPrecision and recall

Submit your manuscripts athttpwwwhindawicom

ScientificaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CorrosionInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Polymer ScienceInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CeramicsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CompositesJournal of

NanoparticlesJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Biomaterials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

NanoscienceJournal of

TextilesHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Journal of

NanotechnologyHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

CrystallographyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CoatingsJournal of

Advances in

Materials Science and EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Smart Materials Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MetallurgyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

MaterialsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Nano

materials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal ofNanomaterials

Advances in Materials Science and Engineering 5

Means11989811198982and variances1205902

112059022of 119878(119909 119910) respectively

inside and outside the contour ldquo119862rdquo can be calculated asfollows

1198981=

intΩ

119878 (119909 119910) sdot 119867 (120601 (119909 119910)) 119889119909 119889119910

intΩ

119867(120601 (119909 119910)) 119889119909 1198891199101205902

1

=

intΩ

(119878 (119909 119910) minus 1198981)2

sdot 119867 (120601 (119909 119910)) 119889119909 119889119910

intΩ

119867(120601 (119909 119910)) 119889119909 119889119910

1198982=

intΩ

119878 (119909 119910) sdot (1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

intΩ

(1 minus 119867 (120601 (119909 119910))) 119889119909 1198891199101205902

2

=

intΩ

(119878 (119909 119910) minus 1198982)2

sdot (1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

intΩ

(1 minus 119867 (120601 (119909 119910))) 119889119909 119889119910

(11)

1198642(120601) is a regularization energy added to keep a smooth

contour of the segmented region so we should minimize thelength of the contourwhich can be calculated by the followingrelation

1198642(120601 (119909 119910)) = int

Ω

1003816100381610038161003816nabla119867 (120601 (119909 119910))1003816100381610038161003816 119889119909 119889119910 (12)

When evolving level set function usually it may be toosteep or flat so we add a third energy term119864

3(120601) to reinitialize

the level set function and to maintain the signed distanceproperty We use a distance regularized level set evolution(DRLSE) proposed by Wang et al [24]

1198643(120601 (119909 119910)) = int

Ω

119875 (1003816100381610038161003816nabla120601 (119909 119910)

1003816100381610038161003816) 119889119909 119889119910 (13)

where 119875 is called a double well potential function defined by

119875 (119911) =

1

21205872(1 minus cos (2120587119911)) if 119911 lt 1

1

2(119911 minus 1) if 119911 gt 1

(14)

As a result the total energy function 119864(120601) can be writtenas

119864 (1206011198981 1198982 1205901 1205902) = 120572119864

1(1206011198981 1198982 1205901 1205902)

+ 1205731198642(120601) + 120574119864

3(120601)

(15)

We have added 120572 120573 120574 as controlling parameters We notethat the parameter 120572 controls the signed distance propertyof the level set function and the parameter 120573 governs itssmoothness The parameter 120574 permits attracting the level setfunction to defect regions

Now we keep all the variables fixed except for 120601 min-imization of the total energy 119864(120601) with respect to 120601 isequivalent to solving the gradient descent flow equation

120597120601

120597119905= minus

120597119864 (120601)

120597120601 (16)

We obtain the following evolution equation

120597120601

120597119905= 120572 div (119889119901 (1003816100381610038161003816nabla120601

1003816100381610038161003816) nabla120601) 120573120575 (120601) div(nabla120601

1003816100381610038161003816nabla1206011003816100381610038161003816

)

+ 120574120575 (120601) (1198651minus 1198652)

(17)

120575(120601) is the Dirac function 1198651 1198652 119889119901 are defined as

follows

119865119894= log (120590

119894) +

(119878 (119909 119910) minus 119898119894)2

21205902

119894

119894 = 1 2

119889119901 (119911) =1198751015840

(119911)

119911

(18)

The level set evolution equation (17) is composed of 3terms the first term in the right side allows keeping theregularity of the level set function the second term preservesa smooth contour and the third term segments the image intodefect and background

Using a simple finite difference explicit scheme to (17) weobtain the discrete level set formulation as follows

120601119899+1

= 120601119899

+ Δ119905 [120572 div (119889119901 (1003816100381610038161003816nabla1206011198991003816100381610038161003816) nabla120601

119899

)

+ 120573120575 (120601119899

) div(nabla120601119899

1003816100381610038161003816nabla1206011198991003816100381610038161003816

) minus 120574120575 (120601119899

) (1198651minus 1198652)]

(19)

where 119899 Δ119905 are the index iteration number and time steprespectively

The main stages of the proposed scheme for segmentingdefects can be summarized as follows

Step 1 Input image

Step 2 Select a region of interest (ROI)

Step 3 Initialize parameters Δ119905 119899 120572 120573 120574

Step 4 Compute off-center saliency map 119878(119909 119910)

Step 5 Compute the initial level set mask 1206010

Step 6

For a fixed number of iterations 119899Compute average and variance119898

11198982 1205901 1205902

Update level set function 120601 with (19)

Step 7 Keep the zero level of the function 120601

3 Experimental Results

In this section the proposed level set active contour model istested and evaluated on a set of radiography images of weldjoints obtained from Federal institute of Material Researchand Testing (Bam) [25] The radiograph films have beenscanned with the scanner LS85 SDR from Lumisys mostly

6 Advances in Materials Science and Engineering

ROI

(a)

RD 0

1

(b)

RD 0

1

(c)

RD 0

1

(d)

Figure 4 Weld defect detection with the proposed active contour (a) the whole weld radiogram (b) ROI selected image (c) off-centersaliency map and (d) final segmentation with proposed level set active contour (blue color)

ROI

(a)

RD 0

2

(b)

RD 0

2

(c)

RD 0

2

(d)

Figure 5 Weld defect detection with the proposed active contour (a) the whole weld radiogram (b) ROI selected image (c) off-centersaliency map and (d) final segmentation with proposed level set active contour (red color)

Advances in Materials Science and Engineering 7RD

03

RD 0

4RD

05

(a)

RD 0

3RD

04

RD 0

5

(b)RD

03

RD 0

4RD

05

(c)

Figure 6 Comparative defect detection tests (a) ROI selected image (b) result of LIF active contour and (c) result of the proposed method

in high density mode The original 12-bit data depth wasrescaled to 8 bits with a linear LUT proportional to opticalfilm density by visual adjustment to the image content Thepixel size is 403 micron (630 dpi) and the images are 8-bit gray values Due to the big size of radiography weldimages the nonuniform illumination and the small defectshape it is difficult to detect the presence of small defectsand determine accurately their sizes during the radiogramvisualization Consequently for the seeking of simplifyingthe task one could begin by selecting the region of interestROI which can be considered as the parts of the imagewhere the radiograph interpreters suspect the presence ofimperfectionsThe selection of the ROI prevents the operatorto make treatments on the irrelevant regions of the imageMoreover it allows reducing the computing time for real-time applications noting that the technique of ROI local-ization is commonly used by researchers in several worksAfter the selection of an ROI region the obtained images areresized to 250 times 350

We present in Figures 4 and 5 the main steps of theproposed segmentation algorithm on two weld radiographyexamples First row corresponds to thewhole weld radiogramthen a region of interest (ROI) is selected Next images inFigures 4(b) and 5(b) show the selected region to processincluding various defects like porosities and slags distributedat the weld middle The off-center saliency map is shownin the next Figures 4(c) and 5(c) as we can see defects are

enhanced and the background is suppressed Figures 4(d)and 5(d) depict the defect contour detection with a blue andred colour obtained with the proposed off-center saliencylevel set active contour Even though the low contrast andthe nonuniform illumination we can see that the most partsof defects are segmented with low false detections We notethat parameters of the proposed method are set as Δ119905 = 01120572 = 004 120573 = 0001 times 255 times 255 120574 = 15 and 119899 = 30

31 Comparative Test 1 We show a comparison between theproposed model and a level set active contour dependingonly on gray level image intensity proposed by Zhang et al[17] and called the local image fitting active contour LIFLIF model Matlab implementation is given in authorrsquos web-site (httpkaihuazhangnetJ papersPR 10rar) Figure 6(b)presents results of the local fittingmodel LIF and the results ofthe proposed active contour model are shown in Figure 6(c)Parameters of the proposed algorithm are fixed as Δ119905 = 01120572 = 002 120573 = 0001 times 255 times 255 120574 = 10 and 119899 = 30

Through a visual evaluation the superiority of the pro-posed level set active contour can be verified Local fittingactive contour LIF [17] fails to obtain satisfactory resultsdue to the influence of the clutter in the background andthe low contrast of defects Much false detection and a lotof defects are not detected The best results are obtainedwith the proposed level set active contour The most parts ofdefects are identified with low false detections A quantitative

8 Advances in Materials Science and EngineeringRD

06

RD 0

7RD

08

RD 0

9

(a)

RD 0

6RD

07

RD 0

8RD

09

(b)

RD 0

6RD

07

RD 0

8RD

09

(c)

Figure 7 Comparative defect segmentation tests (a) ROI selected image (b) results of Songmethod and (c) results of the proposedmethod

comparison between the proposed active contour and LIFmethod is shown in the Table 1 We compute a segmentationevaluation measure used by many researchers called the 119865-measure [26] To compute this evaluation criterion we needideal image segmentation and the proposed segmentationresults the two results should be given in binary masks Notethat ground truth or ideal segmentations used are obtained

from expert visual interpretation We note also that a highervalue (max 1) denotes accurate segmentation The detectionrates are shown in Table 1 which demonstrates the highperformance of our method comparing with LIF approachMoreover the computation time (cpu-time) is computed fortwo methods The step time algorithm is fixed to 01 forthe two methods Algorithms are implemented on Matlab

Advances in Materials Science and Engineering 9

Table 1 Evaluation of segmentation accuracy and computationtime

119865-measure Cpu-Time (sec)LIF Prop method LIF Prop method

RD02 052 076 354 41RD 03 042 081 254 38RD04 062 087 186 35

Table 2 Evaluation of cpu-time convergence

Cpu-time (sec)Song Prop method

RD06 826 385RD07 941 481RD 08 967 523RD09 966 477

R2010a (on a PCCore i5 29GHz) Obtained values show thatour proposed model is very fast compared with local activecontour LIF

32 Comparative Test 2 We show in this paragraph a com-parative test between the proposed method and saliencyconvex active contour model proposed by Song and Yan[18] known as SCACM Images in Figure 7(a) correspondsto regions of interest (ROI) selected from different weldradiogram films containing various kinds of defects likeporosities and horizontal and vertical thin cracks Figure 7(b)presents SCACM method detection results and Figure 7(c)shows the segmentation results with the proposed method

Obtained results in Figure 7 show that Song methoddetects only defects with small and round shapes due to theinfluence of the local window size used in this method In theother side our method does not need to use a local windowas is explained before we use a global Gaussian distributionapproximation of the off-center saliency intensities Theproposed method detects defects with various forms andsizes with low false detections Most parts of defects areextracted despite their low contrasts Using Song methodmany low contrasted defects aremissed and not identified Asin previous experiment the convergence time of twomethodsevaluated in Table 2 shows that proposed model achievesdefect segmentation in a fast time comparing with Songmethod We choose parameters of the proposed algorithm as120572 = 001 120573 = 0001 times 255 times 255 120574 = 13 and 119899 = 30 We notealso that we have used the Matlab implementation given bySCACM authors in their website (httpfacultyneueducnyunhyanWebpage20for20articleSCACMDemoSCACMrar)

4 Conclusion

Detection of small defects in low contrasted radiographyimages corrupted with uneven illumination is very compli-cated Usually proposed methods in this area have limitedresults In this work our aim is to improve robustness of

weld defect segmentation in radiography images in order toobtain satisfactory results We have developed a new methodbased on level set active contour guided with an off-centersaliency map The segmentation is achieved when an energyfunction is minimized Different tests on weld radiographyimages with various kinds of defects prove efficiency androbustness of the proposed method A comparison test isshown with two kinds of level set active contour models Thefirst depends only on gray level intensity and the second usesa saliencymap to derive the segmentation active contourThecomparison tests reveal that the proposed method permitsovercoming the problem of nonuniform illumination andthe low contrast level in radiography weld images Moreoverour method shows a fast convergence time comparing withother methods However some weakness of the proposedmethod must be studied in a future work To reduce humaninteraction it is recommended to develop a method to selectautomatically the region of the interest It is also interestingto validate this segmentation method to another radiographyimage application like medical area

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] I Valavanis and D Kosmopoulos ldquoMulticlass defect detectionand classification in weld radiographic images using geometricand texture featuresrdquo Expert Systems with Applications vol 37no 12 pp 7606ndash7614 2010

[2] J Zapata R Vilar and R Ruiz ldquoPerformance evaluation of anautomatic inspection system of weld defects in radiographicimages based on neuro-classifiersrdquo Expert Systems with Appli-cations vol 38 no 7 pp 8812ndash8824 2011

[3] C Stolojescu-Crisan and S Holban ldquoA comparison of X-rayimage segmentation techniquesrdquo Advances in Electrical andComputer Engineering vol 13 no 3 pp 85ndash92 2013

[4] A Mahmoudi and F Regragui ldquoWelding defect detection bysegmentation of radiographic imagesrdquo in Proceedings of theWorld Congress on Computer Science and Information Engineer-ing (WRI rsquo09) vol 7 pp 111ndash115 Los Angeles Calif USAMarch2009

[5] D Mery and M A Berti ldquoAutomatic detection of weldingdefects using texture featuresrdquo Insight Non-Destructive Testingand Condition Monitoring vol 45 no 10 pp 676ndash681 2003

[6] M A Carrasco and D Mery ldquoSegmentation of welding defectsusing a robust algorithmrdquo Materials Evaluation vol 62 no 11pp 1142ndash1147 2004

[7] E S Amin ldquoApplication of artificial neural networks to evaluateweld defects of nuclear componentsrdquo Journal of Nuclear andRadiation Physics vol 3 no 2 pp 83ndash92 2008

[8] S-B Zhou A-Q Shen and G-F Li ldquoConcrete image segmen-tation based on multiscale mathematic morphology operatorsand Otsu methodrdquo Advances in Materials Science and Engineer-ing vol 2015 Article ID 208473 11 pages 2015

[9] S Osher and N Paragios Geometric Level Set Methods inImaging Vision and Graphics Springer 2003

10 Advances in Materials Science and Engineering

[10] X-F Wang D-S Huang and H Xu ldquoAn efficient local Chan-Vese model for image segmentationrdquo Pattern Recognition vol43 no 3 pp 603ndash618 2010

[11] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998

[12] T F Chan and L A Vese ldquoActive contours without edgesrdquo IEEETransactions on Image Processing vol 10 no 2 pp 266ndash2772001

[13] DMumford and J Shah ldquoOptimal approximations by piecewisesmooth functions and associated variational problemsrdquo Com-munications on Pure and Applied Mathematics vol 42 no 5pp 577ndash685 1989

[14] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009

[15] X Liu S-J Peng Y-MCheung Y Y Tang and J-XDu ldquoActivecontours with a joint and region-scalable distributionmetric forinteractive natural image segmentationrdquo IET Image Processingvol 8 no 12 pp 824ndash832 2014

[16] K Zhang L Zhang H Song and W Zhou ldquoActive contourswith selective local or global segmentation a new formulationand level set methodrdquo Image and Vision Computing vol 28 no4 pp 668ndash676 2010

[17] K Zhang H Song and L Zhang ldquoActive contours driven bylocal image fitting energyrdquo Pattern Recognition vol 43 no 4pp 1199ndash1206 2010

[18] K Song and Y Yan ldquoMicro surface defect detection methodfor silicon steel strip based on saliency convex active contourmodelrdquoMathematical Problems in Engineering vol 2013 ArticleID 429094 13 pages 2013

[19] L Itti and C Koch ldquoComputational modelling of visual atten-tionrdquo Nature Reviews Neuroscience vol 2 no 3 pp 194ndash2032001

[20] R Achantay S Hemamiz F Estraday and S SusstrunkyldquoFrequency-tuned salient region detectionrdquo in Proceedings ofthe IEEE Computer Society Conference on Computer Vision andPatternRecognition (CVPR rsquo09) pp 1597ndash1604 IEEE June 2009

[21] X Hou and L Zhang ldquoSaliency detection a spectral residualapproachrdquo in Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition (CVPR rsquo07) pp 1ndash8 Minneapo-lis Minn USA June 2007

[22] S Montabone and A Soto ldquoHuman detection using a mobileplatform and novel features derived from a visual saliencymechanismrdquo Image and Vision Computing vol 28 no 3 pp391ndash402 2010

[23] P Viola and M J Jones ldquoRobust real-time face detectionrdquoInternational Journal of Computer Vision vol 57 no 2 pp 137ndash154 2004

[24] L Wang L He A Mishra and C Li ldquoActive contours drivenby local Gaussian distribution fitting energyrdquo Signal Processingvol 89 no 12 pp 2435ndash2447 2009

[25] httpwwwbamdeenindexhtm[26] httpsenwikipediaorgwikiPrecision and recall

Submit your manuscripts athttpwwwhindawicom

ScientificaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CorrosionInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Polymer ScienceInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CeramicsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CompositesJournal of

NanoparticlesJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Biomaterials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

NanoscienceJournal of

TextilesHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Journal of

NanotechnologyHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

CrystallographyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CoatingsJournal of

Advances in

Materials Science and EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Smart Materials Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MetallurgyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

MaterialsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Nano

materials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal ofNanomaterials

6 Advances in Materials Science and Engineering

ROI

(a)

RD 0

1

(b)

RD 0

1

(c)

RD 0

1

(d)

Figure 4 Weld defect detection with the proposed active contour (a) the whole weld radiogram (b) ROI selected image (c) off-centersaliency map and (d) final segmentation with proposed level set active contour (blue color)

ROI

(a)

RD 0

2

(b)

RD 0

2

(c)

RD 0

2

(d)

Figure 5 Weld defect detection with the proposed active contour (a) the whole weld radiogram (b) ROI selected image (c) off-centersaliency map and (d) final segmentation with proposed level set active contour (red color)

Advances in Materials Science and Engineering 7RD

03

RD 0

4RD

05

(a)

RD 0

3RD

04

RD 0

5

(b)RD

03

RD 0

4RD

05

(c)

Figure 6 Comparative defect detection tests (a) ROI selected image (b) result of LIF active contour and (c) result of the proposed method

in high density mode The original 12-bit data depth wasrescaled to 8 bits with a linear LUT proportional to opticalfilm density by visual adjustment to the image content Thepixel size is 403 micron (630 dpi) and the images are 8-bit gray values Due to the big size of radiography weldimages the nonuniform illumination and the small defectshape it is difficult to detect the presence of small defectsand determine accurately their sizes during the radiogramvisualization Consequently for the seeking of simplifyingthe task one could begin by selecting the region of interestROI which can be considered as the parts of the imagewhere the radiograph interpreters suspect the presence ofimperfectionsThe selection of the ROI prevents the operatorto make treatments on the irrelevant regions of the imageMoreover it allows reducing the computing time for real-time applications noting that the technique of ROI local-ization is commonly used by researchers in several worksAfter the selection of an ROI region the obtained images areresized to 250 times 350

We present in Figures 4 and 5 the main steps of theproposed segmentation algorithm on two weld radiographyexamples First row corresponds to thewhole weld radiogramthen a region of interest (ROI) is selected Next images inFigures 4(b) and 5(b) show the selected region to processincluding various defects like porosities and slags distributedat the weld middle The off-center saliency map is shownin the next Figures 4(c) and 5(c) as we can see defects are

enhanced and the background is suppressed Figures 4(d)and 5(d) depict the defect contour detection with a blue andred colour obtained with the proposed off-center saliencylevel set active contour Even though the low contrast andthe nonuniform illumination we can see that the most partsof defects are segmented with low false detections We notethat parameters of the proposed method are set as Δ119905 = 01120572 = 004 120573 = 0001 times 255 times 255 120574 = 15 and 119899 = 30

31 Comparative Test 1 We show a comparison between theproposed model and a level set active contour dependingonly on gray level image intensity proposed by Zhang et al[17] and called the local image fitting active contour LIFLIF model Matlab implementation is given in authorrsquos web-site (httpkaihuazhangnetJ papersPR 10rar) Figure 6(b)presents results of the local fittingmodel LIF and the results ofthe proposed active contour model are shown in Figure 6(c)Parameters of the proposed algorithm are fixed as Δ119905 = 01120572 = 002 120573 = 0001 times 255 times 255 120574 = 10 and 119899 = 30

Through a visual evaluation the superiority of the pro-posed level set active contour can be verified Local fittingactive contour LIF [17] fails to obtain satisfactory resultsdue to the influence of the clutter in the background andthe low contrast of defects Much false detection and a lotof defects are not detected The best results are obtainedwith the proposed level set active contour The most parts ofdefects are identified with low false detections A quantitative

8 Advances in Materials Science and EngineeringRD

06

RD 0

7RD

08

RD 0

9

(a)

RD 0

6RD

07

RD 0

8RD

09

(b)

RD 0

6RD

07

RD 0

8RD

09

(c)

Figure 7 Comparative defect segmentation tests (a) ROI selected image (b) results of Songmethod and (c) results of the proposedmethod

comparison between the proposed active contour and LIFmethod is shown in the Table 1 We compute a segmentationevaluation measure used by many researchers called the 119865-measure [26] To compute this evaluation criterion we needideal image segmentation and the proposed segmentationresults the two results should be given in binary masks Notethat ground truth or ideal segmentations used are obtained

from expert visual interpretation We note also that a highervalue (max 1) denotes accurate segmentation The detectionrates are shown in Table 1 which demonstrates the highperformance of our method comparing with LIF approachMoreover the computation time (cpu-time) is computed fortwo methods The step time algorithm is fixed to 01 forthe two methods Algorithms are implemented on Matlab

Advances in Materials Science and Engineering 9

Table 1 Evaluation of segmentation accuracy and computationtime

119865-measure Cpu-Time (sec)LIF Prop method LIF Prop method

RD02 052 076 354 41RD 03 042 081 254 38RD04 062 087 186 35

Table 2 Evaluation of cpu-time convergence

Cpu-time (sec)Song Prop method

RD06 826 385RD07 941 481RD 08 967 523RD09 966 477

R2010a (on a PCCore i5 29GHz) Obtained values show thatour proposed model is very fast compared with local activecontour LIF

32 Comparative Test 2 We show in this paragraph a com-parative test between the proposed method and saliencyconvex active contour model proposed by Song and Yan[18] known as SCACM Images in Figure 7(a) correspondsto regions of interest (ROI) selected from different weldradiogram films containing various kinds of defects likeporosities and horizontal and vertical thin cracks Figure 7(b)presents SCACM method detection results and Figure 7(c)shows the segmentation results with the proposed method

Obtained results in Figure 7 show that Song methoddetects only defects with small and round shapes due to theinfluence of the local window size used in this method In theother side our method does not need to use a local windowas is explained before we use a global Gaussian distributionapproximation of the off-center saliency intensities Theproposed method detects defects with various forms andsizes with low false detections Most parts of defects areextracted despite their low contrasts Using Song methodmany low contrasted defects aremissed and not identified Asin previous experiment the convergence time of twomethodsevaluated in Table 2 shows that proposed model achievesdefect segmentation in a fast time comparing with Songmethod We choose parameters of the proposed algorithm as120572 = 001 120573 = 0001 times 255 times 255 120574 = 13 and 119899 = 30 We notealso that we have used the Matlab implementation given bySCACM authors in their website (httpfacultyneueducnyunhyanWebpage20for20articleSCACMDemoSCACMrar)

4 Conclusion

Detection of small defects in low contrasted radiographyimages corrupted with uneven illumination is very compli-cated Usually proposed methods in this area have limitedresults In this work our aim is to improve robustness of

weld defect segmentation in radiography images in order toobtain satisfactory results We have developed a new methodbased on level set active contour guided with an off-centersaliency map The segmentation is achieved when an energyfunction is minimized Different tests on weld radiographyimages with various kinds of defects prove efficiency androbustness of the proposed method A comparison test isshown with two kinds of level set active contour models Thefirst depends only on gray level intensity and the second usesa saliencymap to derive the segmentation active contourThecomparison tests reveal that the proposed method permitsovercoming the problem of nonuniform illumination andthe low contrast level in radiography weld images Moreoverour method shows a fast convergence time comparing withother methods However some weakness of the proposedmethod must be studied in a future work To reduce humaninteraction it is recommended to develop a method to selectautomatically the region of the interest It is also interestingto validate this segmentation method to another radiographyimage application like medical area

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] I Valavanis and D Kosmopoulos ldquoMulticlass defect detectionand classification in weld radiographic images using geometricand texture featuresrdquo Expert Systems with Applications vol 37no 12 pp 7606ndash7614 2010

[2] J Zapata R Vilar and R Ruiz ldquoPerformance evaluation of anautomatic inspection system of weld defects in radiographicimages based on neuro-classifiersrdquo Expert Systems with Appli-cations vol 38 no 7 pp 8812ndash8824 2011

[3] C Stolojescu-Crisan and S Holban ldquoA comparison of X-rayimage segmentation techniquesrdquo Advances in Electrical andComputer Engineering vol 13 no 3 pp 85ndash92 2013

[4] A Mahmoudi and F Regragui ldquoWelding defect detection bysegmentation of radiographic imagesrdquo in Proceedings of theWorld Congress on Computer Science and Information Engineer-ing (WRI rsquo09) vol 7 pp 111ndash115 Los Angeles Calif USAMarch2009

[5] D Mery and M A Berti ldquoAutomatic detection of weldingdefects using texture featuresrdquo Insight Non-Destructive Testingand Condition Monitoring vol 45 no 10 pp 676ndash681 2003

[6] M A Carrasco and D Mery ldquoSegmentation of welding defectsusing a robust algorithmrdquo Materials Evaluation vol 62 no 11pp 1142ndash1147 2004

[7] E S Amin ldquoApplication of artificial neural networks to evaluateweld defects of nuclear componentsrdquo Journal of Nuclear andRadiation Physics vol 3 no 2 pp 83ndash92 2008

[8] S-B Zhou A-Q Shen and G-F Li ldquoConcrete image segmen-tation based on multiscale mathematic morphology operatorsand Otsu methodrdquo Advances in Materials Science and Engineer-ing vol 2015 Article ID 208473 11 pages 2015

[9] S Osher and N Paragios Geometric Level Set Methods inImaging Vision and Graphics Springer 2003

10 Advances in Materials Science and Engineering

[10] X-F Wang D-S Huang and H Xu ldquoAn efficient local Chan-Vese model for image segmentationrdquo Pattern Recognition vol43 no 3 pp 603ndash618 2010

[11] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998

[12] T F Chan and L A Vese ldquoActive contours without edgesrdquo IEEETransactions on Image Processing vol 10 no 2 pp 266ndash2772001

[13] DMumford and J Shah ldquoOptimal approximations by piecewisesmooth functions and associated variational problemsrdquo Com-munications on Pure and Applied Mathematics vol 42 no 5pp 577ndash685 1989

[14] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009

[15] X Liu S-J Peng Y-MCheung Y Y Tang and J-XDu ldquoActivecontours with a joint and region-scalable distributionmetric forinteractive natural image segmentationrdquo IET Image Processingvol 8 no 12 pp 824ndash832 2014

[16] K Zhang L Zhang H Song and W Zhou ldquoActive contourswith selective local or global segmentation a new formulationand level set methodrdquo Image and Vision Computing vol 28 no4 pp 668ndash676 2010

[17] K Zhang H Song and L Zhang ldquoActive contours driven bylocal image fitting energyrdquo Pattern Recognition vol 43 no 4pp 1199ndash1206 2010

[18] K Song and Y Yan ldquoMicro surface defect detection methodfor silicon steel strip based on saliency convex active contourmodelrdquoMathematical Problems in Engineering vol 2013 ArticleID 429094 13 pages 2013

[19] L Itti and C Koch ldquoComputational modelling of visual atten-tionrdquo Nature Reviews Neuroscience vol 2 no 3 pp 194ndash2032001

[20] R Achantay S Hemamiz F Estraday and S SusstrunkyldquoFrequency-tuned salient region detectionrdquo in Proceedings ofthe IEEE Computer Society Conference on Computer Vision andPatternRecognition (CVPR rsquo09) pp 1597ndash1604 IEEE June 2009

[21] X Hou and L Zhang ldquoSaliency detection a spectral residualapproachrdquo in Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition (CVPR rsquo07) pp 1ndash8 Minneapo-lis Minn USA June 2007

[22] S Montabone and A Soto ldquoHuman detection using a mobileplatform and novel features derived from a visual saliencymechanismrdquo Image and Vision Computing vol 28 no 3 pp391ndash402 2010

[23] P Viola and M J Jones ldquoRobust real-time face detectionrdquoInternational Journal of Computer Vision vol 57 no 2 pp 137ndash154 2004

[24] L Wang L He A Mishra and C Li ldquoActive contours drivenby local Gaussian distribution fitting energyrdquo Signal Processingvol 89 no 12 pp 2435ndash2447 2009

[25] httpwwwbamdeenindexhtm[26] httpsenwikipediaorgwikiPrecision and recall

Submit your manuscripts athttpwwwhindawicom

ScientificaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CorrosionInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Polymer ScienceInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CeramicsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CompositesJournal of

NanoparticlesJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Biomaterials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

NanoscienceJournal of

TextilesHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Journal of

NanotechnologyHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

CrystallographyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CoatingsJournal of

Advances in

Materials Science and EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Smart Materials Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MetallurgyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

MaterialsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Nano

materials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal ofNanomaterials

Advances in Materials Science and Engineering 7RD

03

RD 0

4RD

05

(a)

RD 0

3RD

04

RD 0

5

(b)RD

03

RD 0

4RD

05

(c)

Figure 6 Comparative defect detection tests (a) ROI selected image (b) result of LIF active contour and (c) result of the proposed method

in high density mode The original 12-bit data depth wasrescaled to 8 bits with a linear LUT proportional to opticalfilm density by visual adjustment to the image content Thepixel size is 403 micron (630 dpi) and the images are 8-bit gray values Due to the big size of radiography weldimages the nonuniform illumination and the small defectshape it is difficult to detect the presence of small defectsand determine accurately their sizes during the radiogramvisualization Consequently for the seeking of simplifyingthe task one could begin by selecting the region of interestROI which can be considered as the parts of the imagewhere the radiograph interpreters suspect the presence ofimperfectionsThe selection of the ROI prevents the operatorto make treatments on the irrelevant regions of the imageMoreover it allows reducing the computing time for real-time applications noting that the technique of ROI local-ization is commonly used by researchers in several worksAfter the selection of an ROI region the obtained images areresized to 250 times 350

We present in Figures 4 and 5 the main steps of theproposed segmentation algorithm on two weld radiographyexamples First row corresponds to thewhole weld radiogramthen a region of interest (ROI) is selected Next images inFigures 4(b) and 5(b) show the selected region to processincluding various defects like porosities and slags distributedat the weld middle The off-center saliency map is shownin the next Figures 4(c) and 5(c) as we can see defects are

enhanced and the background is suppressed Figures 4(d)and 5(d) depict the defect contour detection with a blue andred colour obtained with the proposed off-center saliencylevel set active contour Even though the low contrast andthe nonuniform illumination we can see that the most partsof defects are segmented with low false detections We notethat parameters of the proposed method are set as Δ119905 = 01120572 = 004 120573 = 0001 times 255 times 255 120574 = 15 and 119899 = 30

31 Comparative Test 1 We show a comparison between theproposed model and a level set active contour dependingonly on gray level image intensity proposed by Zhang et al[17] and called the local image fitting active contour LIFLIF model Matlab implementation is given in authorrsquos web-site (httpkaihuazhangnetJ papersPR 10rar) Figure 6(b)presents results of the local fittingmodel LIF and the results ofthe proposed active contour model are shown in Figure 6(c)Parameters of the proposed algorithm are fixed as Δ119905 = 01120572 = 002 120573 = 0001 times 255 times 255 120574 = 10 and 119899 = 30

Through a visual evaluation the superiority of the pro-posed level set active contour can be verified Local fittingactive contour LIF [17] fails to obtain satisfactory resultsdue to the influence of the clutter in the background andthe low contrast of defects Much false detection and a lotof defects are not detected The best results are obtainedwith the proposed level set active contour The most parts ofdefects are identified with low false detections A quantitative

8 Advances in Materials Science and EngineeringRD

06

RD 0

7RD

08

RD 0

9

(a)

RD 0

6RD

07

RD 0

8RD

09

(b)

RD 0

6RD

07

RD 0

8RD

09

(c)

Figure 7 Comparative defect segmentation tests (a) ROI selected image (b) results of Songmethod and (c) results of the proposedmethod

comparison between the proposed active contour and LIFmethod is shown in the Table 1 We compute a segmentationevaluation measure used by many researchers called the 119865-measure [26] To compute this evaluation criterion we needideal image segmentation and the proposed segmentationresults the two results should be given in binary masks Notethat ground truth or ideal segmentations used are obtained

from expert visual interpretation We note also that a highervalue (max 1) denotes accurate segmentation The detectionrates are shown in Table 1 which demonstrates the highperformance of our method comparing with LIF approachMoreover the computation time (cpu-time) is computed fortwo methods The step time algorithm is fixed to 01 forthe two methods Algorithms are implemented on Matlab

Advances in Materials Science and Engineering 9

Table 1 Evaluation of segmentation accuracy and computationtime

119865-measure Cpu-Time (sec)LIF Prop method LIF Prop method

RD02 052 076 354 41RD 03 042 081 254 38RD04 062 087 186 35

Table 2 Evaluation of cpu-time convergence

Cpu-time (sec)Song Prop method

RD06 826 385RD07 941 481RD 08 967 523RD09 966 477

R2010a (on a PCCore i5 29GHz) Obtained values show thatour proposed model is very fast compared with local activecontour LIF

32 Comparative Test 2 We show in this paragraph a com-parative test between the proposed method and saliencyconvex active contour model proposed by Song and Yan[18] known as SCACM Images in Figure 7(a) correspondsto regions of interest (ROI) selected from different weldradiogram films containing various kinds of defects likeporosities and horizontal and vertical thin cracks Figure 7(b)presents SCACM method detection results and Figure 7(c)shows the segmentation results with the proposed method

Obtained results in Figure 7 show that Song methoddetects only defects with small and round shapes due to theinfluence of the local window size used in this method In theother side our method does not need to use a local windowas is explained before we use a global Gaussian distributionapproximation of the off-center saliency intensities Theproposed method detects defects with various forms andsizes with low false detections Most parts of defects areextracted despite their low contrasts Using Song methodmany low contrasted defects aremissed and not identified Asin previous experiment the convergence time of twomethodsevaluated in Table 2 shows that proposed model achievesdefect segmentation in a fast time comparing with Songmethod We choose parameters of the proposed algorithm as120572 = 001 120573 = 0001 times 255 times 255 120574 = 13 and 119899 = 30 We notealso that we have used the Matlab implementation given bySCACM authors in their website (httpfacultyneueducnyunhyanWebpage20for20articleSCACMDemoSCACMrar)

4 Conclusion

Detection of small defects in low contrasted radiographyimages corrupted with uneven illumination is very compli-cated Usually proposed methods in this area have limitedresults In this work our aim is to improve robustness of

weld defect segmentation in radiography images in order toobtain satisfactory results We have developed a new methodbased on level set active contour guided with an off-centersaliency map The segmentation is achieved when an energyfunction is minimized Different tests on weld radiographyimages with various kinds of defects prove efficiency androbustness of the proposed method A comparison test isshown with two kinds of level set active contour models Thefirst depends only on gray level intensity and the second usesa saliencymap to derive the segmentation active contourThecomparison tests reveal that the proposed method permitsovercoming the problem of nonuniform illumination andthe low contrast level in radiography weld images Moreoverour method shows a fast convergence time comparing withother methods However some weakness of the proposedmethod must be studied in a future work To reduce humaninteraction it is recommended to develop a method to selectautomatically the region of the interest It is also interestingto validate this segmentation method to another radiographyimage application like medical area

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] I Valavanis and D Kosmopoulos ldquoMulticlass defect detectionand classification in weld radiographic images using geometricand texture featuresrdquo Expert Systems with Applications vol 37no 12 pp 7606ndash7614 2010

[2] J Zapata R Vilar and R Ruiz ldquoPerformance evaluation of anautomatic inspection system of weld defects in radiographicimages based on neuro-classifiersrdquo Expert Systems with Appli-cations vol 38 no 7 pp 8812ndash8824 2011

[3] C Stolojescu-Crisan and S Holban ldquoA comparison of X-rayimage segmentation techniquesrdquo Advances in Electrical andComputer Engineering vol 13 no 3 pp 85ndash92 2013

[4] A Mahmoudi and F Regragui ldquoWelding defect detection bysegmentation of radiographic imagesrdquo in Proceedings of theWorld Congress on Computer Science and Information Engineer-ing (WRI rsquo09) vol 7 pp 111ndash115 Los Angeles Calif USAMarch2009

[5] D Mery and M A Berti ldquoAutomatic detection of weldingdefects using texture featuresrdquo Insight Non-Destructive Testingand Condition Monitoring vol 45 no 10 pp 676ndash681 2003

[6] M A Carrasco and D Mery ldquoSegmentation of welding defectsusing a robust algorithmrdquo Materials Evaluation vol 62 no 11pp 1142ndash1147 2004

[7] E S Amin ldquoApplication of artificial neural networks to evaluateweld defects of nuclear componentsrdquo Journal of Nuclear andRadiation Physics vol 3 no 2 pp 83ndash92 2008

[8] S-B Zhou A-Q Shen and G-F Li ldquoConcrete image segmen-tation based on multiscale mathematic morphology operatorsand Otsu methodrdquo Advances in Materials Science and Engineer-ing vol 2015 Article ID 208473 11 pages 2015

[9] S Osher and N Paragios Geometric Level Set Methods inImaging Vision and Graphics Springer 2003

10 Advances in Materials Science and Engineering

[10] X-F Wang D-S Huang and H Xu ldquoAn efficient local Chan-Vese model for image segmentationrdquo Pattern Recognition vol43 no 3 pp 603ndash618 2010

[11] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998

[12] T F Chan and L A Vese ldquoActive contours without edgesrdquo IEEETransactions on Image Processing vol 10 no 2 pp 266ndash2772001

[13] DMumford and J Shah ldquoOptimal approximations by piecewisesmooth functions and associated variational problemsrdquo Com-munications on Pure and Applied Mathematics vol 42 no 5pp 577ndash685 1989

[14] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009

[15] X Liu S-J Peng Y-MCheung Y Y Tang and J-XDu ldquoActivecontours with a joint and region-scalable distributionmetric forinteractive natural image segmentationrdquo IET Image Processingvol 8 no 12 pp 824ndash832 2014

[16] K Zhang L Zhang H Song and W Zhou ldquoActive contourswith selective local or global segmentation a new formulationand level set methodrdquo Image and Vision Computing vol 28 no4 pp 668ndash676 2010

[17] K Zhang H Song and L Zhang ldquoActive contours driven bylocal image fitting energyrdquo Pattern Recognition vol 43 no 4pp 1199ndash1206 2010

[18] K Song and Y Yan ldquoMicro surface defect detection methodfor silicon steel strip based on saliency convex active contourmodelrdquoMathematical Problems in Engineering vol 2013 ArticleID 429094 13 pages 2013

[19] L Itti and C Koch ldquoComputational modelling of visual atten-tionrdquo Nature Reviews Neuroscience vol 2 no 3 pp 194ndash2032001

[20] R Achantay S Hemamiz F Estraday and S SusstrunkyldquoFrequency-tuned salient region detectionrdquo in Proceedings ofthe IEEE Computer Society Conference on Computer Vision andPatternRecognition (CVPR rsquo09) pp 1597ndash1604 IEEE June 2009

[21] X Hou and L Zhang ldquoSaliency detection a spectral residualapproachrdquo in Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition (CVPR rsquo07) pp 1ndash8 Minneapo-lis Minn USA June 2007

[22] S Montabone and A Soto ldquoHuman detection using a mobileplatform and novel features derived from a visual saliencymechanismrdquo Image and Vision Computing vol 28 no 3 pp391ndash402 2010

[23] P Viola and M J Jones ldquoRobust real-time face detectionrdquoInternational Journal of Computer Vision vol 57 no 2 pp 137ndash154 2004

[24] L Wang L He A Mishra and C Li ldquoActive contours drivenby local Gaussian distribution fitting energyrdquo Signal Processingvol 89 no 12 pp 2435ndash2447 2009

[25] httpwwwbamdeenindexhtm[26] httpsenwikipediaorgwikiPrecision and recall

Submit your manuscripts athttpwwwhindawicom

ScientificaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CorrosionInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Polymer ScienceInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CeramicsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CompositesJournal of

NanoparticlesJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Biomaterials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

NanoscienceJournal of

TextilesHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Journal of

NanotechnologyHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

CrystallographyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CoatingsJournal of

Advances in

Materials Science and EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Smart Materials Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MetallurgyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

MaterialsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Nano

materials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal ofNanomaterials

8 Advances in Materials Science and EngineeringRD

06

RD 0

7RD

08

RD 0

9

(a)

RD 0

6RD

07

RD 0

8RD

09

(b)

RD 0

6RD

07

RD 0

8RD

09

(c)

Figure 7 Comparative defect segmentation tests (a) ROI selected image (b) results of Songmethod and (c) results of the proposedmethod

comparison between the proposed active contour and LIFmethod is shown in the Table 1 We compute a segmentationevaluation measure used by many researchers called the 119865-measure [26] To compute this evaluation criterion we needideal image segmentation and the proposed segmentationresults the two results should be given in binary masks Notethat ground truth or ideal segmentations used are obtained

from expert visual interpretation We note also that a highervalue (max 1) denotes accurate segmentation The detectionrates are shown in Table 1 which demonstrates the highperformance of our method comparing with LIF approachMoreover the computation time (cpu-time) is computed fortwo methods The step time algorithm is fixed to 01 forthe two methods Algorithms are implemented on Matlab

Advances in Materials Science and Engineering 9

Table 1 Evaluation of segmentation accuracy and computationtime

119865-measure Cpu-Time (sec)LIF Prop method LIF Prop method

RD02 052 076 354 41RD 03 042 081 254 38RD04 062 087 186 35

Table 2 Evaluation of cpu-time convergence

Cpu-time (sec)Song Prop method

RD06 826 385RD07 941 481RD 08 967 523RD09 966 477

R2010a (on a PCCore i5 29GHz) Obtained values show thatour proposed model is very fast compared with local activecontour LIF

32 Comparative Test 2 We show in this paragraph a com-parative test between the proposed method and saliencyconvex active contour model proposed by Song and Yan[18] known as SCACM Images in Figure 7(a) correspondsto regions of interest (ROI) selected from different weldradiogram films containing various kinds of defects likeporosities and horizontal and vertical thin cracks Figure 7(b)presents SCACM method detection results and Figure 7(c)shows the segmentation results with the proposed method

Obtained results in Figure 7 show that Song methoddetects only defects with small and round shapes due to theinfluence of the local window size used in this method In theother side our method does not need to use a local windowas is explained before we use a global Gaussian distributionapproximation of the off-center saliency intensities Theproposed method detects defects with various forms andsizes with low false detections Most parts of defects areextracted despite their low contrasts Using Song methodmany low contrasted defects aremissed and not identified Asin previous experiment the convergence time of twomethodsevaluated in Table 2 shows that proposed model achievesdefect segmentation in a fast time comparing with Songmethod We choose parameters of the proposed algorithm as120572 = 001 120573 = 0001 times 255 times 255 120574 = 13 and 119899 = 30 We notealso that we have used the Matlab implementation given bySCACM authors in their website (httpfacultyneueducnyunhyanWebpage20for20articleSCACMDemoSCACMrar)

4 Conclusion

Detection of small defects in low contrasted radiographyimages corrupted with uneven illumination is very compli-cated Usually proposed methods in this area have limitedresults In this work our aim is to improve robustness of

weld defect segmentation in radiography images in order toobtain satisfactory results We have developed a new methodbased on level set active contour guided with an off-centersaliency map The segmentation is achieved when an energyfunction is minimized Different tests on weld radiographyimages with various kinds of defects prove efficiency androbustness of the proposed method A comparison test isshown with two kinds of level set active contour models Thefirst depends only on gray level intensity and the second usesa saliencymap to derive the segmentation active contourThecomparison tests reveal that the proposed method permitsovercoming the problem of nonuniform illumination andthe low contrast level in radiography weld images Moreoverour method shows a fast convergence time comparing withother methods However some weakness of the proposedmethod must be studied in a future work To reduce humaninteraction it is recommended to develop a method to selectautomatically the region of the interest It is also interestingto validate this segmentation method to another radiographyimage application like medical area

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] I Valavanis and D Kosmopoulos ldquoMulticlass defect detectionand classification in weld radiographic images using geometricand texture featuresrdquo Expert Systems with Applications vol 37no 12 pp 7606ndash7614 2010

[2] J Zapata R Vilar and R Ruiz ldquoPerformance evaluation of anautomatic inspection system of weld defects in radiographicimages based on neuro-classifiersrdquo Expert Systems with Appli-cations vol 38 no 7 pp 8812ndash8824 2011

[3] C Stolojescu-Crisan and S Holban ldquoA comparison of X-rayimage segmentation techniquesrdquo Advances in Electrical andComputer Engineering vol 13 no 3 pp 85ndash92 2013

[4] A Mahmoudi and F Regragui ldquoWelding defect detection bysegmentation of radiographic imagesrdquo in Proceedings of theWorld Congress on Computer Science and Information Engineer-ing (WRI rsquo09) vol 7 pp 111ndash115 Los Angeles Calif USAMarch2009

[5] D Mery and M A Berti ldquoAutomatic detection of weldingdefects using texture featuresrdquo Insight Non-Destructive Testingand Condition Monitoring vol 45 no 10 pp 676ndash681 2003

[6] M A Carrasco and D Mery ldquoSegmentation of welding defectsusing a robust algorithmrdquo Materials Evaluation vol 62 no 11pp 1142ndash1147 2004

[7] E S Amin ldquoApplication of artificial neural networks to evaluateweld defects of nuclear componentsrdquo Journal of Nuclear andRadiation Physics vol 3 no 2 pp 83ndash92 2008

[8] S-B Zhou A-Q Shen and G-F Li ldquoConcrete image segmen-tation based on multiscale mathematic morphology operatorsand Otsu methodrdquo Advances in Materials Science and Engineer-ing vol 2015 Article ID 208473 11 pages 2015

[9] S Osher and N Paragios Geometric Level Set Methods inImaging Vision and Graphics Springer 2003

10 Advances in Materials Science and Engineering

[10] X-F Wang D-S Huang and H Xu ldquoAn efficient local Chan-Vese model for image segmentationrdquo Pattern Recognition vol43 no 3 pp 603ndash618 2010

[11] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998

[12] T F Chan and L A Vese ldquoActive contours without edgesrdquo IEEETransactions on Image Processing vol 10 no 2 pp 266ndash2772001

[13] DMumford and J Shah ldquoOptimal approximations by piecewisesmooth functions and associated variational problemsrdquo Com-munications on Pure and Applied Mathematics vol 42 no 5pp 577ndash685 1989

[14] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009

[15] X Liu S-J Peng Y-MCheung Y Y Tang and J-XDu ldquoActivecontours with a joint and region-scalable distributionmetric forinteractive natural image segmentationrdquo IET Image Processingvol 8 no 12 pp 824ndash832 2014

[16] K Zhang L Zhang H Song and W Zhou ldquoActive contourswith selective local or global segmentation a new formulationand level set methodrdquo Image and Vision Computing vol 28 no4 pp 668ndash676 2010

[17] K Zhang H Song and L Zhang ldquoActive contours driven bylocal image fitting energyrdquo Pattern Recognition vol 43 no 4pp 1199ndash1206 2010

[18] K Song and Y Yan ldquoMicro surface defect detection methodfor silicon steel strip based on saliency convex active contourmodelrdquoMathematical Problems in Engineering vol 2013 ArticleID 429094 13 pages 2013

[19] L Itti and C Koch ldquoComputational modelling of visual atten-tionrdquo Nature Reviews Neuroscience vol 2 no 3 pp 194ndash2032001

[20] R Achantay S Hemamiz F Estraday and S SusstrunkyldquoFrequency-tuned salient region detectionrdquo in Proceedings ofthe IEEE Computer Society Conference on Computer Vision andPatternRecognition (CVPR rsquo09) pp 1597ndash1604 IEEE June 2009

[21] X Hou and L Zhang ldquoSaliency detection a spectral residualapproachrdquo in Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition (CVPR rsquo07) pp 1ndash8 Minneapo-lis Minn USA June 2007

[22] S Montabone and A Soto ldquoHuman detection using a mobileplatform and novel features derived from a visual saliencymechanismrdquo Image and Vision Computing vol 28 no 3 pp391ndash402 2010

[23] P Viola and M J Jones ldquoRobust real-time face detectionrdquoInternational Journal of Computer Vision vol 57 no 2 pp 137ndash154 2004

[24] L Wang L He A Mishra and C Li ldquoActive contours drivenby local Gaussian distribution fitting energyrdquo Signal Processingvol 89 no 12 pp 2435ndash2447 2009

[25] httpwwwbamdeenindexhtm[26] httpsenwikipediaorgwikiPrecision and recall

Submit your manuscripts athttpwwwhindawicom

ScientificaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CorrosionInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Polymer ScienceInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CeramicsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CompositesJournal of

NanoparticlesJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Biomaterials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

NanoscienceJournal of

TextilesHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Journal of

NanotechnologyHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

CrystallographyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CoatingsJournal of

Advances in

Materials Science and EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Smart Materials Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MetallurgyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

MaterialsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Nano

materials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal ofNanomaterials

Advances in Materials Science and Engineering 9

Table 1 Evaluation of segmentation accuracy and computationtime

119865-measure Cpu-Time (sec)LIF Prop method LIF Prop method

RD02 052 076 354 41RD 03 042 081 254 38RD04 062 087 186 35

Table 2 Evaluation of cpu-time convergence

Cpu-time (sec)Song Prop method

RD06 826 385RD07 941 481RD 08 967 523RD09 966 477

R2010a (on a PCCore i5 29GHz) Obtained values show thatour proposed model is very fast compared with local activecontour LIF

32 Comparative Test 2 We show in this paragraph a com-parative test between the proposed method and saliencyconvex active contour model proposed by Song and Yan[18] known as SCACM Images in Figure 7(a) correspondsto regions of interest (ROI) selected from different weldradiogram films containing various kinds of defects likeporosities and horizontal and vertical thin cracks Figure 7(b)presents SCACM method detection results and Figure 7(c)shows the segmentation results with the proposed method

Obtained results in Figure 7 show that Song methoddetects only defects with small and round shapes due to theinfluence of the local window size used in this method In theother side our method does not need to use a local windowas is explained before we use a global Gaussian distributionapproximation of the off-center saliency intensities Theproposed method detects defects with various forms andsizes with low false detections Most parts of defects areextracted despite their low contrasts Using Song methodmany low contrasted defects aremissed and not identified Asin previous experiment the convergence time of twomethodsevaluated in Table 2 shows that proposed model achievesdefect segmentation in a fast time comparing with Songmethod We choose parameters of the proposed algorithm as120572 = 001 120573 = 0001 times 255 times 255 120574 = 13 and 119899 = 30 We notealso that we have used the Matlab implementation given bySCACM authors in their website (httpfacultyneueducnyunhyanWebpage20for20articleSCACMDemoSCACMrar)

4 Conclusion

Detection of small defects in low contrasted radiographyimages corrupted with uneven illumination is very compli-cated Usually proposed methods in this area have limitedresults In this work our aim is to improve robustness of

weld defect segmentation in radiography images in order toobtain satisfactory results We have developed a new methodbased on level set active contour guided with an off-centersaliency map The segmentation is achieved when an energyfunction is minimized Different tests on weld radiographyimages with various kinds of defects prove efficiency androbustness of the proposed method A comparison test isshown with two kinds of level set active contour models Thefirst depends only on gray level intensity and the second usesa saliencymap to derive the segmentation active contourThecomparison tests reveal that the proposed method permitsovercoming the problem of nonuniform illumination andthe low contrast level in radiography weld images Moreoverour method shows a fast convergence time comparing withother methods However some weakness of the proposedmethod must be studied in a future work To reduce humaninteraction it is recommended to develop a method to selectautomatically the region of the interest It is also interestingto validate this segmentation method to another radiographyimage application like medical area

Conflict of Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

References

[1] I Valavanis and D Kosmopoulos ldquoMulticlass defect detectionand classification in weld radiographic images using geometricand texture featuresrdquo Expert Systems with Applications vol 37no 12 pp 7606ndash7614 2010

[2] J Zapata R Vilar and R Ruiz ldquoPerformance evaluation of anautomatic inspection system of weld defects in radiographicimages based on neuro-classifiersrdquo Expert Systems with Appli-cations vol 38 no 7 pp 8812ndash8824 2011

[3] C Stolojescu-Crisan and S Holban ldquoA comparison of X-rayimage segmentation techniquesrdquo Advances in Electrical andComputer Engineering vol 13 no 3 pp 85ndash92 2013

[4] A Mahmoudi and F Regragui ldquoWelding defect detection bysegmentation of radiographic imagesrdquo in Proceedings of theWorld Congress on Computer Science and Information Engineer-ing (WRI rsquo09) vol 7 pp 111ndash115 Los Angeles Calif USAMarch2009

[5] D Mery and M A Berti ldquoAutomatic detection of weldingdefects using texture featuresrdquo Insight Non-Destructive Testingand Condition Monitoring vol 45 no 10 pp 676ndash681 2003

[6] M A Carrasco and D Mery ldquoSegmentation of welding defectsusing a robust algorithmrdquo Materials Evaluation vol 62 no 11pp 1142ndash1147 2004

[7] E S Amin ldquoApplication of artificial neural networks to evaluateweld defects of nuclear componentsrdquo Journal of Nuclear andRadiation Physics vol 3 no 2 pp 83ndash92 2008

[8] S-B Zhou A-Q Shen and G-F Li ldquoConcrete image segmen-tation based on multiscale mathematic morphology operatorsand Otsu methodrdquo Advances in Materials Science and Engineer-ing vol 2015 Article ID 208473 11 pages 2015

[9] S Osher and N Paragios Geometric Level Set Methods inImaging Vision and Graphics Springer 2003

10 Advances in Materials Science and Engineering

[10] X-F Wang D-S Huang and H Xu ldquoAn efficient local Chan-Vese model for image segmentationrdquo Pattern Recognition vol43 no 3 pp 603ndash618 2010

[11] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998

[12] T F Chan and L A Vese ldquoActive contours without edgesrdquo IEEETransactions on Image Processing vol 10 no 2 pp 266ndash2772001

[13] DMumford and J Shah ldquoOptimal approximations by piecewisesmooth functions and associated variational problemsrdquo Com-munications on Pure and Applied Mathematics vol 42 no 5pp 577ndash685 1989

[14] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009

[15] X Liu S-J Peng Y-MCheung Y Y Tang and J-XDu ldquoActivecontours with a joint and region-scalable distributionmetric forinteractive natural image segmentationrdquo IET Image Processingvol 8 no 12 pp 824ndash832 2014

[16] K Zhang L Zhang H Song and W Zhou ldquoActive contourswith selective local or global segmentation a new formulationand level set methodrdquo Image and Vision Computing vol 28 no4 pp 668ndash676 2010

[17] K Zhang H Song and L Zhang ldquoActive contours driven bylocal image fitting energyrdquo Pattern Recognition vol 43 no 4pp 1199ndash1206 2010

[18] K Song and Y Yan ldquoMicro surface defect detection methodfor silicon steel strip based on saliency convex active contourmodelrdquoMathematical Problems in Engineering vol 2013 ArticleID 429094 13 pages 2013

[19] L Itti and C Koch ldquoComputational modelling of visual atten-tionrdquo Nature Reviews Neuroscience vol 2 no 3 pp 194ndash2032001

[20] R Achantay S Hemamiz F Estraday and S SusstrunkyldquoFrequency-tuned salient region detectionrdquo in Proceedings ofthe IEEE Computer Society Conference on Computer Vision andPatternRecognition (CVPR rsquo09) pp 1597ndash1604 IEEE June 2009

[21] X Hou and L Zhang ldquoSaliency detection a spectral residualapproachrdquo in Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition (CVPR rsquo07) pp 1ndash8 Minneapo-lis Minn USA June 2007

[22] S Montabone and A Soto ldquoHuman detection using a mobileplatform and novel features derived from a visual saliencymechanismrdquo Image and Vision Computing vol 28 no 3 pp391ndash402 2010

[23] P Viola and M J Jones ldquoRobust real-time face detectionrdquoInternational Journal of Computer Vision vol 57 no 2 pp 137ndash154 2004

[24] L Wang L He A Mishra and C Li ldquoActive contours drivenby local Gaussian distribution fitting energyrdquo Signal Processingvol 89 no 12 pp 2435ndash2447 2009

[25] httpwwwbamdeenindexhtm[26] httpsenwikipediaorgwikiPrecision and recall

Submit your manuscripts athttpwwwhindawicom

ScientificaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CorrosionInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Polymer ScienceInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CeramicsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CompositesJournal of

NanoparticlesJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Biomaterials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

NanoscienceJournal of

TextilesHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Journal of

NanotechnologyHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

CrystallographyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CoatingsJournal of

Advances in

Materials Science and EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Smart Materials Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MetallurgyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

MaterialsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Nano

materials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal ofNanomaterials

10 Advances in Materials Science and Engineering

[10] X-F Wang D-S Huang and H Xu ldquoAn efficient local Chan-Vese model for image segmentationrdquo Pattern Recognition vol43 no 3 pp 603ndash618 2010

[11] C Xu and J L Prince ldquoSnakes shapes and gradient vector flowrdquoIEEE Transactions on Image Processing vol 7 no 3 pp 359ndash3691998

[12] T F Chan and L A Vese ldquoActive contours without edgesrdquo IEEETransactions on Image Processing vol 10 no 2 pp 266ndash2772001

[13] DMumford and J Shah ldquoOptimal approximations by piecewisesmooth functions and associated variational problemsrdquo Com-munications on Pure and Applied Mathematics vol 42 no 5pp 577ndash685 1989

[14] L Wang C Li Q Sun D Xia and C-Y Kao ldquoActive contoursdriven by local and global intensity fitting energy with applica-tion to brain MR image segmentationrdquo Computerized MedicalImaging and Graphics vol 33 no 7 pp 520ndash531 2009

[15] X Liu S-J Peng Y-MCheung Y Y Tang and J-XDu ldquoActivecontours with a joint and region-scalable distributionmetric forinteractive natural image segmentationrdquo IET Image Processingvol 8 no 12 pp 824ndash832 2014

[16] K Zhang L Zhang H Song and W Zhou ldquoActive contourswith selective local or global segmentation a new formulationand level set methodrdquo Image and Vision Computing vol 28 no4 pp 668ndash676 2010

[17] K Zhang H Song and L Zhang ldquoActive contours driven bylocal image fitting energyrdquo Pattern Recognition vol 43 no 4pp 1199ndash1206 2010

[18] K Song and Y Yan ldquoMicro surface defect detection methodfor silicon steel strip based on saliency convex active contourmodelrdquoMathematical Problems in Engineering vol 2013 ArticleID 429094 13 pages 2013

[19] L Itti and C Koch ldquoComputational modelling of visual atten-tionrdquo Nature Reviews Neuroscience vol 2 no 3 pp 194ndash2032001

[20] R Achantay S Hemamiz F Estraday and S SusstrunkyldquoFrequency-tuned salient region detectionrdquo in Proceedings ofthe IEEE Computer Society Conference on Computer Vision andPatternRecognition (CVPR rsquo09) pp 1597ndash1604 IEEE June 2009

[21] X Hou and L Zhang ldquoSaliency detection a spectral residualapproachrdquo in Proceedings of the IEEE Conference on ComputerVision and Pattern Recognition (CVPR rsquo07) pp 1ndash8 Minneapo-lis Minn USA June 2007

[22] S Montabone and A Soto ldquoHuman detection using a mobileplatform and novel features derived from a visual saliencymechanismrdquo Image and Vision Computing vol 28 no 3 pp391ndash402 2010

[23] P Viola and M J Jones ldquoRobust real-time face detectionrdquoInternational Journal of Computer Vision vol 57 no 2 pp 137ndash154 2004

[24] L Wang L He A Mishra and C Li ldquoActive contours drivenby local Gaussian distribution fitting energyrdquo Signal Processingvol 89 no 12 pp 2435ndash2447 2009

[25] httpwwwbamdeenindexhtm[26] httpsenwikipediaorgwikiPrecision and recall

Submit your manuscripts athttpwwwhindawicom

ScientificaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CorrosionInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Polymer ScienceInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CeramicsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CompositesJournal of

NanoparticlesJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Biomaterials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

NanoscienceJournal of

TextilesHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Journal of

NanotechnologyHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

CrystallographyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CoatingsJournal of

Advances in

Materials Science and EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Smart Materials Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MetallurgyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

MaterialsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Nano

materials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal ofNanomaterials

Submit your manuscripts athttpwwwhindawicom

ScientificaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CorrosionInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Polymer ScienceInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CeramicsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CompositesJournal of

NanoparticlesJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Biomaterials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

NanoscienceJournal of

TextilesHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Journal of

NanotechnologyHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

CrystallographyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

CoatingsJournal of

Advances in

Materials Science and EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Smart Materials Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

MetallurgyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

MaterialsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Nano

materials

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal ofNanomaterials