surface defect detection method using ... - j-stage home

© 2014 ISIJ 2598

ISIJ International, Vol. 54 (2014), No. 11, pp. 2598–2607

Surface Defect Detection Method Using Saliency Linear Scanning Morphology for Silicon Steel Strip under Oil Pollution Interference

Ke-Chen SONG,* Shao-Peng HU, Yun-Hui YAN and Jun LI

School of Mechanical Engineering & Automation, Northeastern University, Shenyang, Liaoning 110819, P.R. China.

(Received on January 8, 2014; accepted on July 16, 2014)

Surface defect detection of silicon steel strip is an important section for non-destructive testing systemin iron and steel industry. To detect the interesting defect objects for silicon steel strip under oil pollutioninterference, a new detection method based on saliency linear scanning morphology is proposed. In theproposed method, visual saliency extraction is employed to suppress the clutter background. Meanwhile,a saliency map is obtained for the purpose of highlighting the potential objects. Then, the linear scanningoperation is proposed to obtain the region of oil pollution. Finally, the morphology edge processing is pro-posed to remove the edge of oil pollution interference and the edge of reflective pseudo-defect. Experi-mental results demonstrate that the proposed method presents the good performance for detectingsurface defects including wipe-crack-defect, scratch-defect and small-defect.

KEY WORDS: surface defect; oil pollution; silicon steel; saliency extraction; linear scanning.

1. IntroductionAs an important metal material, silicon steel strip is mainly

used as core material in transformers and dynamos. Unfor-tunately, the surface of the silicon steel strip inevitably existdifferent types of defects, e.g., wipe-crack and scratches.What’s worse is that these defects are covered in oil pollu-tion. Although blowing equipment can remove some oil pol-lution, there are still many defects suffered the interferenceof oil pollution. In order to detect these defects, a variety ofnon-destructive testing (NDT) systems are developed. Sincethe visual-based inspection technology has the characteris-tics of the real-time, this technology has been widely usedin defect automatic inspection system such as flaw,1) weld2)

and surface crack.3)

At present, there are a few studies on surface defect auto-matic inspection in iron and steel industry. For instance, Yunet al.4) developed the univariate dynamic encoding algo-rithm for searches (uDEAS) to detect the cracks, and Pan etal.5) exploited an engineering-driven rule-based detection(ERD) method for bleed detection in visual images whichlie in the low signal-to-noise ratio. In addition, Landströmet al.6) focused on automated detection of longitudinalcracks in steel slabs based on morphology theory, andBulnes et al.7) introduced the clustering method to detectperiodical defects. These four studies mentioned abovefocus on locating the positions of the defects while somestudies focus on defect features extraction such as gabor fil-ters,8) wavelet filters,9) and multi-scale geometric analysis(MGA).10) Furthermore, the extracted features can be clas-sified by diversified classifiers including Bayesiannetwork11) and process knowledge based support vector(PK-MSVM).12)

Despite the several studies mentioned above haveachieved good performance in their specific type of defect,

it is difficult to employ these methods directly to detect thesurface defects for silicon steel strip under oil pollutioninterference. In order to detect the interesting defect objects,three important challenges are need to overcome. For adefect image of silicon steel strip, the background of defectimage is quite complex and has random distribution in thewhole image. Therefore, the cluttered background increasesthe difficulty of detection the defects in image, i.e., the clut-tered background is the first challenge. Since the interestingsurface defect objects are covered by oil pollution, it is nec-essary to remove the useless edge of oil pollution. Hence,how to eliminate the oil pollution interference is the secondchallenge. In addition, due to the reflection of oil pollutionwhen capturing an image, the obtained defect image con-tains some reflective pseudo-defects. Moreover, thesereflective regions are very brightly and easily affect thelocation of defects. Consequently, the interference of reflec-tive pseudo-defects is the third challenge.

In order to solve the three challenges, a new detectionmethod based on saliency linear scanning morphology(SLSM) is proposed to detect surface defects of silicon steelstrip. In the proposed method, firstly, visual saliency extrac-tion is employed to suppress the clutter background. There-fore, a saliency map is obtained for the purpose of highlightingthe potential objects. Then, the saliency map is filtered andconverted into a binary image. Thirdly, the linear scanningoperation is proposed to obtain the region of oil pollutioninterference. Finally, the morphology edge processing isproposed to remove the edge of oil pollution interferenceand the edge of reflective pseudo-defect.

The rest of this paper is organized as follows: Section 2presents the hardware configuration of image acquisitionand the image analysis of surface defect. Section 3 introduc-es the proposed saliency linear scanning morphology meth-od in detail. Then Section 4 elaborates the experiments anddiscusses the experimental results. Finally, Section 5 con-cludes the paper.* Corresponding author: E-mail: [email protected]

DOI: http://dx.doi.org/10.2355/isijinternational.54.2598

ISIJ International, Vol. 54 (2014), No. 11

2599 © 2014 ISIJ

2. Image Acquisition and Analysis of Surface Defect2.1. Image Acquisition

The hardware configuration of the image acquisitionmainly includes two components: light source and camera.The layout of the acquisition equipment is shown in Fig. 1.The light source provides the illumination to make the sur-face defects visible and helps to capture the surface defectimages with camera. Since the light-emitting diode (LED)has many advantages such as little power and longevity, itis used to provide stable illumination. In this work, the LEDis made by CCS incorporated company of Japan with themodel number HLND-1200-SW2. In order to obtain themicro surface defect image, an area scan CCD camera isused. The camera used here is made by Basler of Germanywith the model number acA640-90uc. It has a resolution of658 × 492 pixels with frame rate 90 fps and color image. Inthis work, the captured color image is resized as 640 × 480pixels for the purpose of calculating expediently for defectdetection. In addition, a 55-mm focal length lens is installedon the camera.

2.2. Image Analysis of Surface DefectOne of the original surface defect images of silicon steel

strip under oil pollution interference is shown in Fig. 2(a).As it could be seen in Fig. 2(a), the original defect image

represents an area of 21.5 × 16.1 mm2 in real silicon steelstrip. It is observed that the defect image contains fourcomponents: the interesting surface defect objects and threeinterference components (i.e. clutter background, oil pollutioninterference and reflective pseudo-defect). The interestingsurface defect objects are to be detected while the interfer-ence components are useless.

To illustrate the three interference components, severalinterference regions are cropped from original defect image(i.e. Fig. 2(a)). Figures 2(c) and 2(e) show the clutteredbackground interference and oil pollution interferencerespectively. Moreover, Figs. 2(g) and 2(i) show the inter-ference of reflective pseudo-defect. In addition, the Cannyoperator13) is employed to extract the image edge. The edgeextraction image of original defect image is shown in Fig.2(b). Meanwhile, the edge extraction image of clutter back-ground interference and oil pollution interference are shownin Figs. 2(d) and 2(f) respectively. Furthermore, Figs. 2(h)and 2(j) show the edge extraction images of reflective pseudo-defect.

From Figs. 2(c) and 2(d), we can observed that the back-ground of defect image is quite complex, which is a difficultproblem for detecting the defects. Despite the fact that theclutter background has certain characters of texture, thesecharacters are more random than regular texture. The sur-face plot of the Fig. 2(c) is shown in Fig. 3. As it could beseen in Fig. 3, the clutter background has random distribu-tion in the whole image.

Obviously, Figs. 2(e) and 2(f) show the interesting sur-face defect objects are covered by oil pollution. Althoughthe edge of oil pollution is totally extracted, it is useless todetect the surface defects. Therefore, the useless edge of oilpollution is also a problem for detecting the defects.

In addition, due to the reflection of oil pollution whencapturing an image, the obtained defect image containssome reflective pseudo-defects (i.e. Figs. 2(g) and 2(i)).Moreover, these reflective regions are very brightly and eas-ily be extracted the edges. Hence, these edges of pseudo-defect, to some extent, increase the difficulty of detectionthe defects in image.Fig. 1. Schematic of the image acquisition.

Fig. 2. Sample image of surface defect of silicon steel strip under oil pollution interference. (a) The original surface defectimage of silicon steel strip under oil pollution interference. (b) The edge extraction image of original defect image.(c) The region image of clutter background interference. (d) The edge extraction image for figure (c). (e) Theregion image of oil pollution interference. (f) The edge extraction image for figure (e). (g) A region image ofreflective pseudo-defect. (h) The edge extraction image for figure (g). (i) Another region image of reflectivepseudo-defect. (j) The edge extraction image for figure (i).

© 2014 ISIJ 2600


3. Proposed Saliency Linear Scanning MorphologyMethod

In this section, the proposed method is described in foursubsections. Firstly, a saliency extraction approach is intro-duced in Section 3.1 for the purpose of highlighting thepotential objects to get a saliency map. Then, the saliencymap is filtered and converted into a binary image in Section3.2. Furthermore, the linear scanning operation is proposedto obtain the region of oil pollution interference in Section3.3. Finally, the morphology edge processing is proposed toremove the edge of oil pollution interference and the edgeof reflective pseudo-defect in Section 3.4.

3.1. Saliency ExtractionSaliency extraction is an important step in machine vision,

which has been applied in many tasks including object detec-tion and image segmentation. An excellent saliency extractionmethod can well highlight the potential objects to get a salien-cy map. However, most of the current saliency methods oftengenerate saliency maps that have low resolution or poorlydefined borders. Furthermore, some methods may generatemaps that have ill-defined object boundaries.

In order to avoid these drawbacks mentioned above,Achanta et al.14) exploited a frequency-tuned approach toestimate center-surround contrast using color and luminancefeatures. Although the frequency-tuned approach createdfull resolution saliency maps with well-defined boundariesof salient objects, it may fail to correctly highlight thesalient regions in the presence of salient objects and com-plex textured background. In this work, the symmetric

surround saliency15) is employed to improve the frequency-tuned approach.

For an input image I (x, y), the symmetric surround salien-cy value S (x, y) is obtained as:

................ (1)where I f (x, y) is the corresponding image Lab color spacevector value in the Gaussian blurred version (using a N × Nseparable binomial kernel) of the original image, and isthe L2 norm. Here, the L2 norm is the Euclidean distance. Inthe Lab color space, each pixel location is a [L, a, b]T vector.Different from the mean image feature vector of frequency-tuned approach,14) Iμ (x, y) is the average Lab vector of thesub-image whose center pixel is at position (x, y).15)

To illustrate the calculation procedure of the symmetricsurround saliency, Fig. 4 presents the schematic of thesaliency extraction for input defect image. Firstly, the orig-inal defect image Fig. 4(a) is blurred by N × N Gaussian fil-ter window (here, N is set as 5). Then, Lab color space imag-es (i.e. Figs. 4(c) and 4(d)) of original image (i.e. Fig. 4(a))and filter image (i.e. Fig. 4(b)) are obtained by convertingcolor space from RGB to Lab. Finally, the symmetric sur-round saliency value S (x, y) is obtained with Eq. (1).

For the input defect image Fig. 4(a), the obtained saliencymap and its surface plot are shown in Fig. 5. As it could beseen in Fig. 5(a), the saliency map image by saliency extrac-tion is properly able to suppress the textured backgroundand amplify the difference between the interesting defectobject and the textured background.

3.2. Filtering and ThresholdingIn view of the extracted saliency map in Section 3.1 con-

tains some interference information, it is necessary to filterprocess for saliency map. In this work, the open-close filter-ing based on morphological is employed to filter the inter-ference information. Since the saliency map is a gray-scaleimage, the definition of the morphological operation16) issuitable for gray-scale image. The dilation operation anderosion operation are firstly presented before the definitionof the open-close filtering.

The gray-scale dilation of S by structuring element e,denoted S ⊕ e, is defined as

.... (2)

where De is the domain of e, S (x, y) is assumed to equal –∞Fig. 3. Surface plot of the defect sample image in Fig. 2(c).

S x y x y x yf, , ,( ) ( ) ( )= −I Iμ

S e x y

S x x y y e x y x y De

⊕( )( )= − ′ − ′( ) + ′ ′( ) ′ ′( )∈{ }

,max , , ,

Fig. 4. Schematic of the saliency extraction for the input defect image. (a) The original image. (b) The filter image. (c)The Lab color space image of original image. (d) The Lab color space image of filter image.


2601 © 2014 ISIJ

outside the domain of S.Similar to the definition of gray-scale dilation, the gray-

scale erosion of S by structuring element e, denoted ,is defined as

.... (3)

where De is the domain of e, S (x, y) is assumed to be +∞outside the domain of S.

The morphological opening of image S by structuring ele-ment e, denoted , is simply erosion of S by e, followedby dilation of the result by e:

........................... (4)Similarly, the morphological closing of S by e, denoted

, is a dilation followed by erosion:........................... (5)

In this work, to enhance the filter effectiveness, open-close filtering is performed with a series of structuring ele-ments of increasing size, i.e., alternating sequential filtering.Therefore, the filter image If (x, y) is obtained after the open-close filtering for S (x, y). Figure 6 presents the filtering pro-

cess for saliency map (here, ‘disk’ type structuring elementsis used to perform filter operation, the radius range of alter-nating sequential filtering is from 2 to 3). Figure 6(b) showsthe filter image If (x, y) using open-close filtering operation.Figs. 6(a) and 6(c) show the edge extraction image forsaliency map and filter image respectively. Through thecomparison of Figs. 6(a) and 6(c), we can clearly see thatsome interference information is filtered by open-close fil-tering operation.

After open-close filtering operation, the filter image If (x,y) is need to convert into a binary image for the purpose oflinear scanning in Section 3.3. Before the thresholding oper-ation, the If (x, y) values are firstly normalized to [0-1]. Toobtain a binary image, we tried the Otsu’s17) method, whichthe result image is shown in Fig. 7(a). For Otsu’s method,the threshold value is acquired automatically. However, theresult image presents that some useful information are lostin Fig. 7(a). In order to avoid the issue mentioned above, asimple but effective global thresholding method is employedto obtain the binary image. Different from the Otsu’s meth-od, this global thresholding method determines the thresholdvalue according to the experience of artificial selection.

The definition of the global thresholding method is as fol-lows:

S e

S e x y

S x x y y e x y x y De

( )( )= + ′ + ′( ) − ′ ′( ) ′ ′( )∈{ }

,min , , ,

S e

S e S e e= ( ) ⊕

S e•S e S e e• = ⊕( )

Fig. 5. Saliency map and its surface plot. (a) The saliency map. (b) The surface plot of saliency map.

Fig. 6. The filtering process for saliency map. (a) The edge extraction image for saliency map. (b) The filter image usingopen-close operation. (c) The edge extraction image for filter image. (d) The surface plot for filter image.

© 2014 ISIJ 2602


................. (6)

where T is a threshold value according to the experience ofartificial selection.

Therefore, the thresholding image I t (x, y) is shown in Fig.7(b) (here, the threshold value is set as 0.05). Furthermore,a median filter (here, the filter window is set as 29) isemployed to filter the interference. Hence, a new binaryimage Itm (x, y) is obtained after the filter for image It (x, y).Through the comparison of Figs. 7(a) and 7(c), we canclearly see that more useful information are retained usingthe global thresholding method.

3.3. Linear ScanningDespite the thresholding operation gives rise to a useful

binary image, the region of oil pollution interference cannotbe completely extracted. To obtain the region of oil pollu-tion interference, a linear scanning operation is proposed inthis subsection. The linear scanning operation mainly con-tains two sections: the horizontal orientation linear scanningand the vertical orientation linear scanning. For a binaryimage, the details of this operation are as follows.

For each row, finding the coordinate of zero value, then allof the values between the first zero value and the last zero valueare set as zero value. A new binary image Ih (x, y) is obtainedafter the horizontal linear scanning for image Itm (x, y).

The operation for each column is the same as the opera-tion for each row. Hence, the vertical linear scanning imageIv (x, y) is obtained. Furthermore, the image Iv (x, y) is thefinal linear scanning image Ils (x, y).

To illustrate the linear scanning process, Fig. 8 gives anexample of image Itm (x, y). To see more details, Fig. 8(b)presents the partial region of image Itm (x, y) (i.e. Fig. 8(a)).The process of horizontal linear scanning for Fig. 8(b) isshown in Fig. 8(c), and the result image is displayed in Fig.8(d). Meanwhile, Fig. 8(e) shows the process of vertical lin-ear scanning for Fig. 8(d), and the result image is presentedin Fig. 8(f).

In addition, for the completed image Itm (x, y), the resultimage (i.e. image Ih (x, y)) after horizontal linear scanningis shown in Fig. 9(a), and the result image (i.e. image Iv (x,y)) after vertical linear scanning is shown in Fig. 9(b). More-over, Fig. 9(b) is the final linear scanning image I ls (x, y).

Meanwhile, the edge extraction image Iee (x, y) for Ils (x,y) is shown in Fig. 9(c). Furthermore, the region image Ire(x, y) (i.e. Fig. 9(d)) is extracted according to the edge imageIee (x, y). From Fig. 9(d), we can clearly see that the regionof oil pollution interference is completely extracted.

In order to avoid the interference of noise, a median filter(here, the filter window is set as 3) is employed to filter theregion extraction image Ire (x, y). And the result image Iref(x, y) is shown in Fig. 9(e).

To extract the edge of surface defects, Canny operator13)

is employed to detect the surface defects for image Iref (x, y)

(here, the threshold value is set as 0.25). And the resultimage Ice (x, y) is shown in Fig. 9(f).

In addition, it’s important to note that the linear scanningoperation is only for defect images under oil pollution inter-ference. In other words, the linear scanning operation cannotbe used to deal with the normal images that have no oil pol-lution.

3.4. Morphology Edge ProcessingSince the extracted edge image of surface defects in Sec-

tion 3.3 includes the edge of oil pollution interference andthe edge of reflective pseudo-defect, an edge processingoperation is presented in this subsection. In order to removethe interference edge, image subtraction operation isemployed initially. Figure 10 presents the image subtractionoperation for edge extraction image Iee (x, y) and Ice (x, y).Figs. 10(a) and 10(b) show the edge extraction image Iee (x,y) and Ice (x, y) respectively, the result image using imagesubtraction operation is shown in Fig. 10(c). As it could beseen in Fig. 10(c), the oil pollution edge has not been com-pletely removed. Furthermore, the edge of reflective pseu-do-defect is still there.

In order to solve this issue mentioned above, a new imagesubtraction operation based on dilation and logical operationis proposed in this subsection. Different from the definitionof the dilation operation for gray-scale image, the edgeimage is binary image. For a binary image, mathematically,dilation is defined in terms of set operation. The dilation ofIee (x, y) by e, denoted Iee ⊕ e, is defined as

Fig. 7. The thresholding operation for a filter image. (a) The binary image using Otsu’s method. (b) The binary imageusing global thresholding method. (c) The result binary image after median filter for (b).

I x yI x y TI x y Ttf

f

,,,( ) = ( ) ≥

( ) <⎧⎨⎩

01

ifif

Fig. 8. The illustration of the linear scanning process. (a) The binaryimage Itm (x, y). (b) The partial region of image Itm (x, y). (c)The horizontal linear scanning. (d) The result image usinghorizontal linear scanning. (e) The vertical linear scanning.(f) The result image using vertical linear scanning.


2603 © 2014 ISIJ

.................... (7)

where is the reflection of set e, ∅ is the empty set and eis the structuring element.

The dilation edge image Ieed (x, y) is obtained after the dila-tion operation. In order to obtain a dilation edge image Iced (x,y), the Ieed (x, y) is performed the logical OR operation withedge image Ice (x, y). Then, the final edge image Ife (x, y) isobtained after the image subtraction operation for dilationedge image Ieed (x, y) and Iced (x, y). Figure 11 presents the

morphology edge processing for edge extraction image Iee (x,y) and Ice (x, y), here, ‘disk’ type structuring elements is usedto perform dilation operation (the radius is set as 35). Figs.11(a) and 11(b) show the dilation edge image Ieed (x, y) andIced (x, y) respectively, the final edge image Ife (x, y) is shownin Fig. 11(c). As it could be seen in Fig. 11(c), the surfacedefects are almost extracted. In addition, Fig. 11(d) shows theedge of surface defects in original defect image. In order tobetter present the edge of surface defects, the final edge imageIfe (x, y) is performed the dilation operation (the radius is set

Fig. 9. The result images using linear scanning process. (a) The result image Ih (x, y) using horizontal linear scanning. (b)The result image Iv (x, y) (i.e. the final linear scanning image Ils (x, y)) using vertical linear scanning. (c) The edgeextraction image Iee (x, y). (d) The region extraction image Ire (x, y). (e) The result image Iref (x, y) using median fil-ter. (f) The result image Ice (x, y) using Canny operator.

Fig. 10. The image subtraction operation for edge extraction image Iee (x, y) and image Ice (x, y). (a) The edge extractionimage Iee (x, y). (b) The edge extraction image Ice (x, y). (c) The result image using image subtraction operation.

Fig. 11. The morphology edge processing for edge extraction image Iee (x, y) and Ice (x, y). (a) The dilation edge image Ieed(x, y). (b) The dilation edge image Iced (x, y). (c) The final edge image Ife (x, y). (d) The edge of surface defects inoriginal defect image. (e) The result image after dilation operation for Ife (x, y). (f) The dilation edge of surfacedefects in original defect image.

I e z e Iee z ee⊕ = ( ) ≠ ∅{ }ˆ ∩

e

© 2014 ISIJ 2604


as 5), which is shown in Fig. 11(e). Moreover, Fig. 11(f)shows the dilation edge of surface defects in original image.

As a summary, the scheme of the proposed saliency linearscanning morphology method for detecting surface defectsof silicon steel strip under oil pollution interference is pre-sented in Fig. 12.

4. Experimental Results and DiscussionIn this section, we evaluate the performance of the pro-

posed method for detecting surface defects including wipe-crack-defect, scratch-defect and small-defect. Meanwhile,the proposed method is compared with other methods on thedetection of surface defects.

4.1. Implementation DetailsThe important basic parameters of the proposed method

are set as follows: the size of the Gaussian blur window Nin Section 3.1 is set as 5; the open-close filtering in Section3.2 is performed with a series of ‘disk’ type structuring ele-ments (The radius range of alternating sequential filtering isfrom 2 to 3); the threshold value in thresholding operationis set as 0.05, furthermore, the median filter window in Sec-tion 3.2 is set as 29; the median filter window in Section 3.3is set as 3; the threshold value of Canny operator in Section

3.3 is set as 0.25; the ‘disk’ type structuring elements is usedto perform dilation operation in Section 3.4, furthermore,the radius is set as 35.

Moreover, the morphological gradient16) and Cannyoperator13) are compared with the propose method fordetecting surface defects. Furthermore, the code of thesemethods is run in Matlab 7.10 (R2010a) software on Pentium(R) Dual-Core machine with 2.8 GHz and 4 GB of memoryand the Windows XP operating system. In addition, the democode of the proposed SLSM can be down load from ourhomepage: http://faculty.neu.edu.cn/yunhyan/SLSM.html.

4.2. Wipe-crack-defect DetectionWipe-crack-defect is one of the most common defects

types in surface defect of silicon steel strip. In view of theoil pollution have different shape types, the proposed SLSMand other methods are performed in different sample images.First of all, for a new oil pollution shape type of wipe-crack-defect, the main instruction images of proposed method areshown in Fig. 13. As it could be seen in Fig. 13(f ), the sur-face defects are almost extracted using the proposed SLSMfor a new shape type of wipe-crack-defect.

In order to further evaluate the performance of the proposedSLSM method for other shape types of wipe-crack-defect, sixdifferent sample images of wipe-crack-defect are employed to

Fig. 12. The scheme of the proposed saliency linear scanning morphology method.

Fig. 13. The main instruction images of proposed method for wipe-crack-defect. (a) The original image of wipe-crack-defect. (b) The saliency map of wipe-crack-defect. (c) The binary image after thresholding operation. (d) Theregion extraction image of wipe-crack-defect. (e) The final edge image of wipe-crack-defect. (f) The dilationedge of wipe-crack-defect in original image.


2605 © 2014 ISIJ

compare the performance of different methods. Figure 14shows the result images of the different methods for wipe-crack-defect. As is shown in Figs. 14(b) and 14(c), althoughthe morphological gradient and Canny operator detect severalwipe-crack-defects in the cluttered background, too muchinterference edges are also detected. On the contrary, in Fig.14(d), the proposed SLSM not only detects all of the surfacedefects, but also removes the interference edges.

4.3. Scratch-defect DetectionScratch-defect is another common defect type in surface

defect of silicon steel strip. The shape of the scratch-defectis like a strip line. Similar to the wipe-crack-defect, for anew oil pollution shape of scratch-defect, the main instruc-tion images of proposed method are shown in Fig. 15. Asexpected, in Fig. 15(f), the surface scratch-defects are almostextracted using the proposed SLSM method even in the

Fig. 14. Result images of the different methods for wipe-crack-defect. (a) The original image of wipe-crack-defect. (b)The result images of morphological gradient. (c) The result images of Canny operator. (d) The result images ofSLSM method.

Fig. 15. The main instruction images of proposed method for scratch-defect. (a) The original image of scratch-defect. (b) Thesaliency map of scratch-defect. (c) The binary image after thresholding operation. (d) The region extraction image ofscratch-defect. (e) The final edge image of scratch-defect. (f) The dilation edge of scratch-defect in original image.

© 2014 ISIJ 2606


interference of reflective pseudo-defects.In addition, to further evaluate the performance of the pro-

posed SLSM method for other shape types of scratch-defect,six different sample images of scratch-defect are employed tocompare the performance of different methods. Figure 16shows the result images of the different methods for scratch-defect. As is shown in Fig. 16(a), all of the original imagesof scratch-defect suffer from the interference of reflectivepseudo-defects. From Figs. 16(b) and 16(c), we can see that

too much interference edges of reflective pseudo-defects areextracted, despite the morphological gradient and Cannyoperator detect several scratch-defects in the cluttered back-ground. However, in Fig. 16(d), the proposed SLSM not onlydetects all of the surface scratch-defects, but also removes theinterference edges of reflective pseudo-defects.

4.4. Small-defect DetectionExcept for the wipe-crack-defect and scratch-defect,

Fig. 16. Result images of the different methods for scratch-defect. (a) The original image of scratch-defect. (b) The resultimages of morphological gradient. (c) The result images of Canny operator. (d) The result images of SLSM method.

Fig. 17. The main instruction images of proposed method for small-defect. (a) The original image of small-defect. (b) Thesaliency map of small-defect. (c) The binary image after thresholding operation. (d) The region extraction imageof small-defect. (e) The final edge image of small-defect. (f) The dilation edge of small-defect in original image.


2607 © 2014 ISIJ

some small-defects are also common defects in surfacedefect of silicon steel strip. For example, Fig. 17(f) showsa small defect sample image of rust point. The final edgeimage of rust point using the proposed SLSM method is pre-sented in Fig. 17(e). From Fig. 17(e), we can clearly see thatthe edge of rust point is completely displayed on the originalsmall-defect image.

In addition, four different sample images of small-defectare employed to compare the performance of different meth-ods. Figure 18 shows the result images of the differentmethods for small-defect. As expected, in Fig. 18(d), theproposed SLSM detects all of the surface defects. Therefore,these good results in Fig. 18 further confirm the perfor-mance of the proposed SLSM.

5. ConclusionIn this research, a new detection method based on salien-

cy linear scanning morphology is proposed to detect theinteresting defect objects for silicon steel strip under oil pol-lution interference. Since the background of defect image isquite complex, the visual saliency extraction is employed tosuppress the clutter background. In view of the extractedsaliency map contains some interference information, theopen-close filtering is performed. Then, the global thresh-olding method is employed to obtain a binary image. For thebinary image, the linear scanning operation is proposed toobtain the region of oil pollution. Meanwhile, the morphol-ogy edge processing is introduced to remove the edge of oilpollution interference and the edge of reflective pseudo-defect. Moreover, three typical surface defect of silicon steelstrip, i.e., wipe-crack-defect, scratch-defect and small-defect, are used to evaluate the performance of the proposedmethod. Experimental results show that the proposed methodis effective and suitable for detecting the surface defectsunder oil pollution interference. However, it should be notethat the linear scanning operation is only for defect images

under oil pollution interference. In other words, the linearscanning operation cannot be used to deal with the normalimages that have no oil pollution.

Future perspectives of this work include: extension ofdefect sample and classification of the surface defect.

AcknowledgmentThis work is supported by the National Natural Science

Foundation of China (51374063) and the FundamentalResearch Funds for the Central Universities (N120603003,N13081001).

REFERENCES

1) Alaknanda, R. S. Anand and P. Kumar: NDT&E Int., 42 (2009), 2.2) J. Shao, D. Du, B. Chang and H. Shi: NDT&E Int., 46 (2012), 14.3) X. Li, H. Jiang and G. Yin: NDT&E Int., 62 (2014), 6.4) J. P. Yun, S. H. Choi, J. W. Kim and S. W. Kim: NDT&E Int., 42

(2009), 389.5) E. Pan, L. Ye, J. J. Shi and T. S. Chang: J. Manuf. Sci. E-T ASME,

131 (2009), 061008.6) A. Landström and M. Thurley: IEEE J. STSP, 6 (2012), 866.7) F. G. Bulnes, R. Usamentiaga, D. F. García and J. Molleda: Sensors,

12 (2012), 10788.8) R. Medina, F. Gayubo, L. M. González-Rodrigo, D. Olmedo, J.

Gómez-García-Bermejo, E. Zalama and J. R. Perán: Int. J. Adv.Manuf. Tech., 57 (2011), 1087.

9) S. Ghorai, A. Mukherjee, M. Gangadaran and P. K. Dutta: IEEE TInstrum. Meas., 62(2013), 612.

10) K. Xu, Y. Ai and X. Wu: Int. J. Min. Met. Mater., 20 (2013), 37.11) F. Pernkopf: Pattern Anal. Appl., 7 (2004), 333.12) K. Agarwal, R. Shivpuri, Y. Zhu, T. S. Chang and H. Huang: Expert.

Syst. Appl., 38 (2011), 7251.13) J. Canny: IEEE T Pattern Anal., 6 (1986), 679.14) R. Achanta, S. Hemami, F. Estrada and S. Susstrunk: Proc. of 27th

IEEE Conf. on Computer Vision and Pattern Recognition, IEEE, NewYork, (2009), 1597.

15) R. Achanta and S. Susstrunk: Proc. of 17th IEEE Int. Conf. on ImageProcessing, IEEE, New York, (2010), 2653.

16) E. Dougherty and R. Lotufo: Hands-on Morphological ImageProcessing, SPIE Press, Bellingham, (2003), 1.

17) N. Otsu: IEEE Trans. Syst., Man, Cybern., Syst., 9 (1979), 62.

Fig. 18. Result images of the different methods for small-defect. (a) The original image of small-defect. (b) The resultimages of morphological gradient. (c) The result images of Canny operator. (d) The result images of SLSMmethod.

surface defect detection method using ... - j-stage home

Documents