- Email: [email protected]
An infrared thermal image processing framework based on superpixel algorithm to detect cracks on metal surface
Infrared Physics & Technology 67 (2014) 266–272
Contents lists available at ScienceDirect
Infrared Physics & Technology journal homepage: www.elsevier.com/locate/infrared
An infrared thermal image processing framework based on superpixel algorithm to detect cracks on metal surface Changhang Xu a, Jing Xie a,b,⇑, Guoming Chen a, Weiping Huang b a b
Department of Safety Science and Engineering, College of Mechanical and Electronic Engineering, China University of Petroleum (East China), Qingdao 266580, China College of Engineering, Ocean University of China, Qingdao 266071, China
h i g h l i g h t s A proper superpixel algorithm was selected for infrared thermal image processing. Proper texture features of superpixels were selected for clustering. A new infrared thermal image processing frame was proposed to detect cracks automatically. Experiments were implemented to verify the effectiveness of the image processing frame.
a r t i c l e
i n f o
Article history: Received 3 April 2014 Available online 10 August 2014 Keywords: Infrared thermography Defect detection Superpixel algorithm Surface crack Image segmentation
a b s t r a c t Infrared thermography has been used increasingly as an effective non-destructive technique to detect cracks on metal surface. Due to many factors, infrared thermal image has low definition compared to visible image. The contrasts between cracks and sound areas in different thermal image frames of a specimen vary greatly with the recorded time. An accurate detection can only be obtained by glancing over the whole thermal video, which is a laborious work. Moreover, experience of the operator has a great important influence on the accuracy of detection result. In this paper, an infrared thermal image processing framework based on superpixel algorithm is proposed to accomplish crack detection automatically. Two popular superpixel algorithms are compared and one of them is selected to generate superpixels in this application. Combined features of superpixels were selected from both the raw gray level image and the high-pass filtered image. Fuzzy c-means clustering is used to cluster superpixels in order to segment infrared thermal image. Experimental results show that the proposed framework can recognize cracks on metal surface through infrared thermal image automatically. Ó 2014 Elsevier B.V. All rights reserved.
1. Introduction Infrared (IR) thermography is a non-destructive and non-contact technique which is widely applied in predictive and preventive maintenance programs [1–4]. In this technique, there are two ways to get the IR thermal image, the active and passive heating [5]. When IR thermography is applied to detect cracks on metal surface, the former heating way is used widely [6–9]. When a metallic part with surface cracks is heated, the heat will diffuse differently through cracks and sound area. Cracks will prevent the heat
⇑ Corresponding author at: Department of safety science and engineering, College of mechanical and electronic engineering in China University of Petroleum (East China), Changjiang West Road 66#, Economic Development Zone, Qingdao, Shandong province, China. Tel.: +86 0532 86983071. E-mail addresses: [email protected] (C. Xu), [email protected] (J. Xie), [email protected] (G. Chen), [email protected] (W. Huang). http://dx.doi.org/10.1016/j.infrared.2014.08.002 1350-4495/Ó 2014 Elsevier B.V. All rights reserved.
diffusion and lead to a heat convergence. Consequently, the area around cracks will show higher temperature than sound area. This abnormal temperature distribution can be captured using an IR thermal imager. Therefore, cracks can be recognized by finding out the abnormal temperature area in the IR thermal image by an engineer. However, one limitation of the technique is that to determinate the location and size of cracks by glancing over the whole thermal video is laborious. Moreover, the inspection result is easily affected by the engineer. Consequently, to construct a proper image processing framework to pick out crack areas from sound ones is needed eagerly to realize the automatic crack detection using IR thermography. Using image processing technique, picking out cracks from sound areas can be accomplished by segmenting the image into crack areas and sound ones. Many image segmentation algorithms have been widely used in image processing field, such as mean shift [10], normalized cuts [11], watersheds [12], graph-based
267
C. Xu et al. / Infrared Physics & Technology 67 (2014) 266–272
(a)
(b)
(c) Fig. 1. Experimental equipment.
Table 1 The characters of SAY-HY6850 IR. Measurement range
10 °C–600 °C (standard) 40 °C–2000 °C (extensions)
Sensitivity Accuracy Infrared image resolution Spatial resolution Field angle
0.08 °C ±2% 320 240 1.3 Mrad 24° 18°
(d)
(e) [13] and so on. All these classic segmentation algorithms process an image as gather of pixels. Ren X proposed another method for image segmentation named superpixel algorithm [14]. Using existing segmentation method, superpixel algorithms over-segment an image into perceptually meaningful atomic regions which can be used to replace the rigid structure of the pixel grid. Superpixels can capture image redundancy and greatly reduce the complexity of subsequent image processing tasks. Moreover, superpixel algorithms can supply additional information, such as mean and variance of gray levels in a superpixel, to subsequent segmentation algorithm. We applied the superpixel algorithm to the IR thermal image segmentation in order to improve the accuracy of detection and speed up the detection process. Many segmentation methods have been used to generate superpixels. The Normalized Cuts algorithm was used to generate superpixels in [13]. AchantaIn et al. proposed a superpixel algorithm named simple linear iterative clustering (SLIC) based on K-means in [15]. David Martin and Charless Fowlkes’ boundary detector was used in [16,17]. In this paper, the watersheds algorithm was compared to SLIC which is the latest popular superpixel algorithm and the former one was selected to generate superpixels in this application.
(f) Fig. 3. IR thermal frames (a)–(c) are frames obtained at 100 ms, 200 ms, 250 ms from specimen 1 and (d)–(f) are frames obtained at 100 ms 160 ms 200 ms from specimen 2.
After over-segmenting one thermal image into superpixels, the next work is to cluster the similar superpixels and obtain the final segmentation of the thermal image. Fuzzy c-means (FCM) clustering algorithm in [18] was used to cluster superpixels. We chose proper features of superpixels as the input of FCM algorithm. Compared with visual image, IR thermal image has low signal to noise ratio due to a great deal of noise coming from electronic noise, photon noise, aberrations of the optical system, environment temperature disturbance, and some other random factors [19]. When only using the superpixels’ features in raw gray level image, the segmentation result is found to be very poor. Because much noise can be eliminated by using a high-pass filter, the combined features of superpixels in both raw gray level image and high-pass filtered image were selected as the input of the clustering algorithm. We recognized the crack cluster as the one with biggest mean gray level. Finally, tiny noise still existing was removed using morphologic operator. We validated our frame by conducting experiment on test specimens with surface cracks manufactured by wire electrical discharge machining. The accurate information of cracks was automatically determined in a short time using the proposed thermal image processing framework. 2. Experimental procedure
Fig. 2. Specimens with cracks.
The experimental setup shown in Fig. 1 was made of a computer, an inductor, an IR thermal imager and two specimens. The
268
C. Xu et al. / Infrared Physics & Technology 67 (2014) 266–272
Fig. 4. Flowchart of the proposed thermal image processing framework.
(a)
(a)
(b) (b) Fig. 6. Superpixels generated by Watershed from (a) the original gray level image and (b) the corresponding gradient image.
(c) Fig. 5. Superpixels generated by SLIC with different desired number of superpixel and compactness factor: (a) 100 and 10 (b) 200 and 10 (c) 1599 and 10.
inductor generates the eddy current in specimens. It can supply repeatable, reliable heating with precise power control. The SATHY6850 IR thermal imager was used to capture the IR thermal video of the specimens. Its main characters are listed in Table 1.
The two specimens are shown in Fig. 2. They are made of C45E4 steel and the sizes of them are 90 30 6.5 mm and 296 25 8 mm respectively. On the specimens, there are a few of surface cracks manufactured using a wire cutting machine which can assure the width of them to be about 0.5 mm. The depths of cracks are about 4 mm. The specimens are clamped by a fixture and the imager is put on a tripod. In this way, the distance between the specimens and
Table 2 Mean gray level of each cluster. Frame in Fig. 3
Cluster 1
Cluster 2
Cluster 3
Cluster 4
Cluster 5
Cluster 6
(a) (b) (c) (d) (e) (f)
0.5216 0.5616 0.0857 0.7597 0.1384 0.7644
0.0909 0.0569 0.5017 0.7147 0.2607 0.4652
0.5645 0.0796 0.6853 0.8509 0.5342 0.7435
0.5536 0.5234 0.5355 0.3076 0.6948 0.8745
0.8799 0.8615 0.0741 0.5394 0.7812 0.5387
0.1333 0.4573 0.5135 0.3633 0.3322 0.2943
C. Xu et al. / Infrared Physics & Technology 67 (2014) 266–272
(a) raw IR image
(b) k-means segmentation result
(c) threshold segmentation result
(d) result of the proposed segmentation method
(e) final detection result of the proposed framework Fig. 7. Result of proposed method and some other segmentation algorithm.
the lens of imager can be fixed accurately. The specimens were excited using the inductor for 200 ms, and then the imager recorded the whole heating process of 3 s. 150 frames were captured during this process for each specimen. 3. Proposed method In order to relieve the operator from the laborious, time-consuming work, an automatic thermal image processing framework was proposed. The framework mainly includes four steps: superpixels generation, image segmentation, crack recognition and postprocessing. It works in sequence as shown in Fig. 4. We began by choosing a proper algorithm to generate superpixels for the application of crack detection in Section 3.1. Features of superpixels for the segmentation using FCM algorithm were selected in Section 3.2. In Section 3.3, we introduce a method to recognize the crack clusters from sound clusters. Morphologic operator was applied to eliminate some tiny interference in the recognized crack cluster. 3.1. Superpixels generation Due to the continuity of heat diffusion, the thermal features of a small area in IR thermal image are always approximately uniform.
269
Hence, superpixels are more proper basic image elements to be processed than individual image pixels. Superpixels can give more statistic information of a small area, such as mean and variance. The additional information will be favorable to the accuracy of segmentation and the accuracy of the detection can be improved. Using superpixels other than individual pixels as proposed element will reduce the complexity of subsequent image segmentation task and the detection process can be accomplished in less time. In superpixel algorithm, using an over-segmentation method, the given IR image is partitioned into sub-regions, each of which is called a superpixel. In this way, the problem of segmenting the image is cast into clustering the superpixels into groups. In order to simplify the image process, the raw IR image is inverted to a gray-level image firstly. There are many approaches to generate superpixels, each with its own advantages and drawbacks that may be better suited to a particular application. The generation speed, ability to adhere to image boundaries, and impact on segmentation performance are general characters to evaluate a superpixel algorithm. For the crack detection application, the adherence to the crack boundaries is the most important requirement for superpixel algorithm. SLIC has become a popular superpixel algorithm since it was proposed [15]. It used K-means to over-segment a color image in the 5d space of color information and image location. As the clustering method is simpler, it is very efficient. In SLIC algorithm, two parameters have to be determined at first, the desired number of superpixel and the compactness factor. The superpixels of Fig. 3(c) generated by the SLIC with three sets of these two parameters are shown Fig. 5. We can see that the two parameters determined by operator have a great effect on the result of superpixels generation. If the number of superpixel is limited too small, the generated superpixels will lose the boundary information of the cracks. Hence, it is important to find proper parameters for a satisfying detection result from SLIC method. Another popular segmentation algorithm, watershed can also be used to generate superpixels. Watershed is usually utilized on two kinds of images, the original image and the corresponding gradient image. Their respective results of Fig. 3(c) are shown in Fig. 6. From the results, we can see that the watershed segmentation of the original image obtains superpixels with too large size and therefor lost detailed information of cracks. The result of watershed on the gradient image (Fig. 6(b)) has a better adherence to cracks boundaries. Hence, the gradient image of IR thermal image is more proper for watershed algorithm.
(a) raw IR frame of specimen 1 (b) FCM clustering on gray level image (c) FCM clustering on the gray level image and high-pass filtered one (d) FCM clustering on superpixels using gray level information (e) segmentation result using the proposed algorithm (f) the final detection result using the proposed frame Fig. 8. Segmentation result of the proposed method and other FCM ones.
270
C. Xu et al. / Infrared Physics & Technology 67 (2014) 266–272
Fig. 9. Detection results using the proposed framework. The left column images are raw infrared thermal images. The right column images are detection results using the proposed framework.
The number of superpixels in Fig. 6(b) is 1599. Compared to the SLIC superpixels with the same number (Fig. 5(c)), the watershed superpixels have a little better adherence to cracks boundaries. Especially for an automatic detection system, the watershed has a superior performance that it avoids the determination of parameters as in the SLIC. In consideration of this important virtue and its equally satisfying result, the watershed algorithm is chosen to over-segment image and generate superpixels in our image process framework. 3.2. Thermal image segmentation After superpixel generation, the image is constructed from superpixels other than original individual pixels. Then, the image should be segmented into a few parts which compose of similar superpixels. We use the FCM clustering algorithm to accomplish the segmentation. The standard FCM objective function for partitioning X ¼ fxk jk ¼ 1; 2; . . . ; ng into c clusters is given by [18]
J¼
c X n X 2 um ik kxk v i k
ð1Þ
i¼1 k¼1
where uik is the membership function charactering the possibility of xk belonging to the i-th cluster. The parameter m is the weighted exponent that controls the fuzziness of the resulting partition and vi is the center of the i-th cluster. The clustering result is obtained by minimizing the objective function. This minimization can be accomplished in an iteration way. For the image segmentation based on superpixels, the data xk points to features of the k-th superpixel and the FCM algorithm will assign each superpixel to a cluster. To select features of superpixels properly is important to get a satisfactory segmentation result. Since each superpixel is a small image patch, the texture characters can be used as features of it. There are many popular texture features, such as mean, variance and entropy. The first three ones represent average brightness, average contrast and random degree respectively. In this
application, the mean and variance of a superpixel denoted by lraw and rraw are chosen as clustering features. When only considering the above two features of superpixels, the detection result is usually very poor for some thermal image frames. The reason lies in the low contrast between cracks and sound areas in these frames. This contrast depends on the capture time of the frame, environment conditions and crack dimensions. To improve the detection performance for some IR image with poor contrast between crack and sound area, some other characters of superpixels have to be added into the segmentation features. High-pass filter can sharpen edges in an image and make the changes of gray level in the image more distinct. Therefore, in the high-pass filtered image, the boundary between blurred cracks and around sound area should be sharpened. There is more information about these cracks in it than in the original gray level image. With observation and analysis, we used superpixels’ mean and variance in the high-pass filtered image denoted by lhpass and rhpass as additional features for clustering. After feature selection, the selected four features were normalized from 0 to 1 firstly. This normalization operation can eliminate the negative effect from different feature scope. Due to different importance of the four features for the superpixel clustering, different weight coefficients should be multiplied to the features respectively to adjust effect of them on the segmentation result. Depending on number of trial results, the weight coefficients for lraw, rraw, lhpass, rhpass were selected as 0.43, 0.07, 0.43 and 0.07 respectively. When the four weighted features of each superpixel have been computed, they will be substituted into the FCM algorithm as xk in Eq. (1). In the objective function of Eq. (1), two parameters have to be determined, the weighted exponent m and the number of clusters c. The weighted exponent is usually chosen as 2 in most applications and the same as in this paper. The cluster number depends on specific application. In this paper, the cluster number is specified as 6 by comparing the clustering results of different values of it. After all parameters and features are determined, the object function will be minimized using iteration method. When the object function arrives the minimum value, the k-th superpixel will clustered into the cluster with largest uik (i = 1, 2, . . .,6). Then each superpixel will be pointed to a specific cluster and the IR thermal image will be segmented into six parts. 3.3. Recognition of cracks cluster Among segmentation clusters, one represents cracks. In order to propose a fully automatic crack detection system, the crack cluster should be picked out from other clusters without manual intervention. Due to the obstruction of heat from cracks and heat accumulation around crack area, the mean temperature of crack area should be higher than surrounding sound surface. In the raw gray level image, there are some other areas have approximately high temperature as crack areas, which results in poor result using threshold segmentation. After clustering, most of these interferences have been disposed and the crack cluster always has the maximum mean lightness corresponding to the highest temperature. As a result, the mean gray level of each cluster was taken as a feature to recognize the crack cluster. The cluster which has maximum mean gray level in the corresponding raw gray image represents cracks. Table 2 shows the mean gray level of each cluster in the segmentation of Fig. 3(c) and the 3rd cluster will be recognized as cracks. 3.4. Postprocessing Between superpixels which generated using watershed algorithm, there exist boundaries with a single pixel width. Therefore,
271
C. Xu et al. / Infrared Physics & Technology 67 (2014) 266–272 Table 3 Crack covering measure. Method
Proposed 1a 2b 3c 4d 5e a b c d e
Specimen 1
Specimen 2
100 ms
200 ms
250 ms
Mean
100 ms
160 ms
200 ms
Mean
0.7183 0.5009 0.5691 0.5009 0.5884 0.2071
0.6369 0.4421 0.3212 0.4402 0.5806 0.2540
0.5158 0.1959 0.2894 0.1987 0.3024 0.1908
0.6237 0.3796 0.3932 0.3799 0.4905 0.2173
0.4936 0.2289 0.1609 0.1985 0.3846 0.1801
0.4560 0.2474 0.2233 0.2355 0.4152 0.1694
0.4347 0.0864 0.1622 0.1952 0.3714 0.1789
0.4614 0.1876 0.1821 0.2097 0.3904 0.1761
k-mean. Threshold. FCM clustering on gray level image. FCM clustering on the gray level image and high-pass filtered one. FCM clustering on superpixels only using gray level information.
the recognized crack cluster has been separated by the ridge lines. To get complete crack information, the separated superpixels representing one crack should be linked together. Closing is a proper morphologic operation for this task. It consists of dilation followed by erosion with the same structuring element. Before applying closing operation, image area only including crack superpixels should be transformed to a binary one. Because the boundary has only one pixel width, disk has radius of one pixel is selected as the structure element. The closing operation can eliminate the bays along the border area. Therefore, the bay of one pixel width between crack superpixels was filled and one crack became an entirety. Due to some remaining tiny noise in the recognized crack cluster, after closing operation, these noises were shown with cracks simultaneously. In order to get more accurate crack information, these noises should be eliminated in the final detection result. Tiny noises and cracks can be labeled as connected regions respectively. Because the noisy area always has very small size compared to cracks, the pixels number which one connected region includes was selected as a criteria to decide whether the connected region is a crack or not. The region with fewer pixels than the specified threshold will be disappeared in the final crack detection result. The specified threshold determines the size of crack that can be detected to some degree. In this paper, it is specified as 30.
low contrast. For the left frame, satisfying detection result can be obtained by using both the former two comparing methods and our proposed one, except FCM clustering of superpixel only using gray level information. However, for the right frame, only the proposed method can get the information of right crack, which has low contrast to sound area. When using the first comparing method, the left crack has been detected with expanded boundary and the right one has not been detected. When the second comparing method was used, even though the boundary of the left crack is detected more accurately, the right crack has not been detected yet. The third comparing method obtains the similar result as the second one. But with the proposed method, two cracks are clustered together. The final detect result shows that two cracks are detected on the right location using the proposed framework, although the dimension of crack which cannot be detected by other methods has some bias from the real size. The final detection results of all the six typical frames using the proposed framework are shown in Fig. 9. From it, we can see that all cracks can be detected using the proposed framework even in frames with poor quality. In order to evaluate the performance of the proposed framework quantitatively, a quantity measure is needed to evaluate the crack detection accuracy. In image processing research, a segmentation result is popularly evaluated using an indicator named covering measure [20]
4. Results and discussion
C¼
1X jsi \ g i j jg jmax jLj g 2G i si 2S jsi [ g i j
ð2Þ
i
To show the superiority of the proposed method to other classic segmentation algorithm, the proposed method, k-means and threshold segmentation were respectively applied to the typical thermal image shown in Fig. 3(f). The segmentation results of the image using the above three algorithms are shown in Fig. 7. From the result, we can see that k-means cannot detect the right crack which has low contrast to sound areas. Threshold segmentation also got an unvalued result with a great deal of noise. However, using the proposed method, three cracks have been all detected. To verify the validity of the proposed thermal image processing framework further, three other segmentation results of FCM clustering on raw gray level image, FCM clustering on raw gray level image and high-pass filtered image, FCM clustering on superpixels of the raw gray level image were compared to that of the proposed method. We picked out two representative frames of specimen 1 (Fig. 3(a) and (c)). In the former frame two cracks are both clear while one crack is blurry in the latter one. Final detection results are listed in Fig. 8. We found that for the left frame in the first row, the contrast between the two cracks and sound area is high. Meanwhile, in the right frame, only one crack has high contrast to sound area and the other crack is very blurry because of the
where G represents the ground truth segmentation which is annotated manually in this paper. S is segmentation to be evaluated. L is the original image to be segmented. || denotes the pixels number of area. jsjsi \[ ggi jj is the overlap measure between si and gi. However, for i i the crack detection application, whether cracks have been segment into one cluster is the key criteria to evaluate the segmentation result. As a result, the covering measure is modified to be crack covering measure in this application,
C¼
jd \ gj jd [ gj
ð3Þ
where g is the ground truth of cracks and d is the detected cracks. When the cracks are detected with accurate size, the crack covering measure will just be 1. In the case that cracks were detected with dilated size or with shrunk size, the crack covering measure will all be less than 1. The measure will be decreased with the increasing of difference between g and d. Therefore, the crack covering measure can evaluate the detection result accurately. In our experiment, two specimens were tested. For each specimen, 150 frames were captured using IR camera and the crack
272
C. Xu et al. / Infrared Physics & Technology 67 (2014) 266–272
covering measure of six typical frames were listed in Table 3. The proposed method gets the maximum crack covering measure for each frame compared with other methods. Using the proposed framework, all cracks have been detected and the average crack covering measure over the typical frames gets the maximum value for each specimen. Therefore, the proposed image process framework has satisfying performance and can detect cracks even in thermal image with poor quality.
5. Conclusions We showed that an automatic system to detect cracks on a metal part is an eager demand to widen the application of IR thermography. A superpixel based image processing framework has been proposed to accomplish this task. Watershed algorithm is chosen to generate superpixels due to its high adherence to the crack boundaries. Combined superpixels’ features from both raw gray level image and high-pass filtered image were selected as clustering features of the FCM algorithm considering the interference coming from various aspects. Compared with some other segmentation methods, the proposed framework can detect crack which has low contrast with sound area. It can weaken the sensitivity of the detection result to the frame quality and realize a whole automatic detection. The location and shape of cracks can be determined accurately by the proposed IR thermal image processing frame. In future work we plan to explore the applicability of the proposed framework to other materials, such as fibre reinforced composite.
Conflict of interest The authors declare that they have no conflict of interests.
Acknowledgements This work was financially supported by Shandong Provincial Natural Science Foundation of China (No. ZR2013EEQ028) and the Fundamental Research Funds for the Central Universities of China (No. 14CX02077A).
References [1] M. Omar, M. Hassan, K. Donohue, K. Saito, R. Alloo, Infrared thermography for inspecting the adhesion integrity of plastic welded joints, NDT E Int. 39 (2006) 1–7. [2] M.S. Jadin, S. Taib, Recent progress in diagnosing the reliability of electrical equipment by using infrared thermography, Infrared Phys. Technol. 55 (2012) 236–245. [3] S. Bagavathiappan, B.B. Lahiri, T. Saravanan, J. Philip, T. Jayakumar, Infrared thermography for condition monitoring – a review, Infrared Phys. Technol. 60 (2013) 35–55. [4] B. Yang, Y. Huang, L. Cheng, Defect detection and evaluation of ultrasonic infrared thermography for aerospace CFRP composites, Infrared Phys. Technol. 60 (2013) 166–173. [5] Maldague, P.V. Xavier, Introduction to NDT by active infrared thermography, Mater. Eval. 60 (2002) 1060–1073. [6] B. Oswald-Tranta, Thermo-inductive crack detection, Nondestruct. Test. Eval. 22 (2007) 137–153. [7] D. Wagner, N. Ranc, C. Bathias, P.C. Paris, Fatigue crack initiation detection by an infrared thermography method, Fatigue Fract. Eng. Mater. Struct. 33 (2009) 12–21. [8] H. Park, M. Choi, J. Park, W. Kim, A study on detection of micro-cracks in the dissimilar metal weld through ultrasound infrared thermography, Infrared Phys. Technol. 62 (2014) 124–131. [9] L.D. Favro, R.L. Thomas, X. Han, Sonic infrared imaging of fatigue cracks, Int. J. Fatigue 23 (2001) 471–476. [10] Y. Cheng, Mean shift, mode seeking, and clustering, IEEE Trans. Pattern. Anal. 17 (1995) 790–799. [11] J. Shi, J. Malik, Normalized cuts and image segmentation, in: Proceedings, 1997 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 1997 pp. 731–737. [12] L. Vincent, P. Soille, Watersheds in digital spaces: an efficient algorithm based on immersion simulations, IEEE Trans. Pattern. Anal. 13 (1991) 583–598. [13] P. Felzenszwalb, D. Huttenlocher, Efficient graph-based image segmentation, Comput. Vision 59 (2004) 167–181. [14] R. Xiaofeng, J. Malik, Learning a classification model for segmentation, in: Proceedings Ninth IEEE International Conference Computer Vision, vol. 11, 2003, pp. 10–17. [15] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, S. Susstrunk, SLIC superpixels compared to state-of-the-art superpixel methods, IEEE Trans. Pattern. Anal. 34 (2012) 2274–2282. [16] Mori, Greg, Recovering human body configurations: combining segmentation and recognition, in: Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, vol. 2, 2004, pp. II326–II333. [17] G. Mori, Guiding model search using segmentation, in: ICCV 2005. Tenth IEEE International Conference on Computer Vision, vol. 1412, 2005, pp. 1417–1423. [18] J.C. Bezdek, S.K. Pal, Fuzzy Models for Pattern Recognition, Piscataway, IEEE Press, NJ, 1991. [19] V. Vavilov, Noise-limited thermal/infrared nondestructive testing, NDT E Int. 61 (2014) 16–23. [20] P. Arbelaez, M. Maire, C. Fowlkes, J. Malik, Contour detection and hierarchical image segmentation, IEEE Trans. Pattern. Anal. 33 (2011) 898–916.
Contents lists available at ScienceDirect
Infrared Physics & Technology journal homepage: www.elsevier.com/locate/infrared
An infrared thermal image processing framework based on superpixel algorithm to detect cracks on metal surface Changhang Xu a, Jing Xie a,b,⇑, Guoming Chen a, Weiping Huang b a b
Department of Safety Science and Engineering, College of Mechanical and Electronic Engineering, China University of Petroleum (East China), Qingdao 266580, China College of Engineering, Ocean University of China, Qingdao 266071, China
h i g h l i g h t s A proper superpixel algorithm was selected for infrared thermal image processing. Proper texture features of superpixels were selected for clustering. A new infrared thermal image processing frame was proposed to detect cracks automatically. Experiments were implemented to verify the effectiveness of the image processing frame.
a r t i c l e
i n f o
Article history: Received 3 April 2014 Available online 10 August 2014 Keywords: Infrared thermography Defect detection Superpixel algorithm Surface crack Image segmentation
a b s t r a c t Infrared thermography has been used increasingly as an effective non-destructive technique to detect cracks on metal surface. Due to many factors, infrared thermal image has low definition compared to visible image. The contrasts between cracks and sound areas in different thermal image frames of a specimen vary greatly with the recorded time. An accurate detection can only be obtained by glancing over the whole thermal video, which is a laborious work. Moreover, experience of the operator has a great important influence on the accuracy of detection result. In this paper, an infrared thermal image processing framework based on superpixel algorithm is proposed to accomplish crack detection automatically. Two popular superpixel algorithms are compared and one of them is selected to generate superpixels in this application. Combined features of superpixels were selected from both the raw gray level image and the high-pass filtered image. Fuzzy c-means clustering is used to cluster superpixels in order to segment infrared thermal image. Experimental results show that the proposed framework can recognize cracks on metal surface through infrared thermal image automatically. Ó 2014 Elsevier B.V. All rights reserved.
1. Introduction Infrared (IR) thermography is a non-destructive and non-contact technique which is widely applied in predictive and preventive maintenance programs [1–4]. In this technique, there are two ways to get the IR thermal image, the active and passive heating [5]. When IR thermography is applied to detect cracks on metal surface, the former heating way is used widely [6–9]. When a metallic part with surface cracks is heated, the heat will diffuse differently through cracks and sound area. Cracks will prevent the heat
⇑ Corresponding author at: Department of safety science and engineering, College of mechanical and electronic engineering in China University of Petroleum (East China), Changjiang West Road 66#, Economic Development Zone, Qingdao, Shandong province, China. Tel.: +86 0532 86983071. E-mail addresses: [email protected] (C. Xu), [email protected] (J. Xie), [email protected] (G. Chen), [email protected] (W. Huang). http://dx.doi.org/10.1016/j.infrared.2014.08.002 1350-4495/Ó 2014 Elsevier B.V. All rights reserved.
diffusion and lead to a heat convergence. Consequently, the area around cracks will show higher temperature than sound area. This abnormal temperature distribution can be captured using an IR thermal imager. Therefore, cracks can be recognized by finding out the abnormal temperature area in the IR thermal image by an engineer. However, one limitation of the technique is that to determinate the location and size of cracks by glancing over the whole thermal video is laborious. Moreover, the inspection result is easily affected by the engineer. Consequently, to construct a proper image processing framework to pick out crack areas from sound ones is needed eagerly to realize the automatic crack detection using IR thermography. Using image processing technique, picking out cracks from sound areas can be accomplished by segmenting the image into crack areas and sound ones. Many image segmentation algorithms have been widely used in image processing field, such as mean shift [10], normalized cuts [11], watersheds [12], graph-based
267
C. Xu et al. / Infrared Physics & Technology 67 (2014) 266–272
(a)
(b)
(c) Fig. 1. Experimental equipment.
Table 1 The characters of SAY-HY6850 IR. Measurement range
10 °C–600 °C (standard) 40 °C–2000 °C (extensions)
Sensitivity Accuracy Infrared image resolution Spatial resolution Field angle
0.08 °C ±2% 320 240 1.3 Mrad 24° 18°
(d)
(e) [13] and so on. All these classic segmentation algorithms process an image as gather of pixels. Ren X proposed another method for image segmentation named superpixel algorithm [14]. Using existing segmentation method, superpixel algorithms over-segment an image into perceptually meaningful atomic regions which can be used to replace the rigid structure of the pixel grid. Superpixels can capture image redundancy and greatly reduce the complexity of subsequent image processing tasks. Moreover, superpixel algorithms can supply additional information, such as mean and variance of gray levels in a superpixel, to subsequent segmentation algorithm. We applied the superpixel algorithm to the IR thermal image segmentation in order to improve the accuracy of detection and speed up the detection process. Many segmentation methods have been used to generate superpixels. The Normalized Cuts algorithm was used to generate superpixels in [13]. AchantaIn et al. proposed a superpixel algorithm named simple linear iterative clustering (SLIC) based on K-means in [15]. David Martin and Charless Fowlkes’ boundary detector was used in [16,17]. In this paper, the watersheds algorithm was compared to SLIC which is the latest popular superpixel algorithm and the former one was selected to generate superpixels in this application.
(f) Fig. 3. IR thermal frames (a)–(c) are frames obtained at 100 ms, 200 ms, 250 ms from specimen 1 and (d)–(f) are frames obtained at 100 ms 160 ms 200 ms from specimen 2.
After over-segmenting one thermal image into superpixels, the next work is to cluster the similar superpixels and obtain the final segmentation of the thermal image. Fuzzy c-means (FCM) clustering algorithm in [18] was used to cluster superpixels. We chose proper features of superpixels as the input of FCM algorithm. Compared with visual image, IR thermal image has low signal to noise ratio due to a great deal of noise coming from electronic noise, photon noise, aberrations of the optical system, environment temperature disturbance, and some other random factors [19]. When only using the superpixels’ features in raw gray level image, the segmentation result is found to be very poor. Because much noise can be eliminated by using a high-pass filter, the combined features of superpixels in both raw gray level image and high-pass filtered image were selected as the input of the clustering algorithm. We recognized the crack cluster as the one with biggest mean gray level. Finally, tiny noise still existing was removed using morphologic operator. We validated our frame by conducting experiment on test specimens with surface cracks manufactured by wire electrical discharge machining. The accurate information of cracks was automatically determined in a short time using the proposed thermal image processing framework. 2. Experimental procedure
Fig. 2. Specimens with cracks.
The experimental setup shown in Fig. 1 was made of a computer, an inductor, an IR thermal imager and two specimens. The
268
C. Xu et al. / Infrared Physics & Technology 67 (2014) 266–272
Fig. 4. Flowchart of the proposed thermal image processing framework.
(a)
(a)
(b) (b) Fig. 6. Superpixels generated by Watershed from (a) the original gray level image and (b) the corresponding gradient image.
(c) Fig. 5. Superpixels generated by SLIC with different desired number of superpixel and compactness factor: (a) 100 and 10 (b) 200 and 10 (c) 1599 and 10.
inductor generates the eddy current in specimens. It can supply repeatable, reliable heating with precise power control. The SATHY6850 IR thermal imager was used to capture the IR thermal video of the specimens. Its main characters are listed in Table 1.
The two specimens are shown in Fig. 2. They are made of C45E4 steel and the sizes of them are 90 30 6.5 mm and 296 25 8 mm respectively. On the specimens, there are a few of surface cracks manufactured using a wire cutting machine which can assure the width of them to be about 0.5 mm. The depths of cracks are about 4 mm. The specimens are clamped by a fixture and the imager is put on a tripod. In this way, the distance between the specimens and
Table 2 Mean gray level of each cluster. Frame in Fig. 3
Cluster 1
Cluster 2
Cluster 3
Cluster 4
Cluster 5
Cluster 6
(a) (b) (c) (d) (e) (f)
0.5216 0.5616 0.0857 0.7597 0.1384 0.7644
0.0909 0.0569 0.5017 0.7147 0.2607 0.4652
0.5645 0.0796 0.6853 0.8509 0.5342 0.7435
0.5536 0.5234 0.5355 0.3076 0.6948 0.8745
0.8799 0.8615 0.0741 0.5394 0.7812 0.5387
0.1333 0.4573 0.5135 0.3633 0.3322 0.2943
C. Xu et al. / Infrared Physics & Technology 67 (2014) 266–272
(a) raw IR image
(b) k-means segmentation result
(c) threshold segmentation result
(d) result of the proposed segmentation method
(e) final detection result of the proposed framework Fig. 7. Result of proposed method and some other segmentation algorithm.
the lens of imager can be fixed accurately. The specimens were excited using the inductor for 200 ms, and then the imager recorded the whole heating process of 3 s. 150 frames were captured during this process for each specimen. 3. Proposed method In order to relieve the operator from the laborious, time-consuming work, an automatic thermal image processing framework was proposed. The framework mainly includes four steps: superpixels generation, image segmentation, crack recognition and postprocessing. It works in sequence as shown in Fig. 4. We began by choosing a proper algorithm to generate superpixels for the application of crack detection in Section 3.1. Features of superpixels for the segmentation using FCM algorithm were selected in Section 3.2. In Section 3.3, we introduce a method to recognize the crack clusters from sound clusters. Morphologic operator was applied to eliminate some tiny interference in the recognized crack cluster. 3.1. Superpixels generation Due to the continuity of heat diffusion, the thermal features of a small area in IR thermal image are always approximately uniform.
269
Hence, superpixels are more proper basic image elements to be processed than individual image pixels. Superpixels can give more statistic information of a small area, such as mean and variance. The additional information will be favorable to the accuracy of segmentation and the accuracy of the detection can be improved. Using superpixels other than individual pixels as proposed element will reduce the complexity of subsequent image segmentation task and the detection process can be accomplished in less time. In superpixel algorithm, using an over-segmentation method, the given IR image is partitioned into sub-regions, each of which is called a superpixel. In this way, the problem of segmenting the image is cast into clustering the superpixels into groups. In order to simplify the image process, the raw IR image is inverted to a gray-level image firstly. There are many approaches to generate superpixels, each with its own advantages and drawbacks that may be better suited to a particular application. The generation speed, ability to adhere to image boundaries, and impact on segmentation performance are general characters to evaluate a superpixel algorithm. For the crack detection application, the adherence to the crack boundaries is the most important requirement for superpixel algorithm. SLIC has become a popular superpixel algorithm since it was proposed [15]. It used K-means to over-segment a color image in the 5d space of color information and image location. As the clustering method is simpler, it is very efficient. In SLIC algorithm, two parameters have to be determined at first, the desired number of superpixel and the compactness factor. The superpixels of Fig. 3(c) generated by the SLIC with three sets of these two parameters are shown Fig. 5. We can see that the two parameters determined by operator have a great effect on the result of superpixels generation. If the number of superpixel is limited too small, the generated superpixels will lose the boundary information of the cracks. Hence, it is important to find proper parameters for a satisfying detection result from SLIC method. Another popular segmentation algorithm, watershed can also be used to generate superpixels. Watershed is usually utilized on two kinds of images, the original image and the corresponding gradient image. Their respective results of Fig. 3(c) are shown in Fig. 6. From the results, we can see that the watershed segmentation of the original image obtains superpixels with too large size and therefor lost detailed information of cracks. The result of watershed on the gradient image (Fig. 6(b)) has a better adherence to cracks boundaries. Hence, the gradient image of IR thermal image is more proper for watershed algorithm.
(a) raw IR frame of specimen 1 (b) FCM clustering on gray level image (c) FCM clustering on the gray level image and high-pass filtered one (d) FCM clustering on superpixels using gray level information (e) segmentation result using the proposed algorithm (f) the final detection result using the proposed frame Fig. 8. Segmentation result of the proposed method and other FCM ones.
270
C. Xu et al. / Infrared Physics & Technology 67 (2014) 266–272
Fig. 9. Detection results using the proposed framework. The left column images are raw infrared thermal images. The right column images are detection results using the proposed framework.
The number of superpixels in Fig. 6(b) is 1599. Compared to the SLIC superpixels with the same number (Fig. 5(c)), the watershed superpixels have a little better adherence to cracks boundaries. Especially for an automatic detection system, the watershed has a superior performance that it avoids the determination of parameters as in the SLIC. In consideration of this important virtue and its equally satisfying result, the watershed algorithm is chosen to over-segment image and generate superpixels in our image process framework. 3.2. Thermal image segmentation After superpixel generation, the image is constructed from superpixels other than original individual pixels. Then, the image should be segmented into a few parts which compose of similar superpixels. We use the FCM clustering algorithm to accomplish the segmentation. The standard FCM objective function for partitioning X ¼ fxk jk ¼ 1; 2; . . . ; ng into c clusters is given by [18]
J¼
c X n X 2 um ik kxk v i k
ð1Þ
i¼1 k¼1
where uik is the membership function charactering the possibility of xk belonging to the i-th cluster. The parameter m is the weighted exponent that controls the fuzziness of the resulting partition and vi is the center of the i-th cluster. The clustering result is obtained by minimizing the objective function. This minimization can be accomplished in an iteration way. For the image segmentation based on superpixels, the data xk points to features of the k-th superpixel and the FCM algorithm will assign each superpixel to a cluster. To select features of superpixels properly is important to get a satisfactory segmentation result. Since each superpixel is a small image patch, the texture characters can be used as features of it. There are many popular texture features, such as mean, variance and entropy. The first three ones represent average brightness, average contrast and random degree respectively. In this
application, the mean and variance of a superpixel denoted by lraw and rraw are chosen as clustering features. When only considering the above two features of superpixels, the detection result is usually very poor for some thermal image frames. The reason lies in the low contrast between cracks and sound areas in these frames. This contrast depends on the capture time of the frame, environment conditions and crack dimensions. To improve the detection performance for some IR image with poor contrast between crack and sound area, some other characters of superpixels have to be added into the segmentation features. High-pass filter can sharpen edges in an image and make the changes of gray level in the image more distinct. Therefore, in the high-pass filtered image, the boundary between blurred cracks and around sound area should be sharpened. There is more information about these cracks in it than in the original gray level image. With observation and analysis, we used superpixels’ mean and variance in the high-pass filtered image denoted by lhpass and rhpass as additional features for clustering. After feature selection, the selected four features were normalized from 0 to 1 firstly. This normalization operation can eliminate the negative effect from different feature scope. Due to different importance of the four features for the superpixel clustering, different weight coefficients should be multiplied to the features respectively to adjust effect of them on the segmentation result. Depending on number of trial results, the weight coefficients for lraw, rraw, lhpass, rhpass were selected as 0.43, 0.07, 0.43 and 0.07 respectively. When the four weighted features of each superpixel have been computed, they will be substituted into the FCM algorithm as xk in Eq. (1). In the objective function of Eq. (1), two parameters have to be determined, the weighted exponent m and the number of clusters c. The weighted exponent is usually chosen as 2 in most applications and the same as in this paper. The cluster number depends on specific application. In this paper, the cluster number is specified as 6 by comparing the clustering results of different values of it. After all parameters and features are determined, the object function will be minimized using iteration method. When the object function arrives the minimum value, the k-th superpixel will clustered into the cluster with largest uik (i = 1, 2, . . .,6). Then each superpixel will be pointed to a specific cluster and the IR thermal image will be segmented into six parts. 3.3. Recognition of cracks cluster Among segmentation clusters, one represents cracks. In order to propose a fully automatic crack detection system, the crack cluster should be picked out from other clusters without manual intervention. Due to the obstruction of heat from cracks and heat accumulation around crack area, the mean temperature of crack area should be higher than surrounding sound surface. In the raw gray level image, there are some other areas have approximately high temperature as crack areas, which results in poor result using threshold segmentation. After clustering, most of these interferences have been disposed and the crack cluster always has the maximum mean lightness corresponding to the highest temperature. As a result, the mean gray level of each cluster was taken as a feature to recognize the crack cluster. The cluster which has maximum mean gray level in the corresponding raw gray image represents cracks. Table 2 shows the mean gray level of each cluster in the segmentation of Fig. 3(c) and the 3rd cluster will be recognized as cracks. 3.4. Postprocessing Between superpixels which generated using watershed algorithm, there exist boundaries with a single pixel width. Therefore,
271
C. Xu et al. / Infrared Physics & Technology 67 (2014) 266–272 Table 3 Crack covering measure. Method
Proposed 1a 2b 3c 4d 5e a b c d e
Specimen 1
Specimen 2
100 ms
200 ms
250 ms
Mean
100 ms
160 ms
200 ms
Mean
0.7183 0.5009 0.5691 0.5009 0.5884 0.2071
0.6369 0.4421 0.3212 0.4402 0.5806 0.2540
0.5158 0.1959 0.2894 0.1987 0.3024 0.1908
0.6237 0.3796 0.3932 0.3799 0.4905 0.2173
0.4936 0.2289 0.1609 0.1985 0.3846 0.1801
0.4560 0.2474 0.2233 0.2355 0.4152 0.1694
0.4347 0.0864 0.1622 0.1952 0.3714 0.1789
0.4614 0.1876 0.1821 0.2097 0.3904 0.1761
k-mean. Threshold. FCM clustering on gray level image. FCM clustering on the gray level image and high-pass filtered one. FCM clustering on superpixels only using gray level information.
the recognized crack cluster has been separated by the ridge lines. To get complete crack information, the separated superpixels representing one crack should be linked together. Closing is a proper morphologic operation for this task. It consists of dilation followed by erosion with the same structuring element. Before applying closing operation, image area only including crack superpixels should be transformed to a binary one. Because the boundary has only one pixel width, disk has radius of one pixel is selected as the structure element. The closing operation can eliminate the bays along the border area. Therefore, the bay of one pixel width between crack superpixels was filled and one crack became an entirety. Due to some remaining tiny noise in the recognized crack cluster, after closing operation, these noises were shown with cracks simultaneously. In order to get more accurate crack information, these noises should be eliminated in the final detection result. Tiny noises and cracks can be labeled as connected regions respectively. Because the noisy area always has very small size compared to cracks, the pixels number which one connected region includes was selected as a criteria to decide whether the connected region is a crack or not. The region with fewer pixels than the specified threshold will be disappeared in the final crack detection result. The specified threshold determines the size of crack that can be detected to some degree. In this paper, it is specified as 30.
low contrast. For the left frame, satisfying detection result can be obtained by using both the former two comparing methods and our proposed one, except FCM clustering of superpixel only using gray level information. However, for the right frame, only the proposed method can get the information of right crack, which has low contrast to sound area. When using the first comparing method, the left crack has been detected with expanded boundary and the right one has not been detected. When the second comparing method was used, even though the boundary of the left crack is detected more accurately, the right crack has not been detected yet. The third comparing method obtains the similar result as the second one. But with the proposed method, two cracks are clustered together. The final detect result shows that two cracks are detected on the right location using the proposed framework, although the dimension of crack which cannot be detected by other methods has some bias from the real size. The final detection results of all the six typical frames using the proposed framework are shown in Fig. 9. From it, we can see that all cracks can be detected using the proposed framework even in frames with poor quality. In order to evaluate the performance of the proposed framework quantitatively, a quantity measure is needed to evaluate the crack detection accuracy. In image processing research, a segmentation result is popularly evaluated using an indicator named covering measure [20]
4. Results and discussion
C¼
1X jsi \ g i j jg jmax jLj g 2G i si 2S jsi [ g i j
ð2Þ
i
To show the superiority of the proposed method to other classic segmentation algorithm, the proposed method, k-means and threshold segmentation were respectively applied to the typical thermal image shown in Fig. 3(f). The segmentation results of the image using the above three algorithms are shown in Fig. 7. From the result, we can see that k-means cannot detect the right crack which has low contrast to sound areas. Threshold segmentation also got an unvalued result with a great deal of noise. However, using the proposed method, three cracks have been all detected. To verify the validity of the proposed thermal image processing framework further, three other segmentation results of FCM clustering on raw gray level image, FCM clustering on raw gray level image and high-pass filtered image, FCM clustering on superpixels of the raw gray level image were compared to that of the proposed method. We picked out two representative frames of specimen 1 (Fig. 3(a) and (c)). In the former frame two cracks are both clear while one crack is blurry in the latter one. Final detection results are listed in Fig. 8. We found that for the left frame in the first row, the contrast between the two cracks and sound area is high. Meanwhile, in the right frame, only one crack has high contrast to sound area and the other crack is very blurry because of the
where G represents the ground truth segmentation which is annotated manually in this paper. S is segmentation to be evaluated. L is the original image to be segmented. || denotes the pixels number of area. jsjsi \[ ggi jj is the overlap measure between si and gi. However, for i i the crack detection application, whether cracks have been segment into one cluster is the key criteria to evaluate the segmentation result. As a result, the covering measure is modified to be crack covering measure in this application,
C¼
jd \ gj jd [ gj
ð3Þ
where g is the ground truth of cracks and d is the detected cracks. When the cracks are detected with accurate size, the crack covering measure will just be 1. In the case that cracks were detected with dilated size or with shrunk size, the crack covering measure will all be less than 1. The measure will be decreased with the increasing of difference between g and d. Therefore, the crack covering measure can evaluate the detection result accurately. In our experiment, two specimens were tested. For each specimen, 150 frames were captured using IR camera and the crack
272
C. Xu et al. / Infrared Physics & Technology 67 (2014) 266–272
covering measure of six typical frames were listed in Table 3. The proposed method gets the maximum crack covering measure for each frame compared with other methods. Using the proposed framework, all cracks have been detected and the average crack covering measure over the typical frames gets the maximum value for each specimen. Therefore, the proposed image process framework has satisfying performance and can detect cracks even in thermal image with poor quality.
5. Conclusions We showed that an automatic system to detect cracks on a metal part is an eager demand to widen the application of IR thermography. A superpixel based image processing framework has been proposed to accomplish this task. Watershed algorithm is chosen to generate superpixels due to its high adherence to the crack boundaries. Combined superpixels’ features from both raw gray level image and high-pass filtered image were selected as clustering features of the FCM algorithm considering the interference coming from various aspects. Compared with some other segmentation methods, the proposed framework can detect crack which has low contrast with sound area. It can weaken the sensitivity of the detection result to the frame quality and realize a whole automatic detection. The location and shape of cracks can be determined accurately by the proposed IR thermal image processing frame. In future work we plan to explore the applicability of the proposed framework to other materials, such as fibre reinforced composite.
Conflict of interest The authors declare that they have no conflict of interests.
Acknowledgements This work was financially supported by Shandong Provincial Natural Science Foundation of China (No. ZR2013EEQ028) and the Fundamental Research Funds for the Central Universities of China (No. 14CX02077A).
References [1] M. Omar, M. Hassan, K. Donohue, K. Saito, R. Alloo, Infrared thermography for inspecting the adhesion integrity of plastic welded joints, NDT E Int. 39 (2006) 1–7. [2] M.S. Jadin, S. Taib, Recent progress in diagnosing the reliability of electrical equipment by using infrared thermography, Infrared Phys. Technol. 55 (2012) 236–245. [3] S. Bagavathiappan, B.B. Lahiri, T. Saravanan, J. Philip, T. Jayakumar, Infrared thermography for condition monitoring – a review, Infrared Phys. Technol. 60 (2013) 35–55. [4] B. Yang, Y. Huang, L. Cheng, Defect detection and evaluation of ultrasonic infrared thermography for aerospace CFRP composites, Infrared Phys. Technol. 60 (2013) 166–173. [5] Maldague, P.V. Xavier, Introduction to NDT by active infrared thermography, Mater. Eval. 60 (2002) 1060–1073. [6] B. Oswald-Tranta, Thermo-inductive crack detection, Nondestruct. Test. Eval. 22 (2007) 137–153. [7] D. Wagner, N. Ranc, C. Bathias, P.C. Paris, Fatigue crack initiation detection by an infrared thermography method, Fatigue Fract. Eng. Mater. Struct. 33 (2009) 12–21. [8] H. Park, M. Choi, J. Park, W. Kim, A study on detection of micro-cracks in the dissimilar metal weld through ultrasound infrared thermography, Infrared Phys. Technol. 62 (2014) 124–131. [9] L.D. Favro, R.L. Thomas, X. Han, Sonic infrared imaging of fatigue cracks, Int. J. Fatigue 23 (2001) 471–476. [10] Y. Cheng, Mean shift, mode seeking, and clustering, IEEE Trans. Pattern. Anal. 17 (1995) 790–799. [11] J. Shi, J. Malik, Normalized cuts and image segmentation, in: Proceedings, 1997 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 1997 pp. 731–737. [12] L. Vincent, P. Soille, Watersheds in digital spaces: an efficient algorithm based on immersion simulations, IEEE Trans. Pattern. Anal. 13 (1991) 583–598. [13] P. Felzenszwalb, D. Huttenlocher, Efficient graph-based image segmentation, Comput. Vision 59 (2004) 167–181. [14] R. Xiaofeng, J. Malik, Learning a classification model for segmentation, in: Proceedings Ninth IEEE International Conference Computer Vision, vol. 11, 2003, pp. 10–17. [15] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, S. Susstrunk, SLIC superpixels compared to state-of-the-art superpixel methods, IEEE Trans. Pattern. Anal. 34 (2012) 2274–2282. [16] Mori, Greg, Recovering human body configurations: combining segmentation and recognition, in: Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, vol. 2, 2004, pp. II326–II333. [17] G. Mori, Guiding model search using segmentation, in: ICCV 2005. Tenth IEEE International Conference on Computer Vision, vol. 1412, 2005, pp. 1417–1423. [18] J.C. Bezdek, S.K. Pal, Fuzzy Models for Pattern Recognition, Piscataway, IEEE Press, NJ, 1991. [19] V. Vavilov, Noise-limited thermal/infrared nondestructive testing, NDT E Int. 61 (2014) 16–23. [20] P. Arbelaez, M. Maire, C. Fowlkes, J. Malik, Contour detection and hierarchical image segmentation, IEEE Trans. Pattern. Anal. 33 (2011) 898–916.