Segmentation of defects from radiography images by the histogram concavity threshold method

578 InsightVol49No10October2007

1. Introduction

Segmentation is a fundamental step in an automatic vision system. Its application can extract regions of interest which are subsequently classified. Thresholding is one of the important techniques for image segmentation based on the similarity in the brightness of objects. By choosing one or more thresholds of grey level, the object regions can be separated from their background for further processing[1]. Thresholding can be classified in to two categories viz, global and local thresholding techniques. The global thresholding technique is further classified into point-dependent and region-dependent techniques. Since global thresholding techniques can be used for measurement of shape and a region’s uniformity, this

technique can be adopted for estimation of volumetric defects such as porosities and inclusions in welds and voids in castings. The global thresholding works on the principle of segmenting an entire image with a single threshold value, whereas the local thresholding technique partitions a given image into sub-images and determines a threshold for each of these images, followed by smoothing of grey levels for eliminating the discontinuities in the image[2]. Automatic thresholding of grey-level values of an image is very useful in automated analysis of morphological images, and it represents the first step in many applications in image characterisation[3]. It is known that by choosing the threshold as the value that maximises the entropy of the one-dimensional histogram of an image, it is possible to separate the desired objects from the background[3]. The automatic thresholding techniques do not take into consideration the spatial correlation between pixels in an image. Since these methods work on the principle of correlation between pixels in the image rather than grey-level values interaction, the performance of these thresholding techniques deteriorates in a single image that has defects of different size but with same grey-level values. A fuzzy partition approach has been used for image thresholding that is based on both intensity distribution and local information among pixels. It has been found that a maximum fuzzy entropy principle is useful to segment the bi-level threshold and fuzzy region. The segmentation and location of small defects in a complex structure is still a very challenging task[4]. During the course of segmentation of small defects, fixing the boundary conditions is often complex and cumbersome.

Segmentation of small defects is carried out using the background subtraction and histogram threshold methods that delineate the feature from the background[5]. For segmentation of defects in cast parts, using real-time radioscopy, the MONDAN filter-based methodology was developed[6]. An automated flaw detection method was developed for aluminium castings, which was based on special filtering and masking segment[7]. A segmentation method based on bottom-up hierarchical analysis was developed to progressively agglomerate pixels/voxels in the image into non-overlapped homogeneous regions[8]. In the bottom-up approach, segmentation is achieved by clustering the regions at each level until no further change can be produced from this operation. For segmentation of 3D images, a 3D deformable model is used to fit to the object of interest. Earlier, several global and local thresholding techniques to segment the grey-scale images using a shape and region uniformity and local partition of an image into several sub-images have been reported[9-13]. For automated visual inspection of wooden pallets, a novel procedure based on the generalisation of the grey-scale histogram was developed[14]. Using the morphological transformation approach to recognise the defects in both spatial and frequency domains, a flaw detection method was developed[15]. It was shown that a suitable image analysis technique could provide a basis for quantitative assessment of micro-cracks in cement-based materials[16] where an automatic thresholding operation was carried out on the grey-level histogram. Using classical stereological techniques, the characteristics of the micro-cracks were quantitatively assessed. It was demonstrated that flaws

DIGITAL RADIOGRAPHY

Segmentation of defects from radiography images by the histogram concavity threshold method

TSaravanan,SBagavathiappan,JPhilip,TJayakumarandBaldevRaj

A histogram concavity-based thresholding approach has been used for segmenting porosities, voids and inclusions from digitised radiography images. Studies on small defects in cylindrical tubes and flat plates show that histogram concavity-based thresholding approach yields better results compared to conventional automatic thresholding techniques like Maximum Entropy and Otsu methods, where heterogeneous background present in the image due to the geometry and large X-ray quantum and structural noise gives poor segmentation. The segmentation procedure is optimised by filtering the abrupt noises with low-pass median filtering, followed by grey-level global thresholding. The histogram concavity-based thresholds provided excellent segmentation of pores and voids present in Electrochemical Hydrogen Sensors (ECHS), cast plates used in refineries, micro-pores in tube-to-tubesheet (TTS) welds of steam generator of fast breeder reactor, tungsten inclusion in Tungsten Inert Gas (TIG) welds, crater crack and external undercut in welds. An accuracy of ± 1-2 pixel sizes (65-130 microns) is obtained for the estimation of defects sizes from the grey-level profiling.

Dr Baldev Raj is the Distinguished Scientist and Director of Indira Gandhi Centre for Atomic Research, Kalpakkam. He is an Hon Member of ICNDT, Hon Fellow, Indian Society for NDT, Hon Fellow, The British Institute of NDT and Founder Member of the Board of Directors, World Federation of NDE Centres, USA.

Dr T Jayakumar is the Head, Non-Destructive Evaluation Division, Indira Gandhi Centre for Atomic Research, Kalpakkam.

Dr John Philip is the Head, SMART Section, Indira Gandhi Centre for Atomic Research, Kalpakkam.

Dr T Saravanan is a Scientific Officer in the Non-Destructive Evaluation Division, Indira Gandhi Centre for Atomic Research, Kalpakkam.

Shri S Bagavathiappan is a Scientific Assistant in the Non-Destructive Evaluation Division, Indira Gandhi Centre for Atomic Research, Kalpakkam.

Corresponding author: Dr John Philip, Head, SMART Section, Non-Destructive Evaluation Division, Inspection Technology Group, Indira Gandhi Centre for Atomic Research, Kalpakkam – 603 102, Tamil Nadu, India. E-mail: [email protected]

Paper submitted 10 June 2007Accepted 14 September 2007

DOI: 10.1784/insi.2007.49.10.578

InsightVol49No10October2007579

could be detected in radiography weld images by morphological based approach using a ‘Canny’ operator to find the edge pixels of an object, followed by morphological transformation to fix the boundaries[17]. It is found that edge linking with watershed-based segmentation procedure could measure the porosity size and shape reasonably well.

An automated tool box has been developed for porosity quantification in 2D micro-CT slices[18]. It was reported that the image binarisation performed on the basis of a new discrimination criterion, segments the images by intentionally de-emphasising the heterogeneity present in the background while more emphasising the grey-level of object to be segmented[19]. Recently, segmentation of radiography images for correct classification of defects was explored[20]. Using morphological operators and high-gradient pixels, a novel pattern recognition technique for the automatic inspection of flaws in polymeric insulators was studied in a systematic approach[21]. Using a template matching robust algorithm, circular casting defects were detected with an accuracy of 94.3%[22]. Fuzzy C means an algorithm was used to segment the defects like cluster porosities and cracks in welded tubes and compared with wavelet thresholding method[23]. As the background subtraction followed by histogram thresholding is not very effective in segmentation of small pores and voids, a histogram concavity-based thresholding procedure was adopted in the present work.

The objective of our present work is the segmentation of pores and voids present in welds and castings and high-density inclusion in welds. For this, an iterative median filtered radiography image was subjected to the segmentation process by a histogram concavity-based approach which fixes the threshold at the shoulder of the histogram. This methodology was first applied on calibration samples with known defect dimensions and the results obtained are found to be in good agreement with the manual threshold (interactive) procedure. The popular global automatic thresholding techniques like Otsu and Maximum Entropy were applied on the radiography images for comparison. These methods do not yield a good segmentation since these methods require a valley in the histogram to threshold between object and background. The histogram concavity threshold was successfully applied to segment the defect features such as porosities and voids in ECHS, pores and voids in cast plates and high density tungsten inclusion in TIG welded tubes.

2. Experimental

Conventional film radiography of the samples was carried out using a 450 kV Balteau constant potential X-ray unit that had a dual focal spot size of 1.2 × 1.2 mm and 3 × 3 mm. Radiography exposure parameters were suitably adjusted for the cylindrical and plate geometry specimens to obtain a radiography density of around 2.0-2.5 in the region of interest (ROI). Microfocal X-ray radiography was carried out on TTS weld samples by Fein Focus FXE 225.20 which had a focal spot size of 15 microns. To have a better control on radiography sensitivity, low energy coupled with fine grain, slow speed Kodak DR50 film was used. A 3 mm lead sheet was used at all other regions except at the region of interest for masking around the specimens to control the scattering. Exposed films were processed manually under standard conditions. Table 1 shows the typical radiography parameters employed for the experiments. All the radiography images were digitised by a laser X-ray film scanner with a scanning resolution of 50 microns and grey-scale resolution of 4096 levels.

3. Results and discussion

3.1 Image pre-processing by median filtering

The median filter replaces a pixel’s grey level with the median value of its neighbourhood. This method is very effective when the noise

pattern is strong and has spike-like components in the digitised radiography image. The primary role of the median filtering is to make points with distinct intensities to appear like neighbourhood pixels. Unlike neighbourhood averaging filters, these median filters are characteristic of preserving the edge sharpness of the image. As pointed out by Halmshaw[24], in X-ray film radiography, the X-ray quanta absorbed in a photographic emulsion layer produce secondary electrons which in turn sensitise the silver grains. It is known that use of spatial filters such as median would considerably reduce the blurring of edges and preserve the boundaries present in the image, thus making it suitable for visual examination and measurements of features[25]. It was shown that noise pixels (high frequency component of an image or a signal) can be removed effectively by low-pass filtering templates which have a kernel size of either 3 × 3 or 5 × 5. These kernels are operated as convolution operators on an image[26]. For the present experiment, all the digitised radiography images are subjected to median filtering to remove the abrupt noises.

Case 1:

Figure 1(a) shows the digitised radiography image of a hole of diameter 0.73 ± 0.02 mm drilled by the Electrical Discharge Machined (EDM) process in a 9Cr-1Mo welded tube. The outer diameter of the tube is 20 mm and the thickness is 3 mm. During radiography exposure, the hole was placed on the film side to reduce the effect of distortion arising out of defect to detector distance and magnification. From Figure 1(a), it is seen that the image of the hole is blurred and its boundary is not clear. This could be due to the total unsharpness, which is around 100 microns. Figure 1(b) shows the segmented image of Figure 1(a) using histogram concavity approach.

Figure 2 shows the grey-level profile across the hole shown in Figure 1(a). From the profile, the diameter was estimated from full width at half maximum of the peak and is found to be 0.90 mm against the actual diameter of 0.73 ± 0.02 mm. The difference in the size of the hole could be due to the total unsharpness (100 microns) and edge spreading at the boundary of the hole.

Parameters

X-ray unit Balteau NDT 450 kV

Voltage 20-450 kV

Current 30 mA max

Focal spot size(Dual focal spot)

1.2 × 1.2 mm, 3.0 × 3.0 mm

Cooling Oil/air

Parameters applied

For holes in the circular welded

tube

For 3 mm hole in the flat plate

For 32 mm thickness cast plate/ECHS/TTS weld

Voltage applied 150 kV 150 kV 240/140/90 kV

Exposure 30 mA.mts 72 mA.mts 105/2/ mA.mts/35 mA.mts

Radiography density

2.0-2.5

Processing Manual, standard

Detector used Fine grain, slow speed Kodak DR50

Intensifying screen

0.05 mm front and 0.10 mm back Pb screen

SFD 1000 mm 700 mm 700/700/25 mm

Magnification 1 1 1/1/3

Table 1. Typical radiography exposure parameter applied for different specimens

Figure 1. (a) Radiography image of a welded tube with 0.73 ± 0.02 mm hole on the weld, (b) its segmented image and (c) histogram of 1(a)


To segment the image shown in Figure 1(a), a grey-level histogram-based method was used which sets the threshold value at the shoulder of the histogram[2]. This provided a good segmentation of the object and is in good agreement with the manual thresholding approach. For images with distinct objects and background, it is possible to select the threshold from the grey-level histogram, using the ‘Mode’ method[2]. For images where valleys are not found in their grey-level histogram, it is possible to define a good threshold at the shoulder of the histogram. This approach is valid since both the valleys and shoulder correspond to the concavities in the histogram. The grey-level histogram for Figure 1(a) is shown in Figure 1(c). The histogram shows that the object and background pixels have grey values grouped into two dominant modes. Object and background pixels were separated out by fixing a threshold value T which separates the object with respect to background. The procedure to extract pixels for a region of interest is to set a value at any point (x, y) for the image f (x, y), where (x, y) > T is called object point, otherwise it is background pixel. The histogram of any object is the characteristic of image consisting of grey objects superimposed on a dark background.

In the case of histogram concavity approach, for a histogram HS, the height of the histogram at grey levels of g0, g1, g2……, gl-1, are h (g0 ), h (g1), h (g2 ) ….., h (gl-1 ), where h(gi ) ≠ 0 for all i. Hence HS has a two-dimensional region. To determine the concavities of HS, its convex hull is constructed. It is the smallest convex polygon HS which contains HS. The concavities of HS are determined from set-theoretic difference HS – HS . Let h(gi ) be the height of HS at grey-level gi. Possible threshold values are grey levels at which h (gi )-h(gi ) has a local maxima.

For Figure 1(a), from the above histogram, the threshold value at the shoulder is at 118. Hence the grey values above 118 are for background and below for the region of interest (ROI). Figure 1(b) is the result of such grey level adjustment. In Figure 1(b), since the pixels corresponding to background were eliminated, ROI alone is seen with clear boundary. To quantify the dimension of the hole, in Figure 1(b), the image was binarised that contains grey value of 1 for object pixels and 0 for background pixels. In the present case, due to a larger contrast in intensity between the object and the background, the thresholding operation is easier. The grey-level adjustment for setting the threshold value by manual interactive procedure is in good agreement with the above result. Hence, it is understood that the boundary conditions set on the basis of a number of trials for these samples yielded a good segmentation of defect features using the global thresholding approach. From Figure 1(c), it is inferred that due to the distinct large background pixels, the histogram is represented with a broad peak for background whereas due to the small object area, the valley region is missing. The segmented hole diameter is measured as 0.89 mm from the grey-

level profile as indicated in Figure 3 against the actual diameter of 0.73 ± 0.02 mm. The error is 20% which is attributed to the total unsharpness associated with the radiography arrangement.

3.2 Segmentation through popular automatic thresholding methods

3.2.1 Maximum Entropy approach

To segment the weld hole of 0.73 mm, besides histogram concavity and manual thresholding approaches, popular automatic thresholding methods like Otsu and Maximum Entropy were attempted[25]. The entropy maximisation method works on the principle of fixing a threshold (T) in such a way to maximise the information available in the two grey-level distributions of the object and background. This information is measured by the entropy function φ (k), where:

φ ( ) ( / ).log( / ) ( / ).log( / )k0 0 1 1

= − +p w p w p w p wi i i i

kk

k

+∑∑

1

max

min

..(1)

where pi is the probability for a given pixel to have an intensity value equal to i(pi = Ni/N), Ni is the number of pixels that have an i intensity, and N is the total number of pixels. w0 and w1 are the probability to find a pixel intensity lower or higher than:

( ) ;max

min

k0 1

1

w p w pi i

k

k

= =

+

∑∑ .........................(2)

The lower bound thresholding (B (x, y)) is found out for a given image f(x, y), with a grey-level threshold limit ‘s’:where: B (x, y) = 0 if f(x, y) < s ...............................(3) 1 if f(x, y) > s

Entropy maximisation method resulted in a threshold (T) limit in the case of the welded tube (case 1) as 134 whereas for the manual and histogram concavity methods, the value was 118. In the entropy method, the diameter of the hole is estimated as 0.94 mm against the actual hole diameter of 0.73 ± 0.02 mm. The error is around 30%. This high error is attributed to the fact that the circular geometry would lead to a heterogeneous background where these automatic thresholds would perform poorly. Since these methods are sensitive to a single pixel which has a grey value equal to that of the noise pattern, this method is not suitable for images with lower signal-to-noise ratio (SNR).

Otsu method

For an image f (x, y) which has a grey-level range of 0 to L-1, where L is the distinct number of grey levels. Suppose n be the total number of pixels in an image, ni is number of pixels with grey level i, then the probability of pixels which will have intensity i is:

Pi= n i /n ........................................(4)

Figure 2. Grey-level profile of the hole in a weld Figure 3. Grey-level profile of the weld hole after binary thresholding


The average grey level of the entire image is:

m ==

−

∑ ipi

i

L

0

1

.........................................(5)

For single thresholding the pixels in an image are divided into two classes C1 = { 0, 1, 2……,t) and C2 = {t+1, t+2,…..,L-1}, where t is the threshold value. C1 and C2 are corresponding to object and background[28]. The probabilities of the two classes are:

w t pi

i

t

10

( ) ==∑ ......................................(6)

and

w t pi

i

L

2t+1

1

( )–

==∑ .....................................(7)

Using discriminant analysis, the Otsu method will show the optimal threshold by maximising the class variance.

Using the Otsu method, the segmentation process has yielded poor extraction of defect feature for the case of the welded tube. In this case, threshold was set at 191 which includes both defect and background. Since the Otsu method could not perform well for unimodal grey-level distribution and it requires a valley in the grey-level distribution, it has given a false segmentation.

It is clear from the above results that both the Otsu and Maximum Entropy-based approaches could not yield a better threshold for a small defect feature which had a heterogeneous or large background. These methods would perform well for cases where a distinct boundary separation in between grey levels of different classes, which would usually have a valley point in the bi-modal histogram. Hence, for segmentation of micro-porosities and voids with noisy and uneven background, these automatic thresholds which work on the basis of variance between classes are not suitable.

Case 2:

Figure 4(a) shows the digitised radiography image of a 1.02 ± 0.02 mm diameter hole in a weld. In comparison to Figure 1(a), the boundary of the hole in the image is clear due to the larger size of the hole. Figure 4(b) shows the segmented image for the hole using the histogram concavity threshold method.

For estimation of the diameter of the hole in Figure 4(a), a line profile was drawn across the hole and using full width at half maximum (FWHM ) at the grey-level profile, the diameter was measured as 1.08 mm against the actual value of 1.02 ± 0.02 mm. The error in this case is 1 pixel (65 microns) of size. The diameter of the segmented hole in Figure 4(b) is measured using grey-level profile as 1.085 mm. In this case, the other two automatic thresholding techniques have not yielded good results.

Case 3:

Figure 5(a) shows the digitised radiography image of a 3 ± 0.02 mm hole fabricated by EDM process in a stainless steel (SS) flat plate of 8 mm thickness. The depth of the hole is about 50% of the thickness of the plate. From the radiography image, the boundary of the hole is clearly seen due to the larger size of the hole and due to uniform background intensity. Figure 5(b) shows the segmented image of Figure 5(a) for 3 ± 0.02 mm hole by the procedure as given for the case (1). Under the influence of constant background pixels, the brightness factor for the defect feature is easy to be thresholded from the background. From the grey-level profile on the line drawn across the image given in Figure 5(a), the diameter

of the hole is estimated as 3.13 mm against the actual diameter of 3 ± 0.02 mm. The variation is 130 microns which is about 2 pixels in size, which could be accounted for the total unsharpness associated with the system. From the grey-level profile for the line drawn across the binary thresholded image for the 3 mm hole, the diameter is measured as 3.07 mm which is about one pixel variation in size which could be due to the error in fixing the edge pixel between object and background. From Figure 5(c), it is clear that the histogram shows a uni-modal characteristic for the image shown in Figure 5(a). The threshold value for this case is fixed at the shoulder of the histogram as 98. This is in close agreement with the Otsu automatic threshold setting where T is 99. It is found that even the histogram showed a uni-modal distribution of grey levels for the image, both the histogram concavity and Otsu methods exhibited a good segmentation due to a large size of the defect feature with homogeneous background resulted from the plate geometry. Generally, the Otsu and other automatic thresholding methods will yield a better segmentation where the grey-level distribution is clear with bi-modal characteristics. From the above results, it is clear that any automatic and manual threshold techniques would segment the defect features reliably that have distinct grey-level distribution.

3.4 Detection of porosity and voids in castings

A carbon steel plate of dimensions 600 × 80 mm (LXB) and thickness of 32 mm was used for these studies. The plate was manufactured by a casting process and was a failed component from a petrochemical plant.

Figure 6(a) shows the radiography image of the cast plate. The radiography image shows pores and voids of varying size. Figure 6(b) shows the segmented image of Figure 6(a) by histogram concavity-based threshold approach. For separation of the features, the threshold was fixed at the grey level of 71. Then the segmentation was tried with manual threshold where the object pixels separation is in close agreement with the above approach. In Figure 6(a), the circled region has a low density inclusion. From the above procedure, the grey value of the inclusion at the boundary is 72-79 which is almost close to that of the grey value

Figure 4. (a) Radiography image of 1.02 ± 0.02 mm hole in a weld and its segmented image (b)

Figure 5. (a) Digitised radiography image of a 3 ± 0.02 mm hole in a plate, (b) its binary thresholded image and (c) histogram of Figure 5(a)

Figure 6. (a) Digitised radiography image of a cast specimen, (b) its binary thresholded image by histogram concavity method and (c) thresholding by maximum entropy method


of the background pixels. The optical density measured on the boundary of inclusion is also very close to the background region. Therefore, the optimal global thresholding has not resulted in correct extraction of this low-density inclusion as indicated by the circle in Figure 6(b).

Again, segmentation of the defects in Figure 6(a) is attempted by popular automatic multi-level thresholding methods like Otsu and Maximum Entropy. However, due to poor threshold set by the Otsu method, the image was not segmented properly, since this method requires a clear multi-model or bi-modal grey-level distribution and also it requires a distinct variance between defect and background classes. The threshold was set in between 159-255 by Otsu. Since the histogram is represented by a single unimodal grey-level representation, this method gave poor segmentation. Segmentation using maximum entropy approach is shown in Figure 6(c). Here, the threshold limit is set at 91. It was seen that, in addition to correct segmentation of pores and voids, it also extracted some of the background pixels which had grey values close to that of defects (indicated by circles). This could be due to the threshold setting which includes some of the background pixels which have the grey values very close to pores and voids.

From the above studies, it is clear that for segmentation of defects like clustered porosities, isolated pores and micro pores in large structured plates manufactured by casting where pores or holes are predominant defects, the popular automatic thresholding methods are not suitable since it could lead to false indication of defects. In these cases, a histogram concavity-based approach would yield a better segmentation.

3.5 Detection of porosity and voids in ECHS cast electrolyte

Figure 7(a) shows the digitised radiography image of an Electrochemical Hydrogen Sensor (ECHS) which is used as hydrogen leak detector in the secondary sodium circuits of Fast Breeder Reactors (FBR). The ECHS consists of working and reference electrodes and solid electrolyte as the medium for ionic conduction. Electrolytes are prepared by solution casting method where there is a possibility of air entrapment that would result in the occurrence of pores and voids in the electrolyte medium. Hence, the conductivity will be largely affected in the presence of air entrapment and results in poor performance of the meter. In Figure 7(a), it is observed that a large void is present due to air entrapment. For extraction of this void, histogram concavity-based threshold is fixed and the segmented image is shown in Figure 7(b). For this case, the automatic thresholding methods gave poor segmentation. Since the histogram representation of these images does not have a bi-modal characteristics (with valley point), these techniques incorrectly fixed the threshold at the middle of the histogram where in addition to object pixels, background pixels were also present. Figures 7(c) and 7(d) represent radiography images of an elongated cavity in the electrolyte and its binary thresholded image by histogram concavity method.

3.6 Detection of micro-porosities in tube-to-tubesheet (TTS) weld joints

In the steam generator of a prototype fast breeder reactor, the transfer of heat from secondary sodium to water generates steam. Since the sodium-water reaction is exothermic and generates hydrogen which produces high pressure leaks in the sodium system, the integrity of the TTS weld joints separating sodium and water/steam is of paramount importance. Microfocal radiography of the welds was carried out with a projective magnification of 3.0. Figure 8(a) shows the microfocal radiography image of one of the TTS weld joints of the steam generator in which porosities are seen. Using the histogram concavity-based global thresholding technique, the image was segmented and the same is shown in Figure 8(b), where micro-pores are clearly seen. Here, the threshold was fixed at 170

and it is in close agreement with the manual thresholding approach. The Otsu method fixed the threshold at 194 and Maximum Entropy method set the threshold at 174. Due to the poor threshold fixed by these methods, the segmented image contains false indications of defect features.

3.7 Detection of tungsten inclusion in TIG welds

Figure 9(a) shows the radiography image of a welded plate with tungsten inclusion. Since TIG welding process is used, due to entrapment of high-density tungsten from the tip of the electrode, a white spot is seen in the radiography image. In this case, due to a significant grey-level variation of tungsten from the rest of the weld and background plate, binary segmentation resulted in the correct extraction of tungsten inclusion which is shown in Figure 9(b). Histogram concavity method fixes the threshold at 178. Since the Otsu method fixes the threshold at 77, which is the valley point of the histogram, the segmentation is improper. The Maximum Entropy method fixed the threshold at 176 which gave reasonably good separation of the inclusion from the background. However, some of the background pixels are also segmented along with the inclusion.

3.8 Detection of crater crack in welds

Figure 10(a) shows the radiography image of a weld with crater crack (shown by the circle). The crater crack is formed by abrupt weld termination. Initially, crater crack may be star shaped, which may extend up to the edge of the crater. However, it can propagate into longitudinal direction of the weld. From the radiography image, it is seen that the grey-level contrast of the crack is distinct from the background. Hence, for a uniform background intensity, optimising the threshold is easy. The grey levels for the crack appear at the bottom of a uni-modal histogram. The histogram concavity approach sets the threshold at 54 whereas Maximum Entropy and Otsu methods set it at 109 and 122 respectively. Figures 10(b),

Figure 7. (a) Digitised radiography image of an Electrochemical Hydrogen Sensor (ECHS), (b) Binary thresholded image of (a); (c) Another ECHS with elongated cavity and (d) Binary thresholded image of (c)

Figure 8. (a) Microfocal radiography image of a typical TTS weld joint and its binary thresholded image (b)

Figure 9. (a) Radiography image of a weld with tungsten inclusion and (b) binary thresholding applied on Figure 9(a)


(c) and (d) represent the result of segmentation of the image 10(a) by histogram concavity, Maximum Entropy and Otsu methods respectively. In comparison to the histogram concavity method, the automatic threshold method resulted in poor segmentation, which could be due to incorrect threshold setting at the middle in the histogram. This resulted in segmenting a part of the background along with the defect. In Figures 10(c) and (d), the defect is buried in the background.

3.9 Detection of external undercut in welds

Figure 11(a) shows the radiography image of a weld with external undercut shown by a circle. During the course of welding, higher welding current could result in burning of edges of the parent metal. The radiography image of an undercut appears as a dark continuous line at the edge of the weld. It is observed that optimising the threshold is reliable due to the higher grey-level contrast of undercut from the background. Histogram concavity set the threshold at 57 whereas Maximum Entropy and Otsu methods fix the threshold at 136 and 164 respectively. Figure 11(b) shows the result of the segmented image of Figure 11(a) by the histogram concavity method. In comparison to the histogram concavity method, other automatic threshold methods resulted in a poor segmentation due to incorrect threshold fixing so that part of the background close to that defect class is also segmented.

4. Conclusion

The segmentation of defects is carried out by the histogram concavity-based thresholding approach to extract defects such as pores and voids in cast plates and electrochemical hydrogen sensors, micro porosities in tube-to-tube sheet welds and high-density inclusion in TIG welded plates. Studies on small defects in cylindrical tubes and flat plates, shows that the histogram concavity-based thresholding approach yields better results compared to conventional automatic thresholding techniques like

Maximum Entropy and Otsu methods. The segmentation procedure is optimised by filtering the abrupt noises with low-pass median filtering, followed by grey-level global thresholding. It has been found that histogram concavity approach provides the best result for segmenting features such as pores, voids and high-density inclusion.

Acknowledgements

We thank Dr P R Vasudeva Rao and Mr P Kalyanasundaram for support and encouragement.

References

1. R C Gonzalez and P Wintz, ‘Digital Image Processing’, 2nd ed; Addison-Wesley, Reading, Massachusetts, 1987.

2. P K Sahoo, S Soltani and A K C Wong, ‘A survey of thresholding techniques’, Comp. Vision Graphics and Image Processing, 41, pp 233-260, 1988.

3. H D Cheng and Yen-Hung Chen, ‘Fuzzy partition of two-dimensional histogram and its application to thresholding’, Pattern Recognition, 32, pp 825-843, 1999.

4. Duan-hu Shi, Tie Gang, Shuang-yang Yang and Yuan Yuan, ‘Research on segmentation and distribution of small defects in precision weldments with complex structure’, NDT & E International, accepted manuscript.

5. V Kaftandjian, A Joly, T Odievre, M Courbiere and C Hantrais, ‘Automatic detection and characterization of aluminum weld defects, comparison between radiography, radioscopy and human interpretation’, Proc. of 7th Euro Conf. on NDT, Copenhagen, Denmark, pp 1179-1186, May 1998.

6. Yi Sun, Peng Bei, Hong-yu Sun and Ping Zhou, ‘Real-time automatic detection of weld defects in steel pipe’, NDT & E International, 38, No 7, pp 522-528, 2005.

7. D Mery and D Filbert, ‘ A review of methods for automated recognition of casting defects’, Insight, 44, No 7, pp 428-436, 2002.

8. X Shen, M Spann and P Nacken, ‘Segmentation of 2D and 3D images through a hierarchical clustering based on region modelling’, Pattern Recognition, 31, No 9, pp 1295-1309, 1998.

9. S K Fu and J K Mu, ‘A survey on image segmentation’, Pattern Recognition, 13, pp 3-16, 1981.

10. J S Weszka, ‘A survey of thresholding techniques’, Comput. Vision Graphics Image Process. 7 pp 259-265, 1978.

11. W Doyle, ‘Operation useful for similarity-invariant pattern recognition’, J Assoc. Comput. Mach, 9, pp 259-267, 1962.

12. S M X Fernando and D M Monro, ‘Variable thresholding applied to angiography’, Proc. 6th Inter. Conf. on Pattern Recognition, 1982.

13. T Pun, ‘A new method for grey-level picture thresholding using the entropy of the histogram’, Signal Process, 2, pp 223-237, 1980.

14. M A Patricio and Maravall, ‘A novel generalization of the grey scale histogram and its application to the automated visual measurement and inspection of wooden pallets’, Image and vision computing, 25, pp 805-816, 2007.

15. A Movafeghi, M H Kargarnovin, H Soltanian-Zadeh, M Taheri, F Ghasemi, B Rokrok, K Edalati and N Rastkhah, ‘Flaw detection improvement of digitised radiographs by morphological transformations’, Insight, 47, No 10, pp 625-630, 2005.

16. A Ammouche, D Breysse, H Hornain, O Didry and J Marchand, ‘A new image analysis technique for the quantitative assessment of microcracks in cement based materials’, Cement and Concrete Research, 30, pp 25-35, 2000.

17. A Alaknanda and R S Pradeep Kumar, ‘Flaw detection in radiographic weld images using morphological approach’, NDT & E International 39, pp 29-33, 2006.

Figure 11. (a) Radiography image of a weld with external undercut and (b) segmentation by histogram concavity threshold method

Figure 10. (a) Radiography image of a weld with crater crack, (b) segmentation of the crater by histogram concavity threshold on Figure 10(a), (c) Maximum entropy threshold and (d) Otsu threshold method


18. A A Malcolm, Y H Leong, A C Spowage and A P Shacklock, ‘Image segmentation and analysis for porosity measurement’, J. Mat. Processing Tech, accepted manuscript, 2007.

19. Songcan Chen and Daohong Li, ‘Image binarization focusing on objects’, Neurocomputing, 69, 2411-2415, 2006.

20. S M Anouncia and R Saravanan, ‘Non-destructive testing using radiographic images’, Insight, 48, No 10, pp 592-596, 2006.

21. W C Godoi, R R da Silva and V Swinka-Filho, ‘Pattern recognition in the automatic inspection of flaws in polymeric insulators’, Insight, 47, No 10, pp 608-614, 2005.

22. A Ghoreyshi, R Vidal and D Mery, ‘Segmentation of circular casting defects using a robust algorithm’, Insight, 47, No 10, pp 615-617, 2005.

23. X Wang and B S Wong, ‘Segmentation of radiographic images using fuzzy c-means algorithm’, Insight, 47, No 10, pp 631-633, 2005.

24. R Halmshaw, ‘Scattered radiation in industrial radiography’, Brit J NDT, 35, pp 111-118, 2005.

25. R C Gonzalez and R E Woods, Digital Image Processing, 2nd ed, 2002.

26. Gang Wang and T Warren Liao, ‘Automatic identification of different types of welding defects in radiographic images’, NDT & E International, 35, pp 519-528, 2002.

27. S Abutaleb Ahmed, ‘Automatic thresholding of grey-level pictures using two-dimensional entropy’, Comp. Vision Graphics and Image Processing, 47, pp 22-32, 1989.

28. N Otsu, ‘A threshold selection method from grey-level histograms’, IEEE Trans. Syst. Man. Cybern, 9, pp 62-66, 1979.

Enquiry No 710-

Segmentation of defects from radiography images by the histogram concavity threshold method

Documents