AUTOMATIC EVALUATION OF WELDED JOINTS USING IMAGEPROCESSING ON RADIOGRAPHS

Ch. Schwartz

CEA - Centre de Valduc - 21120 IS SUR TILLE - FRANCE

ABSTRACT. Radiography is frequently used to detect discontinuities in welded joints (porosity,cracks, lack of penetration). Perfect knowledge of the geometry of these defects is an important stepwhich is essential to appreciate the quality of the weld. Because of this, an action improving theinterpretation of radiographs by image processing has been undertaken. The principle consists inmaking a radiograph of the welded joint and of a depth step wedge penetrameter in the material. Theradiograph is then finely digitized and an automatic processing of the radiograph of the penetrameterimage allows the establishment of a correspondence between grey levels and material thickness. Analgorithm based on image processing is used to localize defects in the welded joints and to isolatethem from the original image. First, defects detected by this method are characterized in terms ofdimension and equivalent thickness. Then, from the image of the healthy welded joint (that is to saywithout the detected defects), characteristic values of the weld are evaluated (thickness reduction,width).

INTRODUCTION - PRINCIPLE OF THE STUDY

Research for smallest flaws and for perfect knowledge of welded joints is animportant step in the improvement of non-destructive controls. In some specificapplications, radiography is a well-adapted non-destructive technique that allows thequantification of quality of welds (thickness reduction, flaws...)- A method based onindirect thickness measurement has been considered. This method consists in modeling acalibration curve representing thickness of the material been radiographed as a function ofthe density of the film [1]. Such a relationship is possible by the use of a specialpenetrameter (usually called a depth step wedge) made in the same material than thespecimen. The weld of the specimen is placed beside the depth step wedge in order to beradiographed on the same film and to avoid errors from differences in radiation intensityacross the field. The stepped wedge must be as wide as possible and, in order to limit theeffects of the radiation, lead masking is frequently used. It is important to have exactly thesame operating conditions (voltage, intensity, time of exposure, film development...) toestablish a reliable relationship between the film density and the thickness of material(from the image of the depth step wedge) and then to extrapolate it to the weld.

When appropriate experimental conditions are observed, the film is finely digitizedwith a microdensitometer (50 x 50 urn on 256 grey levels). The images are then exploredwith specific procedures using image-processing algorithms and automatic evaluation ofthe weld is possible.

CP657, Review of Quantitative Nondestructive Evaluation Vol. 22, ed. by D. O. Thompson and D. E. Chimenti© 2003 American Institute of Physics 0-7354-0117-9/03/S20.00

689

Grey Level

FIGURE 1. Principle of the calibration curve.

CALIBRATION CURVE MODELING

The first step of the automatic treatment consists in establishing a calibration curvebetween the grey levels of the digitized image as a function of the materialthickness : Gl = f(Th). Average pixel values of each step of the wedge's image arecalculated so that each thickness of the step wedge corresponds with a grey level value(Figure 1).

In order to predict the thickness correspondence for each grey level value on theimage of the weld, data fitting is necessary. By this effect, a model based on theattenuation law has been developed :

~Th n2 + B,.e~lh (1)

where : Th represents material thickness,Gl represents equivalent grey levels,B represents curve coefficients.

This nonlinear curve-fitting problem is solved in the least-squares sense [2]. Thatis, given input data Th = {Thi, ..., Th6}, and the observed output G\ = {On, ..., Gi6},coefficients B = {B0, BI, B2} that "best-fit" the equation F(B, Th) are determined.Actually, it consists in resolving the function :

(2)

The calibration curve is now determined for all the grey level values of the imageof the depth step wedge (Figure 2-a). However, some adjustments are necessary before theautomatic evaluation of the welded joints can be applied. In fact, it appears that the step ofthe wedge, which has the same thickness than the specimen, has not exactly the samedensity on the film.

690

Grey Level Number of pixel

tGrey Level

FIGURE 2. (a) Modeling of the calibration curve - (b) Histogram of the raw image - (c) correction of thecalibration curve.

The calibration curve must be corrected of this drift resulting from the non-isotropyof the X-ray cone beam and from the scattering radiation effects. The average grey levelvalue AGL is extracted form the histogram of the image of the weld (Figure 2-b) andcorresponds directly with the thickness e of the specimen. The characteristic point of theweld (e, AGL) is then put on the calibration graph. The curve generated from the result ofthe stepped wedge is moved to the characteristic point of the weld in order to matchexactly with the different grey level values of the weld (Figure 2-c).

NON-UNIFORMITY ILLUMINATION CORRECTION

Due to a anisotropy of the X-ray cone beam, a non-uniform illumination of thebackground of the image is frequently observed. For example, the background of theimage can be much brighter in the top part of the image than in the bottom part. Thisvariation must be subtracted out from the background of the image in order to not disturbthe results of the automatic evaluation algorithm. Several methods have been tested(mathematical morphology, low filtering...). Good results are obtained by polynomialfitting [3]. Each column of the raw image is fitted by a polynomial function (Figure 3-a).In order to not change the information of the weld during the correction of the non-uniformillumination, the data-fitting algorithm only takes in account points that are not included inthe weld itself, it means points that are far enough from the center of the weld. The non-uniformity illumination of the background is then modeled for all the profiles of the imageof the weld (Figure 3-b). This effect is extracted by subtracting the modeled backgroundimage to the original image of the weld. Then, all the profiles are translated to a realisticgrey level value by simply adding to all the pixels, the value AGL (Figure 3-c).

Grey Level Grey LevelGrey Level

c^w*^

(a)FIGURE 3. (a) Profile in the raw image - (b) non uniformity illumination of the background - (c) Profileafter processing.

691

FLAWS DETECTION AND AUTOMATIC EVALUATION OF THE WELD

A good approach for locating defects in the image of the weld (Figure 4-a) is todetect its edges. Flaws are detected on the image by the use of the Canny-Deriche filter[4] [5]. The optimum edge detector used is the combination of two monodimensional filtersin the two directions of the images. A smoothing filter f(x) and the optimum derivativefilter h(x):

N —a\x\e ' '

h(x) = cxe'a^

The image of the norm of the gradient B is calculated as followed [6] :

\Bx=(A*h(x))*f(y)

(3)

B — A/ BX + By (4)

Edges of the flaws correspond to the local maxima of the norm of gradient B. Theyare directly extracted by thresholding. The noise of the raw images generates the detectionof artifact's edges. In order to limit the detection of these edges, the appearance in theprofil of detected edges in the image are analyzed. The idea is that each profil of a flawlike a porosity is composed of three different areas characterized by specific criteria :1) a "descendant area" characterized by a minimum descendant slope of the profile,2) a flat area,3) a "rising area" characterized by a minimum rising slope of the profile.

If one of the criteria is not fulfilled, the edges are considered as edges of an artifactand are eliminated. The resulting image contains only the edges of the flaws. An automaticalgorithm based on segmentation isolates them form the image (Figure 4-b). The detectedflaws are then characterized in term of size, number and location on the weld. On the otherhand, in order to evaluate the intrinsic weld quality (thickness reduction, width...), a"virtual" image of the weld is created. Each point that is considered as a pixel of a flaw isidentified. For each line of the image where points of flaws are detected, an algorithmdetermines the new grey levels of the pixels of the flaw by joining the two boundaries ofthe edges of the flaw by a cubic-spline function. The result of this reconstruction is shownon Figure 4-c.

FIGURE 4. (a) Raw image - (b) Flaws detected by the automatic processing - (c) Final image withoutflaws.

692

Two different treatments allow evaluating separately the flaws (size, number) andthe weld's characteristics (width, thickness reduction, eventual lack of penetration...).

RESULTS, PERFORMANCE EVALUATION AND FUTURE WORK

This method has been tested on several radiographs of welds of thin sheets ofmetal. The results are compared with the interpretation of the expert and, in most cases, allthe flaws are accurately detected. The size and location of flaws in the weld are correctlyestimated by the expert and by the automatic software. In our application, the lateralresolution is evaluated at 50 um and has been validated by comparison with the expertinterpretation for flaws sizing at least 200 um in diameter. The contrast sensitivitycorresponds to approximately 0.1% of the total thickness of the specimen and the detectionof flaws which thickness correspond to 1% of the total thickness of thin sheets has beenvalidated.

A special penetrameter with artificial flaws (simulating thickness reduction andporosity) associated with a depth step wedge has been manufactured. Different shapes ofthickness reduction have been considered (flat, triangular and circular bottom) and holes ofdifferent size and depth (determined by the flaw thickness of interest) are drilled tosimulate voids in the weld. The stepped wedge and the specimen are radiographed on thesame film (Figure 5) and the automatic algorithms are applied (curve calibration, flawdetection and weld evaluation). First results of a study show that all the flaws (lateral sizevarying from 0.6 mm to 2 mm in diameter - depth representing voids from 0.1 mm to 1mm in thin sheets) in the different shapes of weld reduction profiles are correctly detectedand located by the automatic software.

The next step of the validation will consist in evaluating the size of the smallestflaw (in term of dimensions and thickness) that the automatic software can detect. Thisstudy will also use virtual simulated images of flaws in a weld. This action will allow thedetermination of the limit of detection and the accuracy of the automatic treatments.

CONCLUSION

An automatic image analysis software allowing the detection and the quantificationof flaws in a weld of thin sheet has been developed. The principle is based on a perfectknownledge of the relation between grey levels of the digitized image and thickness ofmaterial. Once this relation is established, image-processing algorithms are applied to theraw image for detecting and segmenting flaws. On the other hand, other algorithms areused to evaluate the weld in term of width, of thickness reduction and to detect an eventuallack of penetration. This method has been tested on several welds of thin sheets of metaland has been validated by expert interpretation. Performance will be evaluated on specialpenetrameter or simulated images representing different flaws within different shapes ofwelds.

FIGURE 5. Digitized image of the standard penetrameter.

693

REFERENCES

1. Thickness measurement radiography, Non Destructive Testing Handbook, AmericanSociety for Nondestructive Testing, pp. 817-821.

2. Press, W.H., Flannery, B.P., Teukolsky, S.A. and Vetterling, W.T., Numerical recipes- The Art of Scientific Computing, Cambridge University Press, 1988.

3. Doering, E. R. and Basart, J. P., Trend removal in X-ray images, Review of progress inQuantitative Nondestructive Evaluation, edited by D.O. Thompson and D.E. Chimenti,Plenum Press, Vol. 7A, 1988, pp. 785-794.

4. Canny, J. F., A computational approach to edge detection, EEE Trans. on PAMI, vol. 8,n° 6, pp. 679-698, 1986.

5. Deriche, R., Using Canny's criteria to derive a recursively implemented optimal edgedetector, Int. Journal of Computer Vision, vol. 1, n°2, pp. 167-187, 1987.

6. Cocquerez, J. P. and Philipp, S., Analyse d'images : filtrage et segmentation, Ed.MASSON pp. 120-129.

694