·Home ·Table of Contents ·Computer Processing and Simulation

A new co-operative segmentation method applied to X-ray images Yacine KABIR,Redouane DRAI signal and image processing Laboratory Centre de soudage et contrôle, Route de Dely Ibrahim, B.P.64, Chéraga, Algeria Tel.: (213) (2) 36 18 54 to 56, Tel/Fax: (213) (2) 36 18 50, Email : kabyac@hotmail.com,redouane_drai@hotmail.com Yazid CHERFA Signal and image processing Laboratory Electronic Institute / University of Blida, Route de Soumaa, B. P.270, Blida, Algeria Tel.: (213) (3) 41 58 53, Fax: (213) (3) 41 78 13, Email:cherfa@hotmail.com Contact

ABSTRACT

The aim of this work which enter in a research project on NDT, is to conceive an aid system for interpretation and taking decisions on imperfections in metallic fusion welds. we have studied and tested several segmentation techniques based on the two approaches ( contour and regions ). A quantitative analysis will be applied to extract some relative geometrical parameters. To the sight of these characteristics, a first classification will be possible.

However there is no unique segmentation, for each application, it is necessary to choice the appropriate technique, with the adapted parameters. This study has permitted to conclude that we cannot favour one method in relation to the other one. The basic idea is to use the two approaches in a single process that exploit the advantages of one technique to overcome the drawbacks of the other one. In this paper we propose a new method by combining estimation, enhancement, edge detection and a region growing in a an intelligent process.
keywords: defects, edge detection, filtering, region growing, segmentation, X-rays image.

1. INTRODUCTION

Due to the degraded quality and the small size of the defects, X-ray films are sometimes difficult to inspect. The interpretation of such images is often likely to be affected by a certain subjectivity of the human operator. It seems very interesting to introduce techniques that are operator independent. The aim of this work which enter in a non destructive testing research project, is to conceive an aid system to interpret and take decisions on metallic fusion welds imperfections. The inspection method used is based on X(or Gamma)-radiography, the propagation of excited photons into the material produce an image on a photographic film. The inspected welded areas of industrial sites can present several kinds of defects, they may occur during the fabrication process, or caused in-service by an excessive load, etc.

We present in this paper, the result of a welding defects characterisation technique of welded joints radiogram images. To this end, the image will be segmented in order to put failures in obviousness. Image segmentation is a fundamental stage in any artificial vision system. It allows producing a compact description that can be characterised by edges or by homogeneous regions. Image edges represent a very rich information and have been extensively studied. The "region" approach is complementary to the "contour" approach, however it does not provide the same results. We have studied and tested several segmentation techniques based on the two approaches. There is no unique segmentation, for each application, it's necessary to choice the appropriate technique, with the adapted parameters. Indeed each technique have its advantages, drawbacks and its limitations. The aim of each one is closely related to the application in question.

A new tendency consist of combining these two approaches in order to obtain a robust segmentation by exploiting the advantages of one method to reduce the drawbacks of the other one. We will present a new method by combining estimation, enhancement, edge detection and a region growing in a an intelligent process. A quantitative analysis to extract some relative geometrical parameters will be applied.

To the sight of these characteristics, a first classification will be possible. We proceed as follow:

• Acquisition and digitising of radiogram image .
• pre-processing (restoration, noise reduction, contrast stretching) in order to enhance the quality of the images;
• segmentation defect selection
• defect contour following to extract some geometrical characteristics
• defect nature decision (e.g. linear or spherical, etc. ).

Fig 1: Flow chart of the NDT project

Step 2 and 3 are executed automatically by the new intelligent co-operative segmenter proposed.

Several practical results will be presented and commented.

This study describes one stage of a software tools realisation, which facilitate and improve the task of the expert for x-ray image interpretation and decision-making. We are interested by the first four stages of this flow chart. Defects intervening in pieces are listed by official norms. For segmentation needs, we have divided the set of defects in two categories, volumetric and linear defects. A defect is considered as linear if its width is twice inferior to its size, all the rest are considered as volumetric defects.

2.PROBLEM PRESENTATION

We use the X or gamma rays power penetrating to detect possible heterogeneities in inspected pieces. These rays are absorbed by the matter crossed, essentially by the photoelectrical effect.

Due to the great number of disruptive factors, the visual information provided by an X or gamma ray image is complex. The processed images are characterised by three particular phenomena:

1. leak contrast between the defect and the image bottom ;
2. image bottom granular aspect, perceived as a bottom noise ;
3. gradient bottom image presence characterising the thickness variation of the inspected sample which can harm the small dimensions and bad contrast defects detection .

The importance of these phenomena varies from one image to an other, according to the nature of the metal, the thickness of the piece and the type of radiation employed. For all the reasons described above, we have to perform some operations, described in section 3:

Pieces and defects nature
Inspected pieces are of variable forms and thickness. Risky zones are those submitted to important physical constraints. Two types of pieces are mainly controlled:

welds, levelled or not levelled;

cast pieces, such as elbows.

Defects intervening in pieces are listed by official norms. For segmentation needs, we have divided the set of defects in two categories, volumetric and linear defects. A defect is considered as linear if its width is twice inferior to the size of the grain, all the rest are considered as volumetric defects.

3.IMAGE SEGMENTATION

The segmentation allows to compact the huge amount of information contained in the original image and preserve only significant information. It can be divided into two categories :

Edge detection, when we are interested by object shapes ;

Region segmentation, when we are interested by the image element contents.

The study of the advantages and drawbacks of the different segmentation methods, leads us to the following concluding remarks :

• The edge approach generally gives a localised contours when transitions are abrupt, but always presents discontinuities;
• An edge detection don't satisfy directly the mathematical formalism used to define a theoretic segmentation; it's based on the duality existing between an edge detection and a region segmentation (a contour define a region boundary, and a region is automatically delimited by a closed contour) ;
• A region segmentation use a similarity criteria to detect homogenous regions, nevertheless their boundaries are not sufficiently precise; the choice of a similarity criterion and its parameters can affect considerably the position of a region boundary.
• The extraction of textured regions is well done by a region segmentation.

A comparison between the two approaches (edge and region segmentation) has shown that neither edge or region segmentation can provide alone an ideal segmentation. The future of image analysis is probably in the co-operation between these two approaches. A co-operation technique permit to exploit the advantages of each method in order to reduce the drawback effects of the other one.

4.PROPOSED CO-OPERATIVE METHOD

The segmentation is obtained by the co-operation of a robust edge detector of type canny and a region growing. The originality of this method reside in the introduction of an adaptation stage allowing a better choice of the edge detector parameters ( scale, binarisation thresholds ) and the homogeneity criteria thresholds related in a great part to the textured aspect of the image. We have modified the classical region growing algorithm; the isotropy of the growing processes with the implemented algorithm corrects the imperfection of the recursive older algorithm, this was possible by avoiding the recursive implementation, and replacing it by the use of a queue structure [5] and the integration of an optimization phase in ordering the choice of pixels liable to be aggregated with the current region in progress.
In order to measure the compatibility and the consistency between the edge and the region maps so obtained, and therefore objectively appreciate our result, we use along the whole process of segmentation, a series of performance measures on both edge and region maps.

The segmentation method is performed according to the following stages:

• Estimation stage
  

  • building the local contrast map; image points are classified into 3 categories: zone highly contrasted (ZHC), homogenous zone (HZ) and weakly contrasted zone (WCZ).
  • creation of a map characterizing the homogeneity image surface, this permit the localization of textured zones. Image points are classified into 2 categories:
    uniform zone(UZ), non-uniform zone (NUZ):
  • noise estimation, we consider only 3 types of noise, additive, multiplicative and impulsive .

• image pre-processing
  

  • contrast stretching if the image is poorly contrasted.
  • Image enhancement by filtering, according to the noise nature affecting the image (previously estimated ):
    • additive (r) gaussian filter;
    • multiplicative (r) mean filter ;
    • impulsive(r) median filter

• Construction of a preliminary edge map (with a robust operator), allowing to obtain high precision on image objects border;
  The co-operation process begin by an edge detection ; the obtained edges are well localized, and corresponds precisely to the real region borders. The segmentation quality depends largely on the region boundaries detected. The edge operator can:

  • generate false edges because of textured regions for example.
  • Miss edges due to a bad lighting .

• Small edge elimination (size varies with the application in question)
  The edge detection operator is sensitive to all gray level variation, it can present wrong responses in textured or noisy zones; in fact, this operator don't take in consideration local characteristics; apart from this fact, sometimes, threshold values are not valid for the whole image. It's then very important to eliminate these false small edge chains, while respecting the type of picture to segment, and the nature of mannered objects. At this step, there's a risk to eliminate true edge chains, and in spite of all, some small chains without significance are still not suppressed because of the threhold values used.

  • Performing a region growing segmentation constrained by the edge map and extraction of the regions boundaries; the growing process start from seeds (randomly, extracted from a preliminary region segmentation,....);

  A growing region segmentation, detects better the content of objects in an image, because of the utilization of homogeneity criteria. These criteria permit the detect zones having the same texture attributes. It will be therefore possible to confirm the true contours due to a discontinuity between two distinct adjacent regions, and to eliminate the false edges, due to a sensitivity to micro-textures, to a degraded gray level, or to a shade, etc... . The growing process doesn't evolve beyond contours.

For the aggregation process, we will use a double criteria; the growth is stopped when the pixel candidate to the aggregation:

• Don't verify the similarity criteria (homogeneity)
• coincides with an edge pixel in the edge map.

A pixel verifying the criteria, and coinciding with a contour, is merged with region in progress , in an ulterior stage.
Because of the edge constraint that the process of region growing undergoes; small regions having the measurements of small edge chains will be created

• fusion of small regions, whose sizes, can be adjusted by the operator, or fixed according to the type of the image being treated
  The region growing phase, generate a big number of small regions because of the choice of parameters used. The edge constraint in the last step, can also give birth, to small isolated regions (figure 2).

  Fig 2: Creation of small regions because of the edge constraint, The region growing process bypass the small edge chains and consequently create small regions, that we have to suppress

  
  The size of small regions to be merged, can be chosen according to an a priori content knowledge of the image treated; a certain number of these small regions is due to the edge constraint (the small chains of contours).

  A region is identified as belonging to this false region type, if all its points coincide with edge points. Therefore, We have two types of small regions to be merged:

  • regions generated by the edge constraint;
  • regions generated by the criteria choice
  
  If a small region must be suppressed, it's merged with the most similar adjacent region.

• Verification of the common Boundary between each pair of adjacent regions, in order to confirm their dissimilarity (fusion of over-segmented regions)
  Over-segmented regions are generally owed to an insufficient illumination, to shades, or to degraded of color or light. Several simple and composed criteria are used to operate this verification. For example, two adjacent regions are merged if:

  • the mean gradient along the common boundary is weak (less than a certain threshold);
  • if there is not, sufficiently of edge points, on the common boundary, it is achieved by computing the ratio of the edge pixel number along the boundary by the length of the common boundary, in the most favorable case this ratio is equal to 1

• Elimination of small chains situated far from the boundary of any region;
  It's possible that some false chains persist; they are suppressed if they are completely drowned in homogeneous regions,

  Fig 3: Closing contour example by par prolongation on region boundary

• contour closing of the edge map controlled by the region segmentation map;
  The edge map presents a lot of discontinuous segments, that must be closed. To correct these imperfections, an operation of closing contour is done, the process of closing is performed by prolongation on region boundaries (boundary following), when edges are sufficiently close to this boundary , otherwise the following operation is made according to the image gradient (figure 3).

• Extraction of connected components from the edge map
  In the connected components image, the closed contours delimit the different regions , each one have a unique label. These components constitute in fact, the regions whose attributes can be easily calculated from the original image

• Evaluation of the compatibility between the two maps, edge and region, using dissimilarity measures
  

5. QUANTITATIVE ANALYSIS :

Geometrical features extraction
We have implemented all segmentation techniques described above, in an interactive software. The operator can chose and adapt the appropriate segmentation technique with the desired parameters. Some geometrical features can be extracted using a contour following such as length, width, surface, form, median axes of the selected defect. The selection is easy and can be done interactively using the mouse.

Example :linear or volumic defect

The geometrical measurements previously extracted help the making decision system to decide for example whether the defect is linear or not. This defect discrimination into two categories is considered as a first attempt for defect classification. To this end we define a linearity ratio (R_L). R_L =Length / width.

If R_L is equal or near to "1", the defect is volumic ( ex: porosity), , otherwise it is a linear defect ( ex: cracks, lack of penetration).

6. RESULTS

We have created a specialised library for image processing tools including a wide variety of edge and region segmentation algorithms and image enhancement techniques on an IBM personnel computer, in C++ language. We have tested their efficiency on x-ray welded joints radiograms containing different defect types. The obtained results are very encouraging. This work represents a preliminary decision assistance stage, that will facilitate the expert task in defect detection.

Fig 4: Application of the co-operative segmentation method on x-rays films Presenting several defects. a1) inclusion defect a2) longitudinal crack a3) porosity a4) lack of penetration bi) region maps (i=1,2,34, ) ci) border maps (i=1,2,3,4) dj) edge maps (i=1,2,3,4)

With the proposed co-operative segmentation, we can remark on figure 4 that problems encountered with segmentation methods based only on one approach (edges or regions) are partially resolved, we have a good localization on region borders thanks to the edge constraint introduced on the region growing process; contours are well closed because of the edge prolongation on region borders. The precision on defects form is enhanced by the use of a robust edge detection. The defect characterisation is therefore made efficiently . In other hand, the region growing segmentation process is stopped at the right location even if the merging criteria parameters are not severe.

According to the post-processing carried out, the two results can be used. A saving structure of the result was created in order to enable a suitable exploitation.

7. REFERENCES

1. J.F. Canny. Finding edges and lines in images. MIT Artificial Intel. Lab. Cambridge, MA, REP. AI-TR-720, 1983.
2. J.P Cocquerez et S.Philipp. Analyse d'images: Filtrage et Segmentation, Edition Masson.
3. R. Deriche, « Using Canny's criteria to derive a recursive implemented optimal edge detector », International journal of computer vision, pp. 167-187, 1987.
4. S.Horowitz and T.Pavlidis, Picture Segmentation by a tree Traversal Algorithm, Journal of the Association for Computing Machinery, vol. 23, N°2, April 1976, pp, 368- 388.
5. Y.Kabir. Segmentation d'Images de films de Radiographie Dédiée au Contrôle Non Destructif. Magister thesis. Univ Blida, Algeria . 1999.
6. N.R Pal & S..K Pal A View on Image Segmentation Techniques; Pattern Recognition, 1993 Vol. 26, N° 9 : 1277-1294
7. Y.Cherfa, Y.Kabir, R.Drai.. Welding defects analysis by image processing . IEEE-SMC. CESA'98 IMACS Multiconference - April 1-4, 1998 Nabeul-Hammamet, TUNISIE.
8. Y.Cherfa, Y.Kabir,R.Drai. X - Rays Image Segmentation for NDT of Welding defects. 7^th ECNDT'98, May 26-29, 1998, Copenhagen, DENMARK

© AIPnD , created by NDT.net

|Home|    |Top|