NDT.net - June 2000, Vol. 5 No. 06 |
TABLE OF CONTENTS |
Papadakis pointed out in 1991 that signal processing and information theory methods were neglected in nondestructive testing (NDT). During the 1990's NDT laboratories, both academic and industrial evaluated signal processing tools or image processing methods. Research extended to artificial intelligence, neural networks, and even more recently to data fusion.
We shall limit this article to the extraction of information from an image or a signal. The LCND laboratory contributes to develop segmentation methods adapted to ultrasonic, eddy current and radioscopic images.
In the first part we shall present defect detection in thick austenitic weld (80 mm) through ultrasonic imaging. The DTVG method (Delta Temporel Variance des Gradients) concurrently exploits two characteristics of the defect in the image : short temporal response (DT criterium) and spatial stability of the echo (criterium of the variance of temporal gradients VG).
The next part concerns automatic thresholding technique using cooccurrence matrix (bidimensionnal histogram). The histogram of an NDT image is often unimodal and does not allow to clearly define a threshold. It is particularly true for ultrasonic images. In numerous cases we can show that bi-dimensional histograms enable to separate defects from noise zones. Very interesting results are obtained with ultrasonic images (BSCAN images) and with eddy current images (CSCAN images).
A third part shows how spectral analysis applied to ultrasonic signal could be used to determine the concrete setting time. Correlation with migrographies shows that the propagation of ultrasounds is modified by the continuous evolution of the microstructure (ettringite, CSH, etc..) during the concrete setting.
The last example describes a method that combine radioscopic images and retroprojection algorithm to permit a solution to test a circular weld. The inner part of the welded component is too absorbent to allow radiographic testing.
The originality of our method consists in taking into account, to solve the problem of material control, the exploitation of a set of information gathered in an image form. We thus resort to criteria originating in the physics of the methods used, to which image processing methods can be applied. This fact allows to establish a link in fully expanding sciences: image processing, mainly used in satellite or medical imaging, and mechanics.
We present, as an illustration of these methods, four examples of digital processes of images obtained by different nondestructive testing solutions.
The first criterion (VG) studies the spatial stability of the maximum temporal gradient for each signal of the ultrasonic image. The variance of the position of the highest gradient is calculated. It implies to choose a window size in which variance is calculated. A defect zone has a higher stability (VG is small) than a noise zone. The second stage consists in getting interested in the vertical distribution of the grey levels: a defect is expected as a determinist and short temporal event, corresponding to a sharp reflection on an interface. The second criterion (DT) takes into account the time difference between the maximum amplitude of each signal and this value reduced by 6dB. In the presence of a defect, the temporal distance is relatively small, whereas it becames important in the presence of a noise as the algorithm selects automatically among the possible solutions which gives the maximum temporal distance.
In order to automate the segmentation and to obtain a binary decision : defect or no defect, two thresholds are automatically calculated from a reference image. This reference image is a noise image where no defects are present. So we obtain the limit of the criterium DT, it means the lowest value possible of the temporal distance DT in the case of an image of noise. Similarly the lowest value of the variance in this image determines the reference level to know if there is a defect or not, using VG criterium [1]. The calculation of DTVG is then a binary one: DTVG is equal to 1 if VG and DT are simultaneously below the two reference limits, which means that a defect is likely to be present. Otherwise, DTVG is null. Method can process images with any kind of incidence (null or oblique). The binary image is shown to the operator, black events corresponding to DTVG =1 (Figures 1 and 2).
Fig 1: BSCAN image | Fig 2: Segmented BSCAN image |
These figures show results of the complete treatment for 3 defects (holes) in an austenoferritic steel sample with a 80 mm thickness. After DTVG processing, the structural noise is cancelled and defects appear unambiguously. Moreover, the complete automation of the method is appreciated.
We use the definition of the matrix exploiting a vector "" of modulus r and angle q. The cooccurrence matrix indicates the distribution of couples of pixels of i and j amplitudes separated by the vector . We build a symmetric matrix. A coefficient of the matrix (C_{ij}) can thus be considered as the probability of obtaining two pixels separated by such that their amplitudes are equal to i and j.
Our ultrasonic images are built from sinusoidal type signals. For a coding between 0 and 255, the greater part of the pixels are close to the mean value 128. The coefficients C_{ij }which correspond to noise to noise couples are thus in the center of the matrix. The vector d is chosen to build a star-like shaped matrix. In the case of filtered signal (positive values only) the matrix has a L-shape (figure 6). Coefficients representative of the couples that contain at least one pixel of defect are along the branches of the matrix and noise is concentrated in the center of the matrix for sinusoïdal signals or in the upper-left corner for filtered signals (figure 6).
A threshold t inside the matrix separates C_{ij} coefficients below or above the threshold. The idea is to follow the evolution of the coefficients in the matrix in relation to the value of t and to find the transition between the center of the matrix and the branches. In order to follow this evolution, we test several measures and we select the Square of Mean Distance to the Center of Gravity Measure (abbreviation DMB) which quantifies the mean dispersion of the coefficients C_{ij}. This method is at present validated on images which have at least one distinguishable echo. The modeling has shown that it needs one echo of SNR higher than 1.6. The choice of vector could be done visually by an operator but a method is also defined to automatically calculated the best choice of vector d thanks to the analyse of the matrix shape for several increasing vectors [3]. On the other hand, when the method is valid, the threshold value is a little higher than the noise and so allows to locate lower amplitude defects. The method can therefore be improved to obtain a satisfactory detection threshold on volumes which contain only low amplitude defects. The problem is to work on matrix shape analysis.
We have shown with continuous studies that the thresholding method using cooccurrence matrices adapts very well to 3D ultrasonic volumes. We have developed a global thresholding method of a volume based on the cumulated cooccurrence matrix: one matrix cumulates the results on all the planes (BSCAN images) [3]. The results correspond to the expectations of qualified operators. This technique is a very interesting contribution to the automatic thresholding of data volumes. The two figures below concern a 3D volume thresholded represented by its plan view (CSCAN)
Fig 3: Initial CSCAN image | Fig 4: Thresholded image ( t=43) |
The generalization of this method enables to process eddy current images. Data is presented as CSCAN with positive values (figure 5).
Fig 5: Eddy current image with four defects | Fig 6: Cooccurrence matrix calculated on image of figure 5 (d=[80 0]) |
It is furthermore possible to improve the thresholding strength through the definition of a cooperation rule between two threshold assessment measures. This rule is also used to reduce time of automated processing by limiting the number of vectors . These measures are calculated from the image cooccurrence matrix. The aim is to build a thresholding method that gives a threshold value but also an indicator which allows the operator to know if the methodology well applies to the processed image.
The indicator is set by learning process. Two thresholds are given by two different measures, and the distribution of coefficients in the matrix is also analysed to validate the L-shape. A validation domain is then set when there is a low difference between the two thresholds and when the degree of confidence in the matrix shape is high. In the figure 7 the limits are defined from 56 images. Thresholding solution is validated if threshold difference is lower than 10 and if degree of confidence is greater than 95% [4].
Fig 7: x axis : thresholds difference y axis : degree of confidence Validation Domain | Fig 8: Validated thresholded image (t = 80) |
This method could be adapted to process radioscopic images as the noise distribution in this kind of images also gives unimodal histogram.
Fig 9: A. image time-PSD. B. PSD at time 70 hours. PSD image of the transmitted signal |
The frequential evolution was then translated into a two-dimensional "frequency - setting time" image in which the amplitude of the Power Spectral Density (PSD) is represented according to a grey level scale. The image analysis brings to the fore the filtering effect obtained in relation to the appearance of the various constituents and the bridges linking the grains. It thus allows us to appreciate the evolution in the concrete setting time.
Scanning electron microscope (SEM) is used to observe more precisely the concrete's microstructure at different moments. A good correlation is obtained between the apparition or disparition of the constituents and the maximum of DSP values determined over the 15-100 kHz bandwidth [5]. The setting time is characterized by an important formation of the CSH constituent (calcium silicate) and this corresponds to the point TP162 of the curve in figure 10. This point could be easily identified by an operator.
Fig 10: Time evolution of the DSP maximum |
Fig 11: Massive formation of CSH and ettringite
G = 3000, time = 9h10
Example of micrographie close to the setting time (TP162) |
Fig 12: Schematic view of the radioscopic inspection system for one angular position | Fig 13: Simulated image of the radioscopic image |
Figure 13 shows a simulated image. The model uses parallel beam as images are obtained with a microfocus tube. Two attenuation coefficients should be known : one for the parent metal, one for the melted zone. We need to know the response curve of the detection system. Complete description of this model can be found in Gueudré and al [8]. The simulated image in figure 13 is very similar to the real one and we can see that very poor contrast is obtained at the bottom of the weld (left side of the image).
Classical filtered retroprojection algorithms [9] used in medical scanners or industrial ones need to have complete projections for a large panel of angle. Figure 12 shows that we have only partial projections of the whole piece as the inner part is too absorbent. We have to construct complete projections by using the partial projection at angle q and the one at angle q +p (the same simulated partial projection fact). Between the two partial projections we complete by zero values and we use an apodisation function to smooth the transitions. In figure 14 we show reconstructed images. The weld shape is well perceived for it appears at an homogeneous zone. Reconstruction artefacts (vertical lines) due to incomplete projections are also clearly seen. The reduction of these artefacts is effective as the algorithm uses apodisation near the limits of the projections [6]. Figure 15 illustrates the increase in contrast in the bottom of the weld. Dotted line is the profile obtained in the radioscopic image, that is the classical image used in nondestructive testing.
Fig 14: Recontruction 1 : 1000 projections without apodisation Reconstruction 2 : 4000 projections with apodisation Two half-reconstructed images. | Fig 15: ligne (b) of figure 14 in the case of reconstruction 2 Normalized central ligne profiles of the weld. |
Reconstructed profile shows a much better transition at the left limit of the weld (indicated by the vertical line in figure 15), but reconstruction is quite noisy due to incomplete projections.
So automatic thresholding solution could be difficult. Nevertheless the combination of radioscopic images and retroprojection algorithms adapted to the hollow projections enables to have a good segmentation solution.
© NDT.net - info@ndt.net | |Top| |