Multiphase PIV Method with Digital Object Separation ...

Report 4 Downloads 40 Views
5th International Symposium on Particle Image Velocimetry Busan, Korea, September 22-24, 2003

PIV’03 Paper 3249

Multiphase PIV Method with Digital Object Separation Methods Markus Honkanen1, Pentti Saarenrinne Abstract This paper describes some digital object separation methods for PIV images of multiphase flow. The methods enable multiphase-PIV measurements in denser multiphase flows with an increased reliability and accuracy of the measurement results. This paper focuses on PIV images of turbulent bubbly flow. Methods are also applied to PIV images with spray droplets and with flocculent particles. Four different digital object separation methods are developed in this study, namely Probability-of-centre, Convex-perimeter, Curvatureprofile and Shen’s method. All of these methods utilize information related to the perimeter of the segment. Probability-of-centre method searches segment points of local maximum for the probability of being the centre of the segment. The local maxima correspond to the number of individual objects in the segment. The last three methods are called breakpoint detection methods, because they search for the connecting points of outlines of individual objects on the perimeter of the segment. The performance of these methods is analyzed and the benefits of using an object separation method are emphasized. An independent velocity and size is measured to all separated bubbles in the bubble group. The improved multiphase-PIV method is applied to turbulent bubbly flow in a mixing vessel with a diameter of 260 mm. Introduction In various industrial process applications the gas void-fractions of the bubbly flow exceed the upper limit of PIV measurements. The upper limit of the measurements should be raised towards denser bubbly flows, if industrial processes are studied. In the PIV image of dense bubbly flow, a considerable number of detected segments are created by a group of bubbles and not by individual bubbles. If two or more bubbles overlap in the image, they are detected as one big bubble and only one velocity and size is measured for the whole group. This causes errors in measured size and velocity distributions of bubbles. Thus, a robust object recognition method equipped with digital object separation methods is needed. The aim of the study is to improve the PIV and digital imaging method to study dense bubbly flows and to measure the properties of bubbles in a turbulent bubbly flow. Measured quantities are velocity fields and relative velocities of the continuous phase and bubbles, as well as the sizes and shapes of the bubbles. The challenges of the study relate to many subjects. Main problems arise with (1) the optical imaging of the flow, (2) the discrimination between two phases, (3) the noise removal due to the disturbed optical access to the flow, (4) the correct measurement of the sizes of the bubble images and (4) relating these measures to real bubble size distribution in the flow, (5) the correct measurement of bubble velocity and (6) the correct comparison between the measured fluid flow field and the bubble velocity. The interaction between the tracer particles and bubbles in the fluid has also to be taken into count. These problems are already considered in previous studies (Honkanen et al. 2002a,b,c and 2003). Now we concentrate on problems (3), (4) and (5). The size of a single, noiseless bubble image can be measured with high accuracy. The accuracy decreases when noise increases and the percentage of overlapping bubble images increases. The accuracy of measured bubble size distribution can be increased by measuring a larger sample of bubbles, but how accurately the distribution can be measured? Gouriet et al. (2002) detected the overlapping bubble segments and simply divided the total area of the segment by the number of involved bubbles to get the size of a single bubble. They concluded that a statistical approach smoothes method’s detection error, observing that overlapped bubbles have often close sizes. However, in their measurements, only 5 % of the bubble segments consisted of many bubbles.

Markus Honkanen1, Pentti Saarenrinne Energy and Process Engineering, Tampere University of Technology, Finland Correspondence to: Postal address: TUT, Energy and Process Engineering, P.O.Box 589, FIN-33101 Tampere, Finland E-mail: [email protected], phone: +358 50 3020828.

1

PIV’03 Paper 3249

Oberdier (1984) has shown that the depth of field increases with the dispersed particle diameter. The larger particles appear to be in focus over a thicker measurement volume than smaller particles. This will result in the measured size distribution becoming biased towards larger particles. In addition, the images of large particles have higher probability to overlap with other particle images, so the measured sizes of large particles are often also overestimated if the digital object separation methods are not used. On the other hand, the large particles have higher probability to fall partly behind the un-focused objects that locate in front of the measurement plane. Thus, image’s effective measurement area for detecting the large particles is smaller than for small particles. An accurate object recognition algorithm does not guarantee physically realistic measurement results. One has to also pay attention to the quality of the measurement technique, data preprocessing and validation of measured data. The novel two-phase PIV measurement technique is presented by Lindken and Merzkirch (2002). Here the technique is applied to study of turbulent bubbly flow in a mixing vessel. The pulsed back lighting (with a diodelaser, 808 nm) is used to detect bubble outlines and the (Nd-YAG) laser light sheet is produced to illuminate the tracer particles. Images are recorded with only one camera. The fluid is seeded with tracer particles, which have a fluorescent dye that can be excited by the illumination source. The camera is equipped with an optical edge filter (>570 nm). Therefore, the camera records only the light scattered by fluorescent tracer particles and the shadows of the bubbles. The light reflected by bubbles is totally filtered.

(b)

(a) Figure 1. a) A bubbly flow PIV-image with a fluid flow vector field. The result of the curvature-profile method is shown in Figure b) with separation lines and c) with fitted ellipses.

(c)

Digital image processing methods are applied to the images in order to discriminate between the two phases of the flow, to enhance the PIV-vector field of the fluid flow and to measure the correct size distribution of bubbles. The example results of bubbly flow in a mixing vessel can be seen in Figure 1. All the recorded bubbly flow images need lots of digital image pre-processing to correctly detect the bubbles and to obtain accurate vector fields of the fluid flow. Bubble shadows and tracer particle images are discriminated based on differences in size and brightness. In order to recognize the bubble shadows, the intensity values of the image higher than the median intensity level are set to the median value. A 2D-median filter (presented by Kiger et al. 2000) is applied to the image and the image is then converted by taking a natural logarithm of the image. A nonhomogenous gas density in the flow causes large changes in the background intensity of the image. The background intensity of the image is equalized effectively by subtracting the sliding background of the logarithmic image. Wu et al. (2002) subtracted the so-called logarithmic background and proved this to provide good results. Recognition of bubbles is difficult because they form random ellipse-like shapes and they have a wide size distribution (in this study from 0.04 mm to 5.0 mm). Correlating an ellipse or a reference bubble image to bubbly flow images has not been successful. Only the basic bubble shapes have been detected. When bubbles are recognized in the image by grey scale thresholding, all the shadow images (in-focus and out-of-focus images) are detected as bubbles. Chigier (1991) presented a method to measure the size of out-of-focus bubbles correctly. In this study only the in-focus bubble images should be recognized. In order to measure a correct relative velocity for a bubble, the bubble must locate centrally in the direction of depth in the measurement plane. When

2

PIV’03 Paper 3249

back lighting is used, the depth of field of the image is controlled by camera aperture. Low f-number keeps the effective thickness of the measurement plane in less than 9 mm in case of bubbly flow with a bubble size range 0.1 - 4 mm (/Honkanen et al. 2002c). The employed object recognition method uses the grey scale information and the local gradient of the grey scales as parameters in the recognition procedure. The local grey scale gradients, i.e. sharp edges of bubble shadows, are detected with a digital 5x5 Sobel filter (/Oberdier, 1984). The images are thresholded by grey scale values and by local grey scale gradient values. A threshold value for the local grey scale gradient is determined by criteria for in-focus bubble image. The information given by greyscale thresholding and by thresholding of local grey scale gradient is combined and only the segments that satisfy both criteria are detected as in-focus bubbles. The thresholding of grey scale gradient is done for each partial perimeter in order to prevent false detections of high grey scale gradients at the middle of the bubble image, which can be caused by residuals of tracer particle images or by penetration of light through the bubble. Digital object separation methods Digital object separation methods look for image segments that consist of many objects and study the overlapping objects individually. If the segment is too complex and objects cannot be recognized, the segment is rejected in order to avoid false recognitions. Four different digital object separation methods are developed in this study, namely Probability-of-centre, Convex-perimeter, Curvature-profile and Shen’s method. These object separation methods utilize the results of the presented object recognition method. The following quantities of detected segment can be recognized: segment area, list of pixels in the segment, bounding box, centroid, perimeter (i.e. segment border line), convex line, solidity (i.e. the ratio of segment area and the area inside the convex line), principal axes, orientation and eccentricity. First of all, all the object separation methods need the information about the perimeter of the segment. The perimeter is presented as a list of x- and y-coordinates of the pixels in the perimeter. In addition, the Convexperimeter method needs the information about the convex line and Probability-of-centre method needs the list of all pixels (b) inside the segment. Thus, the Curvature-profile method and Shen’s method demand the least amount of initial information of the segment. The perimeter of the segment is detected with a 8-point-connectivity algorithm that follows the segment border in the binary image so that the segment has value 1 and other image parts are 0. The resulted perimeter is coarse and the saw-toothed line has to be smoothed with a FFT-filter. FFT-filter can effectively eliminate the high frequencies of the perimeter resulting in a smooth perimeter. The number of preserved frequencies depends on the length of the perimeter. After defining the segment and its perimeter, the procedure continues differently in different object separation methods. Probability-of-centre method Probability-of-centre method searches segment points of local maximum for the probability of being the centre of the segment. The local maxima correspond to the number of individual objects in the segment. The probability of being the centre of a segment is measured for each point inside the segment. If more than one local maximum for the probability appears, the segment consists of more than one object. The probability is calculated from the standard deviation of the distance of the specific pixel to each pixel on the perimeter. When the standard deviation becomes small the probability increases. The surface-plot is used for detecting the local minima of the standard deviation. The method is very sensitive to noise. The shadows of the bubbles in the image have bright holes in the centre causing false recognitions. The computation speed of Probability-of-centre method depends on the size of the segments. If the image consists of many large segments, the computation time may exceed over 2 minutes per image. If the segments are small, the Probability-of-centre method is still at least three-times slower than the Convex-perimeter method, for example. The method detects the each object’s centroid and the diameter of the equivalent circle or it can divide the segment area equally between the detected objects in the segment. However, all the breakpoint detection methods are found to be more efficient and more reliable than this method. Breakpoint detection methods The last three methods are called breakpoint detection methods because they search for the connecting points of outlines of individual objects on the perimeter of the segment. Breakpoint detection methods employ following procedures. 1. detecting the perimeter of the segment 2. locating the connecting points at the perimeter 3. a) pairing the connecting points and separating the segments or b) fitting a sphere or an ellipse to the arcs/parts of the perimeter separated by connecting points 4. calculating object centroids and matching object pairs with a PTV algorithm

3

PIV’03 Paper 3249

5. 6.

averaging object properties over two image frames calculating rotation parameter, defined as the difference of object orientation angles between the image frames The second procedure, locating the connecting points at the perimeter, is the most critical procedure in the object separation. Procedure can be done in different ways. Three methods with three different techniques for locating the connecting points are presented. They are Convex-perimeter method, Curvature-profile method and the Shen’s method. Berg et al. (2001) has also presented a new breakpoint detection method that finds the connecting points by calculating the centres of gravity for the areas over edge pixels in the binary image. Other five object separation procedures can be utilized in all presented methods. Convex-perimeter method The Convex-perimeter method has been developed at Tampere University of Technology and firstly presented in source (Honkanen 2002a). In the Convex-perimeter method, a segment is bounded with a convex line by minimizing the length of the line. The convex line is detected similarly as the perimeter of the segment. If the solidity of a segment (i.e. the ratio of the segment area and the area inside the convex line) is less than the solidity-threshold value i.e. ~0.91-0.92, the segment must include more than one bubble. The solidity check separates effectively the object groups from the single objects. If the segment consists of a group of objects, the procedure continues and the convex line is compared with the perimeter of the segment. The difference between the perimeter and the convex line is measured at every point on the perimeter by measuring the minimal distance of convexline to that point. If the difference is greater than the given threshold value at any point on the perimeter, it is assumed that the segment includes more than one bubble. The local maximums in difference between the perimeter and the convex line represent the connecting points of the bubbles, so long as they are above the given threshold level (threshold ~ 5 pixels). The threshold can be scaled with the segment’s size. There cannot be two connecting points in the perimeter close to each other. Thus, a threshold limit is given for the closeness too. The connecting points can be very close to each other in the physical space, but they cannot have close indices on the perimeter line. If more than one connecting point is detected, the procedure continues with finding the pairs from the group of connecting points. In some cases, when bubble segments are very irregularly shaped, the Figure 2. A bubble image that minimal distance to the convex line corresponds to convex line points that is analyzed with the Convexare on the wrong side of the bubble segment. In this case the connecting perimeter method. The convex point is detected wrongly. This can be seen in Figure 2. The shape of the line is shown in blue color and bubble segment on left is bent and wrong connecting points are detected. the perimeter in green color. Curvature-profile method Pla (1995) recognized partial circular shapes from segmented contours by a breakpoint detection method that utilizes the curvature data of the perimeter of a segment. A same kind of breakpoint detection method is developed in this study, referred to here as Curvature-profile method. The perimeter of a segment is listed as xand y-coordinates (xp(i) and yp(i)) having the origin in the centre point of a segment. The slope of the perimeter is measured in radians as:

 y p (i + 1) - y p (i - 1)  .  (x (i + 1) x (i 1) p   p

α = arctan 

The arctangent function returns angles in range (-π,π) and other angles are wrapped around, resulting in an artificial discontinuity of the slope. The slope is normalized, thereby correcting the discontinuities between two consecutive points greater than π or less than -π. Curvature of the perimeter is the derivative of the slope (Dα). The peaks in the curvature line correspond to the points of sudden changes of the perimeter slope. The negative curvature peaks (i.e. local minima) correspond to the points where the slope turns away from the segment centre. These points are usually the connecting points of two overlapping objects. Pla (1995) described curvature extrema (local minima and maxima) as connecting points, but in most cases the connecting points are the local minima of curvature. The other description, i.e. local maxima, increases the number of wrong detections dramatically. The perimeter of a smaller segment tends to have bigger values of curvature than a bigger segment. Thus, the threshold value of curvature is scaled with the RMS-value of the curvature, which is defined as:

4

PIV’03 Paper 3249

n

Dα RMS =

∑ ( Dα i =1

i

− Dα ) 2

n

.

If the segment consists of many bubbles, the minimum value of curvature must be below the minus RMSvalue of the curvature and also below the threshold value, i.e. –0.1. The point is detected as a connecting point if the curvature at the point is less than in the neighbouring points and it is less than the given threshold value in that point and in the closest neighbouring points. The presented Curvature-profile method detects the connecting points correctly and reliably, but problems may occur if there is a lot of noise in the perimeter data. The noise can be subtracted with the FFT-filter, but then real information can also be lost. Figure 3 shows the data of an example bubble image, shown on left in Figure 1. The perimeter’s slope and the curvature are plotted a) without filtering and b) after using the FFT-filter. The benefits of smoothing the perimeter are noticeable.

(a)

(b)

Figure 3. a) The raw perimeter data (slope and curvature) and b) the smoothed perimeter data. Shen’s method Shen et al. (2000) criticized the Pla’s “curvature-profile” method and stated that the curvature of a curve cannot be calculated accurately. Curvature calculation is sensitive to noise and therefore the perimeter has to be smoothed. The smoothing process makes the corners on the perimeter blunt and more difficult to detect (/Shen et al. 2000). Instead of calculating the curvature of a smoothed perimeter, the connecting points are searched from a raw perimeter data. A connecting point satisfies the following conditions:

( xi −1 ≤ xi ≥ xi +1 ) ∪ ( xi −1 ≥ xi ≤ xi +1 ) ∪ ( y i −1 ≤ yi ≥ yi +1 ) ∪ ( yi −1 ≥ y i ≤ yi +1 )

In other words, the connecting point must be a local extremum point. There are also other extremum points in the perimeter. If the perimeter is rotated by a small angle step continuously from 0° to 360°, the true connecting points appear as local extremum points more often than other points. According to Shen’s numerical experiments for a connecting point, its probability of being a local extremum point is about four to five times of the average probability. If the perimeter data (xp,yp) is rotated by an angle θ, the new coordinates (xp’, yp’) are given by

x p ' = x p cos θ − y p sin θ

y p ' = x p sin θ + y p cos θ where θ goes from 0° to 360°. The local extremum points are calculated for all rotated coordinates (xp’,yp’). Shen’s method locates the connecting points very well, if the detected objects are circular, but for other objects with ellipse-like or irregular shapes, the method determines all the corner points on the perimeter as connecting points. In order to improve the Shen’s method, the curvature has to be calculated at the located connecting points and only the connecting points with a negative curvature are selected. This way the Shen’s method gives the same results as the Curvature-profile method, but its computation time is significantly longer. There are two ways to proceed: a) pairing the connecting points and separating the segments or b) fitting a sphere or an ellipse to the arcs of the perimeter separated by connecting points.

5

PIV’03 Paper 3249

Pairing connecting points and separating the segment parts After locating the connecting points, the connecting points can be paired and the segment parts can be separated from each other with separation lines. It is supposed that there can be only one pair for each connecting point. Whenever a corresponding connecting point is found for a connecting point, the points are marked as “matched”, they will be connected with line that separates the segment in two (i.e. separation line) and they won’t be used in the pairing process again. The slopes of all the possible separation lines are measured. If the Convex-perimeter method is utilized, the normal of the convex line and the normal and tangent lines of the perimeter are measured for all connecting points. By comparing this information the correct pairs of connecting points can be found. The connecting points are paired if one of following 5 criteria is satisfied. In every case, the length of the separation line must be less than the minor axis length of a segment. 1. The slope of the separation line differs less than the given angle-threshold from the normal of the convex line in both connecting points. 2. The slope of the separation line differs less than the given angle-threshold from the normal of the perimeter in both connecting points. 3. The slope of the separation line differs less than the given angle-threshold from the normal of the perimeter in first connecting point and from the normal of the convex line in second connecting point. 4. The tangent of the convex line in first connecting point differs less than given angle-threshold from the normal of the convex line in second connecting point. 5. The separation line must be shorter than length-threshold and less than half of the minor axis length of a segment. The Curvature-profile method and Shen’s method pair the connecting points exactly like the Convexperimeter method, but only the criteria 2 and 5 can be used. There can be also a criterion that, if there are more than five unmatched connecting points in the segment, the segment is discarded. The optimal threshold values for all the criteria can vary over the set of images and even locally in each image. The angle-threshold values can be around 20-25 degrees and length-threshold value around 40 pixels, depending on the object size. Instead of thresholding, a least square method can be used to pair the connecting points in a way that the residual error is minimized. Error is calculated by fitting the data to the above criteria. The data of all separated parts of the segment is measured by cutting the binarised segment, with a value of 1, with separation lines, with a value of 0. The total area of the detected segment will decrease a bit because the areas of the one-pixel-thick separation lines do not belong to any parts of the separated segment. All the separated parts of the segment are labeled and their properties are measured. This method always underestimates the size of the overlapping objects because the overlapping object regions behind, or on top of, the other object are cut off. On the other hand, this method does not assume any particular shape for the object, so all shapes of objects are recognized correctly. If the image segments are very large and complex, the object separation Figure 4. Some example images of algorithm can be repeated for each and every separated part of the overlapping bubbles that are segment, thus ensuring that the separation is done correctly and the separated with separation lines. separated parts really represent individual in-focus bubbles. Figure 4 shows some overlapping bubble images that are separated with the presented method. Fitting a sphere or an ellipse to the arcs of the perimeter Once the perimeter of the studied segment has been divided into parts with connecting points, an ellipse or a sphere is fitted to each perimeter part. Small bubbles are spherical and large bubbles form ellipse-like shapes. If the arc length is less than 1 mm, a sphere is fitted with a Hough transform based method. A sphere is fitted only if the resultant sphere is less than 1 mm in diameter. Otherwise an ellipse is fitted using a direct least square ellipse-fitting method, presented by Fitzgibbon et al. (1999). The method is computationally efficient and highly robust to noise. The method provides reasonable results in all circumstances. The area correlation method presented by Shen et al. (2000), is used for clustering the fitted ellipses and spheres that belong to the same object. The area correlation method defines an area correlation coefficient Cs as follows:

6

PIV’03 Paper 3249

Cs =

S0 , S1 ⋅ S 2

where S1 and S2 are the areas of fitted ellipses or spheres and S0 is their overlapping area. If Cs is larger than the threshold value (0.85), the ellipses/spheres are clustered and the larger ellipse/sphere is selected to represent the object. The method needs only one dimensionless parameter and this parameter is scale invariant (/Shen et al. 2000). The method works also well for perimeter arcs with a wide range of size variation. The fitted objects are validated by comparing the original segment area to the total area of objects detected from the segment. If the area of objects is more than twice the original segment area, the segment results are rejected. This method can measure the bubble sizes correctly even if the bubble images are overlapping 50 %. If the PIV image is full of overlapping bubble images, the total area of detected bubbles can be even larger than the field of view of the image but, in these kind of conditions, PIV measurements of fluid flow become impossible. The disadvantage of the method is that it is sensitive to false connecting points. If there are many wrongly detected connecting points in the perimeter data, the perimeter is divided into too many arcs and the single objects might be falsely divided into parts. This can be seen for example in Figure 1c. The noise on the perimeter data has to be removed and the locating-process of connecting points has to be done carefully using proper threshold values. Figure 5 shows some successful results.

Figure 5. Some example images of overlapping bubbles with ellipses fitted on them. Performance of the methods By comparing the presented four digital object separation methods, we found out that the Curvature-profile method locates the connecting points most accurately and most reliably. It is also least sensitive to noise. Shen’s method is the best approach if spherical bubbles are studied. The Probability-of-centre method performed the worst. If the studied bubbles are larger than few millimeters, it is recommended to pair the connecting points and separate the segments instead of fitting a sphere or an ellipse to the separated arcs of the perimeter. However, for small bubbles the ellipse-fitting procedure gives very accurate results. The presented breakpoint detection methods have a good tolerance against noise, brightness deviations and bright regions inside the shadow, but grey scale threshold value should be selected carefully, so that the out-of-focus images are left out from the analysis. A small sample of PIV-images of bubbles rising in a stagnant water are studied. 25 % of detected bubbles are overlapping in the images. There are 44 groups of overlapping bubbles in total. The Curvature-profile method has detected 36 groups of bubbles and eight groups remain undetected. The method has detected none of the individual bubbles falsely as a group of bubbles and it has correctly separated 30 groups i.e. 68 % of the bubble groups. Figure 6 shows the results for the sample images, i.e. the size distribution of bubbles and average velocities of bubbles in different size classes. If a basic bubble recognition method is used, the computation time is much less. More images can be analyzed during the time, when the Curvature-profile method analyses the small sample. Figure 7 shows the results of the same set of images, calculated with a basic bubble recognition method, but more images are analyzed. The same computation time is consumed in the both analysis. The results show the benefits and disadvantages of using an object separation method. The basic bubble recognition method can analyze much more samples and increase its accuracy by statistical approach, but still some errors remain. The axial velocities of bubbles in Figure 6 correspond to the values from literature (Wallis et al. 1969), while the axial velocities analyzed with the basic bubble recognition method (in Figure 7) deviate much more. The Shen’s method with an ellipse-fitting procedure is applied to the dense bubbly flow image of CO2bubbles in a butanol flow. The results can be seen in Figure 8.

7

PIV’03 Paper 3249

Figure 6. The size distribution and average axial velocities of bubbles in different size classes calculated with the Curvature-profile method.

Figure 7. The size distribution and average axial velocities of bubbles in different size classes calculated with the basic bubble recognition method, with more samples than in Figure 9.

Figure 8. An image of a dense bubbly flow and the bubbles detected by an object separation method. Precision of dispersed particle’s velocity measurement methods The precision of different measurement methods of velocity of a dispersed particle is studied with simulated multiphase flow images consisting of a dispersed particle shadow and bright tracer particles on top of it. Cross correlation based FFT-CC method, simple centroiding method and intensity weighted centroiding method are presented. The centroiding methods find the centroid of the bubble by a) simply measuring the average of the x-

8

PIV’03 Paper 3249

and y-coordinates of the pixels inside the segment and b) by weighting the edge pixel coordinates with the normalized intensity of the pixel. The presented velocity measurement methods attain sub-pixel accuracy. Four main features that affect the measured velocity precision are background noise i.e. tracers, particle size, particle displacement and particle image contrast. The simulation shows that the centroiding methods give inaccurate results for small particles with a diameter less than about 10 pixels. The centroiding methods are not as accurate as the FFT-CC method, but the intensity weighting improves the accuracy of centroiding. For images without background noise the precision error of FFT-CC method remains under 0.015 pixels, the error of intensity weighted centroiding remains under 0.03 pixels and the error of simple centroiding remains under 0.08 pixels, unless the particle is not very small (D~4 pixels) or too large for the interrogation window. Clear pixel locking effects are not detected with varying particle displacements. The bright tracer particle images on the top of the particle shadow edges deform the particle shadow and introduce extra movement of the shadow. Intensity weighted centroiding emphasize the intensities of the edge pixels. When the tracers disturb the edge pixels, the methods give inaccurate results.

Figure 9. The top figure shows the PIV-image as its original form (two-frames overlayed) including the results of a basic bubble recognition method. The bubble velocities are measured with FFT-CC method. Bottom figures represent first and second frame of the PIV-image after image pre-processing and detecting the individual bubbles with object recognition method equipped with the Curvature-profile method. The bubble velocities are measured with a simple centroiding method. The difference between the first and second analysis is huge. If the digital object separation is employed, the individual bubbles are detected more reliably and their velocities are calculated more accurately. If there are many dispersed particle shadows in the interrogation window, the centroiding methods can measure the particle displacements more accurately than FFT-CC method. The extra particles in the image do not only decrease the accuracy of the velocity measurement with a FFT-CC method, but they also decrease the reliability of the measurement decreasing the correct correlation peak height, especially if the studied particle has a low contrast to the background. Figure 9 emphasizes the difference in performance of FFT-CC and centroiding

9

PIV’03 Paper 3249

method. In the study of dense bubbly flow, it is recommended to use particle centroiding methods because the cross-correlation based dispersed particle velocity measurement method do not work in dense multiphase flows. The accuracy of centroiding methods in the measurements of dense bubbly flows increases even more if an appropriate digital object separation method is also used. Conclusions Four novel digital object separation methods are developed and their functionality is tested. Convex-perimeter, Curvature-profile and Shen’s methods have good capabilities in detecting and separating the groups of objects. The benefits of using an object separation method are emphasized. The computation time is the main disadvantage of using an object separation method. Large object groups can be separated if the separation procedure is done iteratively (several times to each separated part of the group). The developed methods can be united in order to improve the performance. For separation of sharp object images, the Curvature-profile method is effective and accurate enough. It is challenging to detect the correct perimeter of unfocused objects and to automatically select the suitable threshold value. These two challenges affect the detection and separation of objects and they are not yet met in this study. References Berg van den E H, Meesters A G C A, Kenter, Schlager W 2002, Automated separation of touching grains in digital images of thin sections. Computers&Geosciences vol. 28, pp. 179-190. Chigier N 1991, Optical imaging of sprays. Prog. Energy Combustion Sci. Vol. 17, pp. 211-262. Fitzgibbon A, Pilu M, Fisher R B 1999, Direct least square fitting of ellipses. Pattern Analysis and Machine Intelligence, IEEE Transactions Vol. 21, pp. 476 -480. Gouriet J-B, Vallauri A, Riethmuller M L, Planquart Ph 2002, Application of flow visualization and PIV to continuous steel casting research. 10th Int. Symp. on Flow Visualization (26-29 August 2002, Kyoto, Japan). Honkanen M 2002a, Turbulent Multiphase Flow Measurements with Particle Image Velocimetry: Application to Bubbly Flows. MSc thesis, Tampere University of Technology, Finland. Honkanen M, Saarenrinne P 2002b, Turbulent bubbly flow measurements in a mixing vessel with PIV. 11th Symp. on Applications of Laser Techniques to Fluid Mechanics (7-10 July 2002, Lisbon, Portugal). Honkanen M, Saarenrinne P 2002c, Calibration of PIV to measure local void-fractions in bubbly flows. Pivnet 2/ ERCOFTAC workshop, Lisbon, 5.-6. July. Honkanen M, Saarenrinne P 2003, PIV methods for turbulent bubbly flow measurements. Pivnet 2/ ERCOFTAC workshop, Zaragoza, 1.-2. April. Kiger, K T, Pan C 2000, PIV Technique for the Simultaneous Measurement of Dilute Two-Phase Flows. Journal of Fluids Eng. vol. 122, pp. 811-818. Lindken R, Merzkirch W 2002, A novel PIV technique for measurements in multiphase flows and its application to two-phase bubbly flows. Experiments in Fluids vol. 33, pp. 814-825. Oberdier L M 1984, An Instrumentation System to Automate the Analysis of Fuel-Spray Images Using Computer Vision. ASTM nr. 848, pp. 123-136. Pla F 1996, Recognition of Partial Circular Shapes from Segmented Contours. Computer Vision and Image Understanding vol. 63, pp. 334-343. Shen L, Song X, Iguchi M, Yamamoto F 2000, A method for recognizing particles in overlapped particle images. Pattern Recognition Letters vol. 21, pp. 21-30. Wallis G B 1969, One-dimensional Two-phase Flow. McGraw-Hill, Inc. pp. 243-281. Wu Q-Z, Jeng B-S 2002, Background subtraction based on logarithmic intensities. Pattern Recognition Letters vol. 23, pp. 1529-1536.

10