International Journal of Mineral Processing 121 (2013) 6–11

Contents lists available at SciVerse ScienceDirect

International Journal of Mineral Processing

journal homepage: www.elsevier.com/locate/ijminpro

A stochastic approach for measuring bubble size distribution via image analysis
W. Kracht ⁎, X. Emery, C. Paredes
ALGES Laboratory, Advanced Mining Technology Center, Universidad de Chile, Chile
Mining Engineering Department, Universidad de Chile, Chile

a r t i c l e i n f o a b s t r a c t

Article history: Gas dispersion properties are key parameters in flotation performance. Flotation rate constant has been
Received 13 April 2012 reported to be strongly influenced by the bubble surface area flux (Sb). In order to determine Sb, both
Received in revised form 20 November 2012 superficial gas velocity (Jg) and the bubble size distribution (BSD) have to be measured. The superficial gas
Accepted 26 February 2013
velocity can be easily measured using the inverted probe method. On the other hand, measuring the bubble
Available online 6 March 2013
size distribution is an intricate process. Among the available methods for measuring BSD, image analysis is
Keywords:
the most popular; nevertheless, it has some limitations, especially when bubbles appear touching each
Bubble size distribution other in the image (bubble clusters). From a classical image analysis point of view, it is usually assumed that
Froth flotation bubble clusters occur as a non-selective process and that not considering them does not bias the measurement.
Image analysis In this work, an alternative method for measuring BSD is presented. The image is represented as a coverage
Boolean model stochastic process, namely a Boolean model with circular objects, which allows determining statistical
diameters, d10 and d32, and the complete BSD directly from the binary image. Both the classical image analysis
and the Boolean modelling were tested first with a series of images generated at a laboratory flotation cell,
and later with a series of simulated images where the actual BSD was known. In both cases, the stochastic
approach proved to be better than classical image analysis in the determination of BSD and statistical diameters.
The results show that even if bubble clusters are non-selective (which was imposed on the simulated images)
large bubbles are more likely to be in a cluster, which implies that clusters must be considered in the measure-
ment to avoid biased estimations.
© 2013 Elsevier B.V. All rights reserved.

1. Introduction which makes difficult to determine the bubble surface area flux. An
alternative to measuring BSD (or Sb) is estimating it, but although
Flotation, as a heterogeneous process, is affected by how well the gas – there are models proposed in the literature for estimating either the
usually air – is dispersed into bubbles and distributed across the flotation mean bubble size or directly the bubble surface area flux (Gorain et
cell (Schwarz and Alexander, 2006). In order to quantify the gas disper- al., 1999; Nesset et al., 2006), they rely on parameters that have to
sion, a series of so-called ‘gas dispersion properties’ are defined, namely, be adjusted with actual BSD data. This, together with the fact that dif-
gas holdup (volumetric fraction of gas in a gas-slurry mix, εg), superfi- ferent BSDs may be represented by the same mean bubble size, and
cial gas velocity (volumetric gas flow rate per cross-sectional area of therefore give the same value of Sb (Maldonado et al., 2008), justify
cell, Jg), bubble size distribution (BSD, characterised by a statistical the need for measuring the bubble size distribution rather than just
bubble diameter db), and the derived parameter bubble surface area estimating it.
flux (Sb = 6Jg / db, where db is usually the Sauter mean diameter d32) Among the available techniques for measuring BSD (Randall et al.,
(Nesset et al., 2006; Gomez and Finch, 2007). Sb represents the flux of 1989; Grau and Heiskanen, 2002; Gomez and Finch, 2007), the
bubble surface area per cross-sectional area in the flotation machine, sampling-followed-by-imaging technique has captured most of the
and is related to the flotation rate constant (Gorain et al., 1997, 1998; attention. The apparatus (Gomez and Finch, 2007) consists of a bubble
Hernandez et al., 2003; Hernandez-Aguilar et al., 2005). sampling tube attached to a sealed viewing chamber, also known as
Among the gas dispersion properties, the gas holdup and superfi- bubble viewer, with rear illumination (Fig. 1). The bubble viewer is
cial gas velocity are relatively easy to measure and methods are well installed on top of the flotation cell and the sampling tube, usually a
described elsewhere (Gomez and Finch, 2007). The bubble size distri- tube of 1/2 inch internal diameter, – with a stopper at the lower end –
bution, on the other hand, is a more complicated variable to measure, is immersed into the cell, in the collection zone; both the bubble viewer
and sampling tube are later filled with water containing frother to
prevent coalescence (Zhang et al., 2009) and the stopper at the end of
⁎ Corresponding author at: ALGES Laboratory, Advanced Mining Technology Center,
the sampling tube is removed to allow bubbles to rise into the viewing
Universidad de Chile, Chile. Tel.: +56 2 2978 4506; fax: +56 2 2978 4985. chamber where images are captured. The images are processed off-line
E-mail address: [email protected] (W. Kracht). to yield BSD data.

0301-7516/$ – see front matter © 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.minpro.2013.02.016
 W. Kracht et al. / International Journal of Mineral Processing 121 (2013) 6–11 7

Viewing chamber
Camera

Lamp

Sampling tube

Fig. 1. Schematic of image-for-sampling device for measuring BSD.

Fig. 3. Example of realisation of a Boolean model with circular objects.

According to Gomez and Finch (2007), “although the accuracy of the distribution is unknown, and, as mentioned above, there is no ‘gold
technique is hard to establish (there is no ‘gold standard’ against which standard’ against which to compare.
to compare) the approach is now widely used, and continues to evolve
as improvements dictated by field experiences are implemented.” 3. A stochastic approach
Although the technique is being widely used, there are some issues
that may affect its accuracy, e.g., sampling at the entrance of the sam- 3.1. Model description
pling tube, bubble evolution along the sampling tube, and limitations
associated with the image analysis process. Some of these issues have Binary images of bubbles, as the one presented above, can be
been addressed (Zhang et al., 2009), however there is still room for modelled as a stochastic process, namely, a set of circular objects, with
improvement, especially in the image analysis process. varying diameters, disposed randomly over the area of the picture.
This idea is formalised by considering the so-called Boolean model
(Matheron, 1975), which is defined as the union of objects placed at
2. Classical image analysis random locations, allowing them to overlap, i.e., to form clusters:

Once the images are recorded, they are converted into binary • The locations where the objects are placed (position of each
images and processed in order to identify objects (segmentation). bubble's centre in the image) form a homogeneous Poisson point
Those objects are later characterised by their position and a series process with a given intensity θ, corresponding to the average num-
of metrics, e.g., area, maximum and minimum diameter, circularity, ber of bubble's centres per unit area (pixel). This intensity is as-
etc. Fig. 2 shows an image of bubbles generated at a flotation machine sumed to be constant over the image.
and its binary version. • The objects (bubbles) in a Boolean model are independent
In Fig. 2 it can be seen that bubbles appear either as single bubbles, realisations of a typical object, which in this case corresponds to a
bubbles truncated by the edges of the image, or as several bubbles disc with diameter given by a cumulative distribution function F
touching each other (bubble clusters). After segmentation, objects to be determined.
in the image are classified into one of these categories by their posi- • The objects (bubbles) and the point process (positions) are assumed
tion and metrics. to be independent. In other words, there is no segregation in the
Bubble clusters are difficult to deal with and require specific image.
procedures to be analysed. In order to avoid the difficulties of sizing
bubbles in bubble clusters, they are assumed to occur as the result Fig. 3 shows a realisation of a Boolean model for a known bubble
of a non-selective process, and they are usually neglected in the diameter distribution (log-normal). The similitude between the image
analysis. The same criterion is applied to truncated bubbles. These in Fig. 3 and a binary image of actual bubbles (Fig. 2 right) suggests
assumptions are not easy to verify because the actual bubble size that the bubble size distribution can be determined by solving the

Fig. 2. Image of bubbles in a flotation machine. Original image (left), and binary image (right).
8 W. Kracht et al. / International Journal of Mineral Processing 121 (2013) 6–11

the geometric covariogram of the object is a mixture of two circular

covariograms given by:

∀h≥0; K ðhÞ ¼ p1 circδ1 ðhÞ þ p2 circδ2 ðhÞ: ð2Þ

This is a nested covariogram, composed of a circular covariogram

with range δ1 and sill p1 and a circular covariogram with range δ2
and sill p2. Eq. (2) can be generalised for a typical object diameter
having a distribution function F. The geometric covariogram in that
case has the form:

þ∞
∀h≥0; K ðhÞ ¼ ∫ circδ ðhÞ F ðdδÞ: ð3Þ
0

Fig. 4. Schematic of the geometric covariogram of a typical object (disc). This is also a nested covariogram where the ranges of the nested
circular covariograms give the set of possible object diameters, and
the sills give their respective probabilities of occurrence. Fig. 5
inverse problem, i.e., given a realisation of a Boolean model with circular shows the geometric covariogram of a circular object (a bubble) of
objects, determine the distribution of the object diameters. 0.8 mm in diameter, and the geometric covariogram of a bubble size
distribution (75% 0.2 mm, 20% 0.4 mm, 5% 0.8 mm).
3.2. Geometric covariogram and diameter distribution of the typical Since there is a one-to-one relationship between the distribution F
object and the nested covariogram K(h) composed of a mixture of circular
covariograms (Hall, 1988), the solution of the inverse problem (deter-
In theory of random sets and stochastic geometry (Serra, 1982; mining F once K(h) is known) represents an alternative to the classical
Molchanov, 2005) the typical object can be described by its geometric image analysis, where objects have to be separated (segmentation)
covariogram. This is a function that depends on a lag distance (h) and and individual measurements carried out in order to determine the
corresponds to the average area of the intersection of the object with distribution.
its translate by h (shaded region in Fig. 4). Eq. (4) presents such a solution to the inverse problem stated above
The geometric covariogram is a function that vanishes when the lag (Emery et al., 2012). The number frequency is easily determined after F
distance is greater than the maximal size of the object (this distance is is known.
known as the range of the covariogram). It also contains information
of object shape and other geometrical properties (area, perimeter,
etc.) (Serra, 1982; Lantuéjoul, 2002; Emery and Lantuéjoul, 2011). 2 þ∞ K″ðhÞdh
∀δ > 0; 1−F ðδÞ ¼ ∫ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð4Þ
π δ h2 −δ2
If the typical object is a disc with fixed diameter δ (i.e., if the distri-
bution function F is a Heaviside step function at δ), then the geometric
covariogram of the object is the so-called circular covariogram (Chilès where K″ is the second-order derivative of K.
and Delfiner, 2012):
3.3. Geometric covariogram and background covariance function
  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 
 2  ffi
 δ arccosh −h 1− h 2 if h≤δ The geometric covariogram of the typical object can be associated
∀h≥0; circδ ðhÞ ¼  2 δ δ δ
: ð1Þ
 0 otherwise with the statistical properties of the Boolean model, represented as a
binary image. The procedure consists in considering the background
of the image (complement of the Boolean model) and defining its
If the typical object is a disc with two possible diameters δ1 non-centred covariance function (C) at lag distance h, as the probability
(with probability p1) and δ2 (with probability p2 = 1 − p1), then that two points distant by h simultaneously belong to the background.

Fig. 5. Geometric covariogram of a 0.8 mm bubble (left), and of a bubble size distribution: 75% 0.2 mm, 20% 0.4 mm, and 5% 0.8 mm (right).
 W. Kracht et al. / International Journal of Mineral Processing 121 (2013) 6–11 9

Fig. 6. Top: actual binary images of bubbles generated at a laboratory flotation cell (image size is 1890 × 1260 pixels). Middle: estimated background covariance functions. Bottom:
estimated bubble size distributions, up to the multiplicative factor theta. The length unit for middle and bottom figs. is the pixel size.

As shown by Lantuéjoul (2002), there exists a one-to-one rela- 3.4. Proposed approach for estimating the diameter distribution and
tionship between the geometric covariogram of the typical object application
and such a covariance function:
The strategy to determine the bubble size distribution can be
∀h≥0; C ðhÞ ¼ expfθK ðhÞ−2θK ð0Þg: ð5Þ summarised as follows (Emery et al., 2012):
Equivalently:
1) Experimentally estimate the covariance function C(h) of the image
∀h≥0; ln½C ðhÞ ¼ θK ðhÞ−2θK ð0Þ ð6Þ background. To this end, let h be a multiple of the pixel inter-
so that distance, P1(h) the set of pairs of pixels oriented along the main
axes of the image and distant by h and P2(h) the subset of P1(h)
∀h≥0; ln½C ðhÞ″ ¼ θK″ðhÞ: ð7Þ such that the two pixels of the pair belong to the image background
10 W. Kracht et al. / International Journal of Mineral Processing 121 (2013) 6–11

(i.e., they are not covered by any bubble). The background covari- the picture. Since the stochastic approach does consider bubble clusters
ance function C(h) is then estimated by dividing the cardinal of and bubbles truncated (given that the centre of the bubble resides in the
P2(h) by the cardinal of P1(h). area of the picture), the measured distribution is almost identical to the
2) Take the logarithm of this covariance function and calculate its actual BSD measured manually (Fig. 7, right).
second-order derivative (using a second-order finite difference). The bubble sizes d10 and d32 for the actual BSD are 1.01 and
This gives an estimate of θK ″(h), as per Eq. (7). 1.29 mm respectively, whereas the same parameters, determined
3) For δ > 0, estimate θ½1−F ðδÞ by integrating 2θK″ ðhÞdh
pffiffiffiffiffiffiffiffiffiffi
ffi over the dis-
2
from classical image analysis are 0.93 and 1.14 mm, which represent
π h −δ2
an underestimation of 7.9% and 11.6%, respectively. This means that
tances h greater than δ, as per Eq. (4). For details on the numerical
the bubble surface area flux (Sb) would be overestimated by 13.2% if
integration process, the reader is referred to Emery et al. (2012).
classical image analysis is used to measure the BSD, and no clusters
or truncated bubbles are considered. On the other hand, when the
The method allows estimating the Poisson intensity (θ) as the sill stochastic approach is used to determine BSD, the statistical bubble
of the function estimated at step 3) and the diameter complementary sizes d10 (0.99 mm) and d32 (1.19 mm) are estimated with a 2.0%
cumulative distribution function (1 − F(δ)) as the same function and 7.8% error respectively. Accordingly, the estimation is better
normalised to a unit sill. than with classical image analysis, but there is still an error associated
with the technique. This may occur because of two reasons. First, the
3.5. Validation stochastic approach not only accounts for the bubbles that are actually
seen in the picture, but also for bubbles that may have been completely
The method was tested with a series of 10 images containing a covered and are not visible in the image, thus are ignored in the manual
total of 2852 bubbles generated at a laboratory flotation cell. The orig- process (these bubbles tend to be smaller than the average). Second, the
inal images were first processed manually to yield the actual bubble actual images (see Fig. 6) may slightly deviate from the Boolean model
size distribution and statistical diameters. Following manual process, assumptions, in particular they may be contaminated by noise (isolated
the binary images were processed using the classical image analysis black pixels in the binary image) or the bubbles may not be homoge-
technique, considering no bubble clusters or bubbles touching the neously distributed over the image (non-constant Poisson intensity θ).
edges of the picture. Finally, the set of binary images was processed However, the stochastic approach has proved to be robust to such
applying the proposed stochastic approach. As an example, Fig. 6 deviations (Emery et al., 2012), so that their effect on the quality of
shows a sample of two of those images (top figures), as well as the the estimation is expected to be small. In contrast, the method is sensi-
corresponding background covariance functions (middle) and bubble tive to the assumption of having circular bubbles, but this assumption is
size distributions (bottom) estimated with the stochastic approach. appropriate in the present case study (Fig. 6). Other anomalies, such as
The final estimate of the bubble size distribution is obtained by uneven lighting or blurred images, can be avoided by using suitable
averaging the curves calculated on each image separately and by camera settings.
normalising to a unit sill. At this point it is worth mentioning that the classical image
Fig. 7 shows the average distributions obtained with each method analysis could be improved by implementing a watershed algorithm
as probability densities. Note that the averaging process over the 10 that makes possible to separate bubbles in clusters (Beucher, 1992).
images reduces the fluctuations observed in the individual distribu- However, such an algorithm is useful only if bubbles are just touching
tions estimated on each image (Fig. 6, bottom), thus yielding a more and not overlapping. The fact that some bubbles are not just touching
robust estimate of the actual BSD. each other but are partially or totally hidden by other bubbles, and
As Fig. 7 (left) shows, large bubble diameters are underestimated that some are truncated by the image edges, makes the classical
and the measured distribution is displaced towards lower bubble method difficult to compensate for the bias, even when a watershed
diameters when bubble clusters and/or bubbles truncated by the algorithm is considered.
edges of the images are not considered. This occurs because large bub- In order to test whether the stochastic approach is actually account-
bles cover a greater area of the image and, therefore, are more likely to ing for all the bubbles in the BSD, even those bubbles that cannot be
appear either overlapped with other bubbles or touching the edges of seen because they are covered by larger ones, a second experiment

Fig. 7. Comparison between manual BSD (obtained manually) and BSD estimated by image analysis neglecting bubble clusters and bubbles truncated by the edges of image (left);
and BSD estimated with the proposed stochastic approach (right).
 W. Kracht et al. / International Journal of Mineral Processing 121 (2013) 6–11 11

Fig. 8. BSD estimated by image analysis neglecting bubble clusters and bubbles truncated by the edges of image (left); and BSD estimated with the proposed stochastic approach.

was run, this time with a simulated BSD. A set of 100 bubble images was was shown that bubble diameters (d10 and d32) are estimated within
simulated as a Boolean model (as in Fig. 3) with bubble sizes following a a 0.5% error when using the stochastic approach, whereas when using
log-normal distribution with d10 (mean bubble size) 1.0 mm and d32 the classical image analysis the diameters are underestimated by
1.37 mm. more than 10%, with the corresponding overestimation of the bubble
Fig. 8 shows the BSDs obtained by applying both the classical surface area flux.
image analysis (Fig. 8, left) and the proposed stochastic approach
(Fig. 8 right) to the 100 simulated images.
References
Again, Fig. 8 shows that when classical image analysis is used to
determine BSD, neglecting bubble clusters and bubbles truncated by Beucher, S., 1992. The watershed transformation applied to image segmentation.
the edges of the image, the measured distribution differs from the Scanning Microsc. Suppl. 6, 299–314.
Chilès, J.P., Delfiner, P., 2012. Geostatistics: Modeling Spatial Uncertainty. Wiley, New
actual (theoretical in this case) BSD. Large bubble diameters are York, NY.
underestimated and the measured distribution is displaced towards Emery, X., Kracht, W., Egaña, A., Garrido, F., 2012. Using two-point set statistics to estimate
lower bubble diameters. The bubble sizes d10 and d32 are underestimated the diameter distribution in Boolean models with circular grains. Math. Geosci. 44 (7),
805–822.
by 10.4% and 14.2% respectively, which means an overestimation in bub- Emery, X., Lantuéjoul, C., 2011. Geometric covariograms, indicator variograms and
ble surface area flux (Sb) of 16.5%. On the other hand, when the stochastic boundaries of planar closed sets. Math. Geosci. 43 (8), 905–927.
approach is used to determine BSD, measured and actual distributions Gomez, C.O., Finch, J.A., 2007. Gas dispersion measurements in flotation cells. Int. J.
Miner. Process. 84 (1–4), 51–58.
are almost identical and the statistical bubble sizes d10 and d32 are esti- Gorain, B.K., Franzidis, J.P., Manlapig, E.V., 1997. Studies on impeller type, impeller
mated within a 0.5% error. Since in this case the BSD obtained with the speed and air flow rate in an industrial scale flotation cell, part 4: effect of bubble
stochastic approach is compared to the actual (theoretical) BSD, which surface area flux on flotation performance. Miner. Eng. 10 (4), 367–379.
Gorain, B.K., Napier-Munn, T.J., Franzidis, J.P., Manlapig, E.V., 1998. Studies on impeller
may differ from what appears in the images (small bubbles hidden by type, impeller speed and air flow rate in an industrial scale flotation cell, part 5:
large ones), the error in the statistical bubble sizes is much lower. validation of k–Sb relationship and effect of froth depth. Miner. Eng. 11 (7),
615–626.
Gorain, B.K., Franzidis, J.P., Manlapig, E.V., 1999. The empirical prediction of bubble
4. Conclusions
surface area flux in mechanical flotation cells from cell design and operating data.
Miner. Eng. 12 (3), 309–322.
Bubble size distributions can be modelled as a stochastic process, Grau, R.A., Heiskanen, K., 2002. Visual technique for measuring bubble size in flotation
namely, a set of circular objects with varying diameters disposed machines. Miner. Eng. 15 (7), 507–513.
Hall, P., 1988. Introduction to the Theory of Coverage Processes. Wiley, New York.
randomly over the image. Since there is a one-to-one relationship Hernandez, H., Gomez, C.O., Finch, J.A., 2003. Gas dispersion and de-inking in a flotation
between the bubble size distribution and the geometric properties column. Miner. Eng. 16 (6), 739–744.
of the bubble images, it is possible to determine the actual BSD with- Hernandez-Aguilar, J.R., Rao, S.R., Finch, J.A., 2005. Testing the k–Sb relationship at the
microscale. Miner. Eng. 18 (6), 591–598.
out applying classical image analysis, but a stochastic approach. Lantuéjoul, C., 2002. Geostatistical Simulation: Models and Algorithms. Springer,
The results show that when classical image analysis is used to deter- Berlin.
mine BSD, both bubble clusters and bubbles truncated by the edges of Maldonado, M., Desbiens, A., del Villar, R., Girgin, H., Gomez, C.O., 2008. On-line estimation
of bubble size distributions using Gaussian mixture models. In: Kuyvenhoven, R.,
the image should be considered. Not considering them biases the mea- Gomez, C., Casali, A. (Eds.), Proceedings of PROCEMIN 2008, V International Mineral
surements because large bubbles are more likely to appear as clusters Processing Seminar. Santiago, Chile, 22–24 October 2008, pp. 389–398.
and/or truncated. This occurs even if bubble clusters occur as a non- Matheron, G., 1975. Random Sets and Integral Geometry. Wiley, New York, NY.
Molchanov, I., 2005. Theory of Random Sets. Springer, London.
selective process, which was imposed on the simulated images. Nesset, J.E., Hernandez-Aguilar, J.R., Acuña, C., Gomez, C.O., Finch, J.A., 2006. Some gas
The stochastic approach gives better results than classical image dispersion characteristics of mechanical flotation machines. Miner. Eng. 19 (6–8),
analysis for the bubble size distribution. When compared against a 807–815.
Randall, E.W., Goodall, C.M., Fairlamb, P.M., Dold, P.L., O'Connor, C.T., 1989. A method
BSD obtained manually (manual process), the estimation error was
for measuring the sizes of bubbles in two- and three-phase systems. J. Phys. E
lower for both d10 and d32. The deviation still observed with the sto- Sci. Instrum. 22, 827–833.
chastic approach may be due to the fact that this method accounts Schwarz, S., Alexander, D., 2006. Gas dispersion measurements in industrial flotation
not only for the bubbles observed in the picture, but also for bubbles cells. Miner. Eng. 19 (6–8), 554–560.
Serra, J., 1982. Image Analysis and Mathematical Morphology. Academic Press, London.
that may have been totally covered and are not visible. In a simulated Zhang, W., Kolahdoozan, M., Nesset, J.E., Finch, J.A., 2009. Use of frother with sampling-
experiment, where the actual bubble size distribution was known, it for imaging bubble sizing technique. Technical note. Miner. Eng. 22 (5), 513–515.