Minerals Engineering, Vol. 10, No. 10, pp.

1075-1083, 1997
© 1997 Elsevier Science Ltd
Pergamon All rights reserved. Printed in Great Britain
80892-.6875(97)00094--0 0892-6875/97 $17.00+0.00

AN IMAGE PROCESSING ALGORITHM FOR MEASUREMENT OF

F L O T A T I O N F R O T H BUBBLE SIZE A N D S H A P E D I S T R I B U T I O N S

N. SADR-KAZEMI and J.J. CILLIERS

Froth and Foam Research Group, Department of Chemical Engineering, UMIST,

P.O. Box 88, Manchester M60 1QD, U.K. E,mail: [email protected]
(Received 28 April 1997; accepted 23 June 1997)

ABSTRACT

The appearance of the upper bubble surface of a flotation froth is used qualitatively as
a basis for control in industrialflotation cells. The complex structures that constitute the
surface of a froth are indicative of the sub-processes occurring within the froth, and
which affect the flotation performance. Being able to accurately and automatically
measure bubble size and shape distributions is therefore an important requirement for the
optimisatwn and control of froth processes. An algorithm has been developed for froth
surface bubble size distribution measurement, and that is largely insensitive to factors
such as froth type, lighting conditions and bubble size.

In this algorithm, reconstruction is applied following the subtraction of a constant value

from the image. Regional maxima are identified and used as markers for the watershed,
which segments the image. Equivalent circular bubble size and shape distributions can
be determined from the number of pixels in each segment.

It is shown that the bubble segmentation algorithm can be generally applied to flotation
froths from a variety of minerals and conditions, and may be suitable for application in
automated industrial froth characterisation and control systems. © 1997 Elsevier Science
Ltd

Keywords
Flotation froths; froth flotation; process control

INTRODUCTION

The appearance of the upper bubble surface of a flotation froth is used qualitatively as a basis for control
in industrial flotation ceils. The complex structures that constitute the surface of a froth are indicative of
the sub-processes occurring within the froth, such as drainage and coalescence. Froth drainage and
coalescence, the presence of selectively and non-selectively collected particles and their effect on froth
stability will affect flotation performance through the froth structure. Being able to accurately and
automatically measure bubble size and shape distributions is therefore an important requirement for the
optimisation and control of froth processes.

1075
1076 N. Sadr-Kazemiand J. J. Cilliers

The flotation performance can be quantitatively related to the froth appearance using measurable parameters
obtained by image processing. Froth-based kinetic models [1,2] have been proposed, which are based on
the flux of bubble surface overflowing the concentrate weir. For accurate measurement of this flux, the
bubble size distribution, rather than an average bubble size, is required. The presence of even a small
fraction of bubbles different in size to the average can significantly affect the specific surface area and
hence the flux. In order to achieve accurate characterisation of flotation froth structures, it is necessary to
develop image processing technique to identify the individual bubbles of the froth.

The surface bubble structure is generally difficult to identify using classical segmentation techniques based
on the grey level intensities of image pixels, since the image is made up of distinct homogeneous objects
with low contrast. This is unlike many image processing applications where the objects of interest from only
part of the image and are distinctly different from the image background. In addition, the feature scale is
different between laboratory and industrial applications, as well as between process stages, for example from
the rougher to the cleaner stages in a circuit. Even in a single mineral flotation operation, day-to-day process
variations and changes in the quality of lighting changes the appearance of the froth. This makes the
development of generic feature extraction algorithms difficult,

This work discusses an algorithm that has been developed for froth surface bubble size distribution
measurement, and that is largely insensitive to factors such as froth type, lighting conditions and bubble
size.

FLOTATION FROTH IMAGE PROCESSING

The rapid progress in computer technology has made the use of machine vision in the control of industrial
flotation operations possible. In general, however, the difficulty has been in relating the information
extracted from the images to the flotation performance observed.

The use of textural analysis for extracting image information has been proposed [3]. A number of grey level
dependance matric features are extracted from the image in the form of single numbers. It is found to be
difficult to attach physical meaning to these numbers, although they qualitatively indicate features such as
fineness, homogeneity and coarseness. These numbers, in turn, could be related to qualitative froth
characteristic classes, such as "ideal", "runny" and "sticky".

In other cases, an empirical modelling approach has been used, often through the use of neural network
models. Moolman et al. [4] train a neural network to classify image characteristics. It is stated that the
addition of a machine vision system would provide valuable tool in the diagnosis of off-specification
performance. However, they recognise that the complexity of the froth makes it extremely difficult to relate
the visual parameters to individual processes or mechanisms which occur in the froth or to build up the
large data bank which would be required to be able to have confidence in the capacity of the neural network
to correctly diagnose faults.

Hargrave et a/.[5] suggest a simple method of occurring visual information from images of the froth
surface. The grey level value across the froth surface was found to show a correlation to ash content and
froth mass flow rates for both a batch laboratory cell and a bank of four pilot-scale cells floating coal. More
recently, Hargrave and Hall [6] were able to relate concentrate grade in a tin plant to the colour of the
images of the froth.

There have been few attempts at estimating the complete bubble size distribution and relating that to the
flotation performance. This requires detection of the edges between the bubbles, and linking of the detected
edges so that the bubbles can be segmented into individual features. Many different types of edge operators
have been developed using techniques such as template matching, threshoiding and statistical tests. Both
the template matching and thresholding edge detectors are simple in design but are not very robust. This
means that as the image quality deteriorates, the performance of these detectors declines significantly.
 Measurement of flotationfroth bubblesize and shape distributions 1077

Woodbum et al. [1] show that a threshold function can be used to characterise the average bubble size of
flotation froths. Whilst the differences in froth structure were distinguishable with this technique, the use
of solely a threshold function was found to underestimate bubble sizes.

Symonds [7] defines bubble boundaries as those points at which an intensity profile shows a minimum.
Symonds finds that morphological segmentation techniques can accurately and reliably extract the bubble
boundaries, resulting in segmented images consisting of isolated regions. These regions are representative
of the individual bubbles and can be processed to provide a statistical description of the surface of the froth.
The procedures, however, are not generally applicable, and have been found to be strongly dependant on
image quality.

An accurate measurement of the bubble size distribution on the surface of the froth is required if the
flotation performance is described using a froth kinetic model, which depends on the froth surface flux
overflowing the weir.

FROTH BASED KINETIC MODEL

The kinetic model is based on the flux of bubble surface overflowing the concentrate weir in a free flowing
froth. This modellJtng strategy was developed by Woodburn et al. [I], who consider the amount of bubble
surface in the froth as the critical parameter which determines the rate of removal of solid species and water
from the flotation cell.

Bikerman [8] notes that bubbles in a foam are stratified, the shell around the air having a very high
viscosity, and that in between these shells exists an interbubble lumella, which has the viscosity of the bulk
solution. The model of Woodburn et al. was recently extended [2] to accommodate this froth structure.

The critical aspect.,; of the model can be summarised as follows. The flux of bubble surface Wa, overflowing
the weir is

= x G!=] x s= (I)

Where a is the w~lume fraction of air supplied to the cell, Qa, that overflows the weir in the form of
unbroken bubbles. The specific surface of the bubbles in the overflowing froth, Sa, can be estimated by
analysis of images of the top surface of the froth, which yields the bubble size and shape distributions.

Due to the difficulty in accurately estimating oc, especially in industrial operations, it is often more
convenient to estinaate the flux of bubble surface from the volume overflow rate of froth, which can also
be determined by image processing. In this case

It'a = [x x ¢ x v ! x h x w] x S a (2)

where vf is the froth velocity of the upper surface in the direction of the weir, and h and w the froth height
and weir width, respectively. The term K takes into account the velocity profile through the flowing foam,
and G is the volume fraction of the air in the overflowing froth.

Associated with the flux of bubble surface are two particle transport mechanisms. The first relates to the
hydrophobic solids, selectively attached to the bubble in the aqueous shell on the curved bubble surface.
The second mode of transport of particles is non-selective and is associated with the water in the interbubble
lamellae, in which both hydrophilic and hydrophobic particles are entrained over the concentrate weir. Water
is removed from the cell in the overflowing froth both in the bubble shell and in the lamellae.

For the use of this model in control, the accurate measurement of the bubble surface flux overflowing the
1078 N. Sadr-Kazemiand J. J. Cilliers

weir is required. This requires an image processing algorithm that can accurately, automatically and robustly
estimate the bubble size distribution of the surface of flotation froths, which is essential for calculating the
flux.

IMAGE PROCESSING PROCEDURES

The Watershed Method

Of the morphological methods, the watershed has been shown to a be robust technique compared to
classical edge detection for various applications, especially, as is the case with the surfaces of froths, when
the image is made up of distinct homogeneous objects with low contrast. Digabel and Lantuejoul [9]
introduced the watershed transformation as a morphological tool. Their data were binary images representing
successive thresholds of the relief of a bituminous surface. A study by Bencher and Lantuejoul [10] led to
the inversion of the original algorithm to extend its applicability.

The principle of the watershed method is a simulation of the rise of water from a set of markers, as shown
in Figure 1. The level p is flooded at a uniform speed from the fronts coming from level p-1 and the local
minima appear at level p. A point of the watershed appears when two distinct fronts join. In Figure 1, A,
B and C represent three minima, of which A and C have been marked as starting points. D and E are two
maxima, D is identified as a watershed, but not E, as it does not flood from two markers. The method
identifies regional rather than local maxima by locating trends in the pixel intensity along different
orientations of scan lines. As a result, it does not have the same dependence on uniform illumination as
other techniques, such as threshold filtering and local gradient techniques (e.g. SOBEL faltering).

D D

k__l l lk_l

B B B
Fig. 1 Watershed Edge Detection

It should be clear from the above example that the critical step for the watershed method to be robust and
generally applicable is to produce an appropriate marker image, as the quality of the marker image used for
the watershed significantly affects the quality of the segmentation.

It is therefore crucial to design robust pre-processing procedures which could be applicable to many
applications, since, once correct markers have been obtained, the watershed transformation can be used to
achieve the desired segmentation automatically.

Greyscale Reconstruction

Vincent and Soille [11] examine the use of morphological greyscale reconstruction, which iteratively maps
from a marker image onto an output image. At each step the marker image is dilated by a structuring
element, and used as the marker image for the next step. This is repeated tmtil convergence, which results
in a more homogenous output image, as shown in Figure 2.

Vincent [12] notes that reconstruction provides a very efficient method to extract regional maxima and
minima from greyscale images. He gives examples of identifying features using reconstruction, and
advocates this as a preprocessing step from an appropriate marker image before applying the watershed. F~w
 Measurement of flotation froth bubble size and shape distributions 1079

the identification of features (or markers) he finds that an arbitrary subtraction followed by a reconstruction
is adequate. He does not, however, use reconstruction as a method for identifying suitable markers for
watershed edge detection.

The bubble identification procedure developed here includes a robust pre-processing technique based on
reconstruction, followed by the more conventional edge detection method using the morphological
watershed. It has been found to successfully isolate bubble edges for a variety of flotation froths.

f(x)
f(x)! A
I I #.~,
Original Image I ,~, ** B j * * Original lmage
i 1[ • I It | | l[ $
i • • * t | I I •

i~ t* i • I u

~f Reconstracted Jmage

................... rker Image

Fig.2 Grey Scale Reconstruction

FROTH BUBBLE EDGE DETECTION ALGORITHM

The following algorithm has been found to be suitable for marker identification for froth surfaces. After
marker extraction, the rest of segmentation can proceed automatically.

In all cases histogram equalisation is applied, which enhances the image by increasing the dynamic range
of the pixel brightnesses. This can have a considerable effect on improving the appearance of the image.
Reconstruction is ~tpplied following the subtraction of a constant value from the image. The subtraction and
reconstruction extracts light structureswithout involving any size or shape criteria. In contrast to Vincent
[12], it has been faund that the value of the constant is critical, but that this value falls within a narrow
range for froths fi'om a variety of sources and conditions. The narrow range of suitable constants for
subtraction is of interest, as it indicates a relatively constant regional contrast for different froths between
the bubble reflections and their edges.

Following the reconstruction, regional maxima are identified using the criteria that pixels of the same value
are connected and surrounding pixels must all have values less than the local maximum. It should be noted
that neither simple thresholding nor regional maxima, to detect the bubble reflections, is adequate to
generate the marker image. The binary image consisting of the regional maxima is used as the marker image
for the watershed, which segments the image into distinct regions, each representing a bubble.

It should be noted that once the segmentation has been achieved, that not only can equivalent circular
bubble size distributions be determined from the number of pixels in each segment, but the shape
distribution and as:~ociated properties such as the elongation can be readily extracted.

The complete algol:ithm is given as a flowsheet in Figure 3. Note that all the components of the algorithm
are standard image processing routines which are available in most packages, thereby obviating the need
for writing software.
1080 N. Sadr-Kazemi and J. J. Cilliers

RESULTS AND DISCUSSION

The following examples show application of the algorithm. In each case the original image, the marker
image and the detected edges, segmenting the original image, are shown.

HISTOGRAMEQUATION ]

SUBTRACTION
r1
Marker ! l
I ]
Marker2
i WATERSHED j

Fig.3 The Segmentation Algorithm

Figure 4 shows a laboratory scale coal flotation froth consisting of a mixture of large and small bubbles.
This shows the accurate detection of bubbles of different sizes, which is often problematic. It should be
noted that in a case like this, the surface flux will be dominated by the small bubbles, and the average
bubble size will not be adequate for modelling purposes°

....ii~

Fig.4 Marker identification and segmentation of a froth consisting of a mixture of large and small bubbles

Figure 5 is an overflowing coal froth from a laboratory scale experiment with a width of 15cm. Here,
although the bubble size is relatively even, the lighting is relatively poor, and the brightness and contrast
changes significantly from the back of the cell (at the righ0 to the weir (at the left). In spite of this, accurate
segmentation is achieved.

Fig.5 Marker identification and segmentation of an overflowing coal froth from a laboratory scale
experiment with a weir width of 15cm.
 Measurement of flotation froth bubble size and shape distributions 1081

Figure 6 shows a froth from a pilot scale fluorspar separation, with a weir width of 23 cm. Note that in this
case, due to the high reflectance of the minerals in the froth, a second histogram equalisation step was
required following the reconstruction and preceding isolation of the maxima.

Fig.6 Marker i’dentification and segmentation of a pilot scale fluorspar sepration, with a weir width of
23cm.

Figures 7 and 8 shows different coal froths (froth 1 and froth 2, respectively) from a Hallimond tube of
diameter 2Scm. Note that the lighting is very different in the two cases, but that the segmentation is
achieved equally well. Figure 9 shows the statistical data from the analysis of the image processing of these
two froths. It can be seen that although the average bubble size is very similar (18mm and 15nm,
respectively), the distributions, and hence the calculated bubble surface flux, is significantly different.

Fig.7 Marker identification and segmentation of a coal froth (froth 1) in a Hallimond tube of diameter
2Scm.

Fig.8 Marker identification and segmentation of a coal froth (froth 2) in a Hallimond tube of diameter
2Scm.

Figure 10 shows an image of a rock pile, and the segmentation of the individual rocks. It can be seen that
the result is satisfactory, and that it may be possible to use this algorithm for the analysis of images from
other minerals processing applications.
1082 N. Sadr-Kazemi and J. J. Cilliers

8 70
..o
~ 60
130
~o 50

E
7=30
=>20
"10
E
o 0
0 20 40 60 80 100
Spherical Bubble Diameter(ram)

-m- Froth 1 .-.- Froth 2]

Fig.9 Results of segmentation and analysis of Figures 7 and 8.

Fig.10 Marker identification and segmentation of a rock pile into individual rocks.

CONCLUSIONS

The accurate and rapid detection of the bubble size and shape distribution on the surface of flotation froths
is important as a quantitative indicator of flotation performance. In conjunction with other measurements,
it can be used for froth control through an appropriate model. The bubble size distribution is further
indicative of the sub-processes occurring within the froth, and may be used to estimate coalescence and
bursting rates. A successful methodology for bubble edge detection under a wide range of conditions has
been developed.

The segmentation algorithm, using reconstruction and the watershed as the critical steps, can be generally
applied to flotation froths from a variety of minerals and conditions. This algorithm is requires no manual
intervention, and may therefore be suitable for application in automated industrial froth charaeterisation and
control systems.

REFERENCES

. Woodburn, E.T., Austin, L.G. & Stockton, J.B., A froth based flotation kinetic model. Chem. Eng.
Res.Des., 72, 1994.
2. Cilliers, J.J., Asplin, R.A. & Woodburn, E.T., Kinetic flotation modelling using froth imaging data.
Frothing in Flotation H (ed. J. Laskowski and E.T. Woodburn), In print, 1997.
 Measurement of flotation froth bubble size and shape distributions 1083

. Moolman, D.W., Aldrich, C., Schmitz, G.P.J. & van Deventer, J.S.J., The interrelationship between
surface froth characteristics and industrial flotation performance. Minerals Engineering, 9, 8,
837-854. 1996.
. Moolman, D.W., Aldrich, C., van Deventer, J.S.J. & Bradshaw, D.B., The interpretation of surface
flotation froths using digital image analysis and neural networks. Chemical Engineering Science,
50, 22, 3501-3513, 1995.
. Hargrave, J.M., Miles, N.J. & Hall, S.T., The use of grey level measurement in predicting coal
flotation performance. Minerals Engineering, 9, 6, 667-674, 1996.
6. Hargrave, J.M. & Hall, S.T., Diagnosis of concentrate grade and mass flowrate in tin flotation from
colour and surface texture analysis. Presented at Minerals Engineering '96, Brisbane, Australia,
26-28 August 1996.
. Symonds, P.J., The investigation of the characterisation of flotation froths and design of a machine
vision system for monitoring the operation of a flotation cell ore concentrator. MSc Thesis,
(University of Cape Town) 1992.
. Bikerm~a, J.J., Foams. Springer-Verlag, New York, 1973.
9. Digabel, H. & Lantuejoul, C., Iterative algorithms. Proc. 2rid European Symp. Quantitative
Analysis of Microstructures in Material Science and Medicine, 1977.
10. Beucher, S. & Lantuejoul, C., Use of watersheds in countour detection. Proc. Int. Workshop Image
Processing, 1979.
11. Vincent, L. & Soille, P., Watershed in Digital Space: an Efficient Algorithm Based on
Immersi~m Simulations. IEEE PAMI, 3, 6, 1991.
12. Vincent, L., Morphological Grey Scale Reconstruction in Image Analysis: Applications and
Efficient Algorithms. IEEE Trans. Image Proc., 2, 1993.