Minerals Engineering 20 (2007) 461–471

This article is also available online at:

www.elsevier.com/locate/mineng

Utilization of optical image analysis and automatic texture

classification for iron ore particle characterisation
E. Donskoi *, S.P. Suthers, S.B. Fradd, J.M. Young, J.J. Campbell,
T.D. Raynlyn, J.M.F. Clout 1
CSIRO Minerals, P.O. Box 883, Kenmore, Qld 4069, Australia

Received 25 August 2006; accepted 13 December 2006

Available online 6 February 2007

Abstract

Optical image analysis is a very convenient tool for obtaining comprehensive information about fine iron ore size fractions. Data can
be obtained on mineral abundances, porosity, particle shape and ore textures with a high level of accuracy. A range of techniques has
been used to characterise iron ore samples on a particle-by-particle basis. Automatic textural classification of iron ore particles was used
to establish classes containing particles with very similar mineral composition and texture. Image analysis coupled with probe analysis
and mineral density measurements provided information about the chemical composition and density of each particle class. The com-
bination of these results enabled a ‘‘virtual feed’’ to be created, which can be a key input into a beneficiation unit model for predicting its
performance. Identification and classification of the textural type of each particle was performed according to the CSIRO-Hamersley
Iron Ore Group Classification Scheme. If more detailed classification is needed, further classification can be performed based on dimen-
sional, chemical or mineral criteria, such as the presence of certain minerals in particles or total iron content. Some deficiencies of the
current image analysis procedures and their further improvement and automation are also discussed.
 2007 Elsevier Ltd. All rights reserved.

Keywords: Iron ores; Mineral processing; Ore mineralogy; Particle morphology; Liberation

1. Introduction tions of different techniques such as sizing, density classifi-

To understand the behaviour of fine iron ores in down-
stream beneficiation operations or in subsequent process-
ing operations such as granulation and sintering, and to
reliably predict such processes, a large amount of informa-
tion about the ore fines needs to be obtained (Clout, 1998;
Donskoi et al., 2006a,b). Generally it should be informa-
tion about particle mineralogy and class, porosity, mineral
associations, texture, hardness, size distribution, mineral
liberation, class densities and mineral composition, as well
as the molecular composition of each mineral. Combina-
*
Corresponding author. Tel.: +61 7 33274158.
E-mail address: Eugene.Donskoi@csiro.au (E. Donskoi).
1
Worked with CSIRO Minerals, now with Fortescue Metals Group
Limited, P.O. Box 6915, East Perth, WA 6892, Australia.
cation by heavy liquids, pycnometry, chemical analysis,
XRD, probe analysis, microscopy and image analysis can
provide the required information.
Determination of the mineralogy of iron ore particles
can be carried out by several methods – indirect or direct
measurements. Indirect methods usually assume that the
minerals in the examined ore are stoichiometric, which is
not always the case. Diagnostic leach tests and calculation
of mineralogy from assays can be included in this category
(Zhang and Whiten, 2001; Benson et al., 2001).
Among the direct techniques are quantitative XRD (e.g.
Mandile and Johnson, 1998), optical image analysis (Danti
et al., 1993) and automated image analysis using scanning
electron microscopy (SEM)-based techniques, such as
QemSCAN (Gottlieb et al., 2000) and the Mineral Libera-
tion Analyser (Gu and Guerney, 2000). The advantage of

0892-6875/$ - see front matter  2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.mineng.2006.12.005
462 E. Donskoi et al. / Minerals Engineering 20 (2007) 461–471
using an electron microscope for these purposes is that,
together with information about mineral liberation and
association, the chemistry of the minerals can be simulta-
neously provided. However, to distinguish reliably between
different iron oxides and hydroxides, the analysis time has
to be significantly increased, which may not be economical.
Also, determination of closed pores and micro porosity in
particles can be problematic if the epoxy used for sample
preparation can not penetrate the pores.
Optical image analysis systems are capable of identifying
porosity in different minerals and differentiating between
various iron ore minerals. This differentiation is based on
the distinctive reflectivities of different iron oxides and
hydroxides.
The image analysis described in this paper is based on
utilization of two software packages: KS400 (ZEISS) and
Recognition2 (see also Donskoi and Clout, 2005). KS400
is used for initial image analysis and calculation of a range
of parameters for individual particles. Recognition2,
developed by CSIRO Minerals, is used for individual
classification of every particle according to its iron ore
textural group, calculation of information about every tex-
tural group and comprehensive characterisation of the
entire group of particles. Information about the abundance
of different iron ore textural types and their chemical com-
position is very important for predicting downstream pro-
cessing performance and product characteristics (Clout,
1998; Box et al., 2002).
The ore texture group identification is performed
according to the CSIRO-Hamersley Iron Ore Group Clas-
sification Scheme described by Box et al. (2002). Classifica-
tion by ore textures provides classes of particles with
similar mineral composition and density. The reason for
this is that mineral abundances and porosity are the main
criteria used for particle classification.
Recognition2 also has the option of classifying particles
by mineral content or by total iron. The ‘‘Selection’’ proce-
dure in Recognition2 can select particles based on their
dimensional, chemical or mineral characteristics. When
linked together, these capabilities enable a set of classes
with distinct characteristics to be built. This set of classes
represents the iron ore size fractions under investigation.
If beneficiation modelling is the goal of the characterisa-
tion, and if the ore will be used as feed to specific benefici-
ation units, the full data set of classes is called the ‘‘virtual
feed’’. During subsequent modelling, the virtual products
can be calculated and compared with the measured proper-
ties of actual products using the same procedures as for the
feed material.
Information obtained from Recognition2 enables ‘‘vir-
tual feeds’’ to be constructed for feeding into a hydrocy-
clone model (Donskoi et al., 2006a) implemented within
the USIM PAC (Caspeo/BRGM, France) steady-state
plant simulator package. The modelling generates calcu-
lated overflow and underflow streams and the abundances
of different ore texture groups in these calculated streams.
These outputs can then be used as inputs for modelling unit
operations following hydrocycloning, such as magnetic
separation or flotation.
2. Initial image analysis in KS400
For optical image analysis each sample (size fraction) is
mounted in an epoxy resin block, which is polished on one
side and image analysed under reflected light (Danti et al.,
1993). Images of particle sections (see Fig. 1) are collected
using the AxioVision (ZEISS) program, and are further
processed using the MINERAL software developed by
CSIRO Minerals (Fig. 2). The MINERAL software is
based on the KS400 (ZEISS) image analysis package.
Improvement of the quality of the image is the initial
stage of image processing within the MINERAL software,
where contrast and sharpness improvement as well as fur-
ther background correction are performed. Background
correction is a special procedure which compensates for
non-uniform illumination across the image. For this pur-
pose, a reference image has to be obtained, which involves
imaging a highly uniform surface without any inclusions or
scratches using the same focal length lens and illumination
a
Kaolinite
Ochreous
Goethite
Pores
Hematite
Vitreous
Goethite
b Kaolinite
Ochreous
Goethite
Pores
Hematite
Vitreous
Goethite
Fig. 1. (a) Original optical image of the 106/+75 lm size fraction of an
iron ore sample, (b) mineral map of the image.
 E. Donskoi et al. / Minerals Engineering 20 (2007) 461–471 463

Fig. 2. Reflectivity thresholding for vitreous goethite using the KS400 MINERAL software.

Fig. 3. Calculation of sample chemical assay using the ‘‘Composition’’ module of Recognition2 for a 25 + 20 lm sample.

as the image to be corrected. A polished highly uniform After the image has been improved, different minerals
piece of quartz is suitable for this purpose. are discriminated and identified with the assistance of
464 E. Donskoi et al. / Minerals Engineering 20 (2007) 461–471
colour and greyscale thresholding and visual identification.
In Fig. 2, thresholding for identification of vitreous goe-
thite within the MINERAL program is illustrated.
Finally, information about every particle is measured
and is saved as a file. This information includes the area
of every mineral, the locked (not in contact with the epoxy
matrix) and free (connected with the epoxy matrix)
perimeter of every mineral, dimensional characteristics of
the particle, grey levels, etc. Subsequently, this information
is read and processed by the Recognition2 software
package.
In terms of sample representativity, it is understand-
able that, the larger the number of particles measured,
the better the statistical figures for particle distribution by
texture and the smaller the random errors. However, there
is still the possibility for systematic error if the set of
particle sections in the blocks is systematically different
from the actual size fraction. It is more likely for this to
happen for very small size fractions, such as CS6 and
CS7 (CS – cyclosized fractions), which have particle sizes
below 9 lm.
To make sure that the set of particles analysed is repre-
sentative of the whole size fraction, the following proce-
dure is used. In advance of image analysis, chemical
information about every major mineral in the ore is
obtained by electron microprobe analysis. Averaged values
from 30 to 100 measurements for every mineral are used.
Within Recognition2 there is a procedure named ‘‘Compo-
sition’’ (see Fig. 3). Using loaded information about aver-
age mineral density, chemical composition of each
mineral (obtained from probe analysis) and mineral abun-
dances (calculated within ‘‘Recognition2’’), the operator
can calculate the assay of the size fraction under investiga-
tion using the ‘‘Composition’’ procedure. If there is a sig-
nificant difference between the assay calculated by
‘‘Composition’’ and the chemical assay obtained from
chemical analysis it would indicate that there is an issue
that needs to be addressed, e.g., making another epoxy
block, analysing more particles or improving sampling
for chemical analysis or image analysis. The calculated
mineral composition of the size fraction can also be com-
pared with results of other direct mineral measurement
methods (e.g., Rietveld XRD Analysis).
Fig. 3 illustrates calculation of the chemical assay from
image analysis of a 25 + 20 lm sample using ‘‘Composi-
tion’’. In Table 1, the calculated assay is compared with the
conventional chemical assay. As can be seen, the two
assays are quite close, which verifies that the set of particles
extracted for image analysis is quite representative for the
sample under investigation.
Table 1
Comparison between conventional chemical assays and assays calculated using the ‘‘Composition’’ module for a 25 + 20 lm sample
Fe (%) SiO2 (%) Al2O3 (%)
Chemical assay 64.95 1.59 1.02
Image analysis 65.05 1.72 1.27
3. Texture classification in Recognition2
Identification and classification of the textural type of
each particle is a very labour consuming process, which
usually involves the microscope operator counting every
particle and allocating it to a certain ore type. This
approach can only provide data on major ore types and
the number of particles of a certain ore type in specific size
fractions, and cannot provide more detailed information
on other characteristics such as liberation, mineral associa-
tions, dimensional characteristics, average density and the
mineral and chemical composition of each ore type.
To refine the classification algorithm in Recognition2,
particle identifications from Recognition2 were compared
with identifications provided by an experienced mineralo-
gist. If these identifications did not coincide, a decision
was made as to whether there was a possibility of identify-
ing the particles correctly simply by changing boundaries
and conditions in the existing recognition scheme (if that
was the case, other particles which were close to this parti-
cle were also reconsidered) or whether a new texture char-
acterisation parameter needed to be introduced into the
image analysis software. In the vast majority of cases where
the two classifications did not agree, it was due to mineral
misidentification in the initial image analysis. At the cur-
rent stage of development of Recognition2, if minerals
are properly identified, then 96–99% of particle identifica-
tions match those provided by experienced mineralogists.
In the very last test of recognition procedure on approxi-
mately two hundred particles where mineralogy was cor-
rectly identified, only one particle was allocated to the
wrong textural class by Recognition2. Two examples of
common image analysis misidentifications are described
in more detail below. The particle shown in Fig. 4 consists
of vitreous goethite with two different degrees of hydration.
The lighter part is actually dehydrated vitreous goethite,
which has been misidentified as hematite during image
analysis. The percentages of different minerals obtained
from image analysis are shown in Table 2. When the reflec-
tivity of dehydrated goethite is compared to that of hema-
tite within the same sample (see Fig. 5), it is clearly much
darker than the hematite. This error resulted in Recogni-
tion2 mistakenly classifying the particle as a vitreous goe-
thite matrix containing martite grains (GMO), instead of
a hard particle mainly consisting of vitreous goethite
(GOV).
The second particle shown here (Fig. 5 and Table 3)
should be classified as a soft to friable hematite particle
(H2F), but Recognition2 classified it as a medium hardness
particle with hematite matrix and some goethite infill
TiO2 (%) MnO2 (%) P2O5 (%) MgO (%) LOI 900 (%)
0.04 0.06 0.14 0.07 3.83
0.02 0.02 0.12 0.07 3.77
 E. Donskoi et al. / Minerals Engineering 20 (2007) 461–471 465

Fig. 5. (a) A friable hematite matrix particle at 20· magnification and (b)
the mineral map obtained after image analysis.

Table 3
Resulting data from image analysis of the particle in Fig. 5, shown as a
percentage of the total particle area
Fig. 4. (a) A vitreous goethite particle at 20· magnification and (b) the
mineral map obtained after image analysis. Porosity Hematite Vitreous Ochreous Particle type
(%) (%) goethite goethite
Recognition2 Geologist
(%) (%)
Table 2
Resulting data from image analysis of the particle in Fig. 4, shown as a 31.31 62.55 6.66 0.87 HGM H2F
percentage of the total particle area
Porosity Hematite Vitreous Ochreous Particle type
(%) (%) goethite goethite
Recognition2 Geologist ognition2 automatically classifies particles into different ore
(%) (%)
0.36 44.71 54.86 0.06 GMO GOV
texture groups. The main information (the number of par-
ticles in every textural group, their percentages by area and
weight) appears on the ‘‘Particles Recognition’’ form (see
(HGM). The two factors contributing to this are overesti- Fig. 7). More detailed information about ore texture
mation of goethite during image analysis and underestima- groups can be calculated using this form and exported to
tion of porosity. The combination of these two errors is a a Word document named ‘‘Recognition Results’’ (all infor-
recurring one throughout image analysis, as porosity is mation calculated within Recognition2 can be exported to
often mistaken for goethite and identified as such. At this file). The calculated table includes the following
higher magnifications (see Fig. 6) it becomes clear that this parameters grouped by iron ore texture type, namely the
is an error, and the particle contains very little goethite. A number of particles, frequency%, area% and weight% of
slight reduction in the amount of goethite as well as an this ore texture group, total iron and specific gravity with
increase in porosity will provide the correct data for Recog- coefficients of variation (standard deviation/mean), mean
nition2 to classify this particle correctly. shape factor and elongation, percentages of each mineral
within each class by area or by weight, and the percentage
of pores by area (see Table 4).
3.1. Classification description If more detailed information about the entire particle
group, specific ore texture class or a combination of several
As already discussed above, after the information from texture classes is needed, the operator can go to the ‘‘Statis-
the MINERAL program is loaded into Recognition2, Rec- tics’’ procedure form in Recognition2 (see Donskoi and
466 E. Donskoi et al. / Minerals Engineering 20 (2007) 461–471

Fig. 6. A friable hematite matrix particle (as shown in Fig. 5) at 100· magnification.

Fig. 7. ‘‘Particles Recognition’’ form (in front of the main Recognition2 interface) with calculated results for the example given in Table 1.

Clout, 2005). Four different analyses can be performed
within the ‘‘Statistics’’ routine, namely particle dimensions,
modal/composition analysis, area association and inter-
growth/liberation analysis.
Particle dimension analysis (see Table 5) provides infor-
mation about the size and shape of particles. It is to be
noted that this is apparent information, which is slightly
different from that for the real particles and hence stereol-
ogical correction ought to be carried out. However, even
after very simple stereological correction, e.g., large section
correction when particle sections with small relative area
are removed, the calculated information is quite close to
reality (Lin et al., 1995). The calculated parameters and
their statistical descriptors are given in the particles
 E. Donskoi et al. / Minerals Engineering 20 (2007) 461–471 467

dimensions table, namely area of the particles, equivalent

(Area%) circle radius – radius of the circle with the same area as
the mean area of the particles, maximum length, maximum
Pores

16.66
26.79

12.34
27.96

14.27
24.49
9.95

0.56
6.45
1.63

6.55
7.13
2.74

0.00
0.11
1.93
3.36
0.00
breadth, shape factor and elongation. Statistical descrip-
tors include mean, variance, standard deviation, coefficient
Quartz
(Wt%)

of variation
P (standard deviation/mean), coefficient of dis-

94.78
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00

0.00
0.00
0.00
0.00
persion ( ðjXi  MjÞ=M, where M is the median) and 5
point distribution characteristics – minimum, first quartile,
Kaolinite

median, third quartile and maximum, and the number of

(Wt%)

0.00
0.00
0.00
0.00
0.00
0.25
0.00
0.00
0.00
0.00
0.00
0.00
0.00
0.00

0.00
99.61

35.82
79.39
cases taken into account.
‘‘Modal/Composition Analysis’’ is subdivided into three
different tables. The first one provides information about
(Wt%)
goeth

48.56

41.45
92.98
0.58
0.79
1.57
0.21
0.92
1.73
4.86
3.84
4.44

2.83

1.08
0.00
0.00
7.25
0.00
Ochr

the presence of different minerals in the particles (see Don-

skoi and Clout, 2005). For every recognisable mineral, the
table summarises the number of occurrences, the total area
(Wt%)
goeth

of all tested particles, volume percentages, surface percent-

14.01
11.87
11.34
52.69
13.34
96.48
58.30

52.42
20.61
2.15
2.04
3.11
2.30
3.65

7.01
4.13
0.39
7.35
Vitr

ages, the density that has been used for the mineral, and
weight percentages.
Hematite
(Wt%)

The second table produced during ‘‘Modal/Composi-

97.27
97.17
95.31
97.49
95.42
84.01
83.26
84.81
42.87
38.10

92.65
0.69
0.24
0.00
0.00
0.00

4.51
0.00

tion Analysis’’ is the ‘‘Variability Analysis’’ table (see Table

6). This table provides statistical analysis data for density
Mean
elong

and the presence of total iron and different minerals by

1.53
1.63
1.62
1.61
1.68
1.58
1.50
1.55
1.72
1.97
1.82
1.76
2.04
2.24
1.92
3.48
1.71
1.91

weight in the sample under investigation. This statistical

analysis can include all particles or only particles that con-
Mean
shape
Example of a particle recognition results table for groupings by ore texture type for the example given in Table 1

tain a specified mineral. The table lists the following

2.11
2.46
2.44
2.20
2.34
2.36
2.38
2.45
2.37
2.91
2.57
2.52
2.80
2.43
1.86

3.16
2.09
fact

parameters for specific gravity, total iron and minerals

present, namely occurrence, mean, standard deviation,
0.315
0.506
0.339
0.449
0.321
0.489
0.356
0.329
0.531
0.566
0.455
0.484
0.504
0.583
0.319
0.974
0.285
0.255
Area
coeff

coefficient of variability and coefficient of dispersion. Also

var

provided are the minimum, first quartile, median, third

quartile and maximum values. This analysis enables the
mineral
Mean

uniformity of mineral composition for a specific group or

area
381
337
314
367
409
386
364
350
360
359
336
382
271
192
352
138
393
462

texture class of particles to be understood. If, from a mod-

elling or characterisation point of view, the creation of clas-
0.014
0.030
0.114
0.015
0.014
0.035
0.032
0.134
0.118
0.174
0.054
0.171
0.268
0.019
0.006
0.040
0.067
0.021
coeff

ses with closer mineral composition is necessary, then

var
SG

further classification can be applied.

Mean

The third table provided by ‘‘Modal/Composition

4.55
4.21
3.68
5.04
4.71
4.79
4.25
3.51
4.10
3.82
3.88
3.28
2.74
2.70
2.60
4.90
3.25
2.80
SG

Analysis’’ gives compositional analyses. This includes the

molecular composition calculated from electron probe
Fe tot

0.005
0.004
0.006
0.004
0.004
0.016
0.016
0.009
0.026
0.064
0.010
0.037
0.019
1.114
1.425
0.009
0.300
0.267
coeff

analyses for every mineral considered and the total cal-

var

culated assay for the group of particles under investi-

gation.
Fe tot
Mean

68.24
68.21
67.95
68.31
68.04
66.83
66.52
66.78
63.18
57.23
59.84
54.94
48.42

67.93
38.12
12.52
3.00
0.39

3.2. Further classifications using Recognition2

1.47
0.69
0.63

1.98

1.31
1.93
6.06
0.66

3.02
1.93
0.12
0.77
0.22
0.54
0.16
47.87

19.92

10.73
Wt%

As outlined above, there are several different ways to

Area%

classify particles in Recognition2. Classification via the

42.72

18.70

12.41
1.45
0.73
0.77

1.89

1.39
2.47
6.65
0.78

4.15
3.17
0.21
1.34
0.20
0.74
0.25

presence of a specific mineral or minerals in iron particles

(liberation analysis by phases) can be applied for the entire
Freq%

42.10

17.47

13.35
1.37
0.79
0.88

1.67

1.37
2.55
6.67
0.79

3.93
4.22
0.39
1.37
0.20
0.59
0.20

group of particles under investigation or for specific texture

classes. In Fig. 8 an example of a liberation output from
No. of

Recognition2 is presented. In this graph, liberation in terms

part

429

178

136
14

17

14
26
68

40
43

14
8
9

2
6
2

of the presence of hematite in the H1M texture group for

the example in Table 1 is provided. Such a classification
Shale-unrecog

would be required if the percentage of textural ore group

Shale-kaol

Shale-VGt
Shale-hem
Part type

H1M, that is nearly pure hard hematite, is under investiga-

Table 4

Quartz
HGM

GMO
GMX
HGH

tion or during beneficiation modelling the percentage of

H2M

H1M

HGF

GOV
GOC
GOL
H2H
H3C
H2F

H1M particles recovered to the product is sought.

468 E. Donskoi et al. / Minerals Engineering 20 (2007) 461–471

Table 5
Example of ‘‘particles dimensions analysis’’ for the example given in Table 1
Area Eqv circ radius Maximum length Maximum breadth Shape factor Elongation
Mean 361.64 10.73 29.52 18.53 2.31 1.68
Variance 28,686.48 53.91 64.77 32.02 0.57 0.26
SD 169.37 7.34 8.05 5.66 0.76 0.51
Coeff var 0.47 0.68 0.27 0.31 0.33 0.30
Coeff dispers 0.38 0.19 0.20 0.24 0.21 0.22
Minimum 0.89 0.53 0.94 0.94 0.00 0.97
First quartile 243.86 8.81 24.91 14.95 1.93 1.33
Median 339.98 10.40 29.38 18.46 2.18 1.57
Third quartile 459.69 12.10 33.72 22.28 2.54 1.88
Maximum 1321.66 20.51 66.35 39.90 7.55 6.05
Number of cases 1019 1019 1019 1019 1011 1019

Table 6
Example of specific gravity, total iron and mineral content variability for the example given in Table 1
Occurrence Mean SD Coeff var Coeff dispers Minimum First quartile Median Third quartile Maximum
Specific gravity 1018 4.494 0.75 0.167 0.114 1.495 4.00 4.829 5.047 5.10
Total Fe 1016 65.031 13.068 0.201 0.107 0.00 59.488 65.887 68.123 68.546
Weight percent
Hematite 791 87.465 21.191 0.242 0.127 0.416 85.424 94.521 99.068 100.00
Vitr goeth 815 24.171 35.738 1.479 2.515 0.147 3.95 10.02 49.209 100.00
Ochr goeth 389 11.90 31.433 2.641 4.214 0.123 1.376 4.153 17.201 100.00
Kaolinite 22 66.789 33.037 0.495 0.207 14.954 80.414 100.00 100.00 100.00
Quartz 4 94.783 5.556 0.059 0.047 88.963 96.856 100.00 100.00 100.00
Pores (Area%) 430 9.135 12.381 1.355 1.633 0.135 1.979 4.97 14.054 58.498

Fig. 9 and Table 7 show calculated results of the classi-
fication of particles in the HGH textural group for the
example given in Table 4. The classification has been per-
formed on the basis of the content of hematite and vitreous
goethite in the particles. Even when the content of hematite
and vitreous goethite in the particles is very similar
(between 75% and 100%, with the majority of particles
being in the 95–100% range), this will create classes with
even closer mineral composition and density.
In Table 7, a typical output for liberation analysis by
phases is shown. It includes area% of grade class (Ar%),
average grade by weight (AvG), weight% of grade class
(Wt%), average specific gravity of class (SG), distribution
by total iron (FeD), average total iron in class (FeG),
recovery of phase A minerals (Rec), which in this case
are hematite and vitreous goethite, cumulative recovery
(C Rec), cumulative yield (C Yld), and number of particles
in each class (N).
In industrial applications, classification by total iron
(Liberation by Total Iron) is also sometimes necessary.
Fig. 10 shows the ‘‘Liberation by Total Iron’’ output and
bar chart for classification by total iron of particles in the
HGH textural group for the example in Table 4. As can
be seen, almost all particles (98.3%) are in the class with
grade 90–100%, which means that all particles in the
HGH group are very similar from iron content perspective.
In addition to the classifications described above,
extraction of particles from the entire group or from
already created classes can be performed with the help
of the Selection module (see Fig. 11). Particle selection
can be based on particle number, type, identification,
location in the image, size and shape characteristics,
mineral weight, areal and perimeter information, and
molecular and elemental composition. This enables crea-
tion of classes of particles with even closer mineral
characteristics.
4. Conclusion and future directions
Optical image analysis is a convenient and reliable tool
for obtaining comprehensive information about iron ore
samples. It enables different degrees of iron oxidation and
hydration to be distinguished. It can also identify the
porosity of different minerals in particles, and hence the
density of particles can be calculated. Use of calculated
chemical composition and mineral abundances within Rec-
ognition2 enables quality control of mineral identification
to be performed.
Information about the percentage of different iron ore
textural types and their chemical composition is very
important in predicting downstream processing perfor-
mance and product characteristics. The Recognition2 soft-
ware can automatically identify different ore textures and
classify particles into different ore texture groups. The main
reason for any misclassification of particles is mineral
misidentification in the initial image analysis. If the image
analysis is correct, then 95–99% of all particles are identi-
fied correctly.
 E. Donskoi et al. / Minerals Engineering 20 (2007) 461–471 469

Fig. 8. ‘‘Liberation by Phases’’ form in Recognition2.

Table 7
Classification of particles in the HGH textural group for the example in
Table 4 based on the content of hematite and vitreous goethite in the
particles
Grade 75–85 85–95 95–100
Ar% 1.94 11.81 86.24
AvG% 84.29 92.12 99.08
Wt% 1.66 11.69 86.65
SG 4.16 4.81 4.89
FeD 1.39 11.55 87.06
FeG 55.78 66.07 67.15
Rec 1.43 10.98 87.59
C Rec 1.43 12.41 100.00
C Yld 100.00 98.57 87.59
By weight
Hem 56.01 82.36 84.77
VGt 28.27 9.76 14.31
OGt 0.76 7.88 0.92
Kln 14.95 0.00 0.00
Fig. 9. Example of classification of all particles in the HGH textural
group on the basis of the presence of hematite and vitreous goethite in the By volume
particles. Hem 45.74 77.72 81.27
VGt 29.43 11.74 17.49
OGt 0.88 10.53 1.25
Kln 23.95 0.00 0.00
Classification by ore texture creates classes of particles
with quite similar mineral composition, and hence the N 1 23 154
N% 0.56 12.92 86.52
density and chemical characteristics are also similar.
470 E. Donskoi et al. / Minerals Engineering 20 (2007) 461–471

Fig. 10. Liberation by total iron output for the HGH textural group for the example in Table 4.

Fig. 11. Example of the ‘‘Selection’’ form in Recognition2.

Procedures such as ‘‘Liberation by Phases’’, ‘‘Liberation by where better predictions can be achieved when specific par-
Total Iron’’ and ‘‘Selection’’ can create particle classes with ticle characteristics in every class are closer. For example,
even closer mineral and chemical composition. This is use- in hydrocyclone modelling it is important that particles in
ful for modelling purposes (see Donskoi et al., 2006a) the same class have close density and size characteristics.
 E. Donskoi et al. / Minerals Engineering 20 (2007) 461–471
In magnetic separation modelling, particles in the same
class need to have similar sizes, density and magnetite or
hematite content.
To have reliable figures on the abundance of different
iron ore textures in a sample, the number of particles
analysed needs to be statistically significant. It depends
on the number of different texture groups within the ore,
the abundance of ore texture groups of interest and the
acceptable statistical error for the study. Even though the
initial image analysis is partially automated, significant
labour is required for the initial image analysis using
KS400, especially if manual editing is necessary to obtain
correct mineral identification. Currently CSIRO Minerals
is working on methods for better mineral identification,
more comprehensive texture classification and full automa-
tion of image analysis.
Acknowledgments
The authors wish to thank staff of the CSIRO Iron Ore
Processing group, for valuable suggestions during the deve-
lopment of Recognition2, especially Daniel Saunders who
took part in the development and James Manuel and Simon
Campbell-Hardwick for providing classification selection
criteria. The authors would also like to thank Ralph
Holmes for valuable discussions and editing of this paper.
References
Benson, S., Showers, G., Louden, P., Rothnie, C., 2001. Quantitative and
process mineralogy at Tiwest. In: Int. Heavy Minerals Conf. AusIMM,
Fremantle, pp. 59–68.
471
Box, J., Phillips, J., Clout J.M.F., 2002. Use of geological material types
for predicting iron ore product characteristics. In: Proceedings of the
first Japan–Australia Symposium on Iron and Steelmaking, Fuwa-
Ward Symposium, ISIJ, Kyoto University, April 4–5, 2002.
Clout, J.M.F., 1998. The effects of ore petrology on downstream
processing performance: a review. In: Mine to Mill 1998 Conf.
AusIMM, Brisbane, pp. 43–50.
Danti, K.J., Cross, K.C., Halsall, C., 1993. A high performance, low cost
image analysis system for IBM compatible computers: practical
applications in mineral processing and geology. In: Programme and
Abstracts ICAM ‘93. Mineralogy in the Service of Mankind, pp. 47–
49.
Donskoi, E., Clout, J.M.F., 2005. ‘‘Recognition’’ – a specialised software
package for iron ore characterisation. In: Iron Ore 2005. AusIMM,
Fremantle, pp. 203–211.
Donskoi, E., Suthers, S.P., Campbell, J.J., Raynlyn, T., Clout, J.M.F.,
2006a. Prediction of hydrocyclone performance in iron ore benefici-
ation using texture classification. In: XXIII International Mineral
Processing Congress, Istanbul, pp. 1897–1902.
Donskoi, E., Clout, J.M.F., Manuel, J.R., Zhang, Y., 2006b. Mathemat-
ical modelling and optimization of iron ore sinter properties. In:
Fourth Int. Conf. on Mathematical Modelling and Computer Simu-
lation, Israel, MMT-2006, pp. 39–49.
Gottlieb, P., Wilkie, G., Sutherland, D., Ho-Tun, E., Suthers, S., Perera,
K., Jenkins, B., Spencer, S., Butcher, A., Rayner, J., 2000. Using
quantitative electron microscopy for process mineralogy applications.
JOM (April), 24–25.
Gu, Y., Guerney, P., 2000. Improving flotation plant performance with
the JKMRC/Philips Mineral Liberation Analyser. In: Flotation 2000.
Miner. Eng. Int., Falmouth.
Lin, D., Gomes, C.O., Finch, J.A., 1995. Comparison of stereological
correction procedures for liberation measurements. Trans. Inst. Min.
Metall. 104, C155–C161.
Mandile, A.J., Johnson, N.W., 1998. Quantitative mineral analysis of ores
and minerals processing products using X-ray diffraction. In: The
Mining Cycle, AusIMM’98. AusIMM, Melbourne, pp. 271–275.
Zhang, N., Whiten, W., 2001. Determining mineral composition from
assays. In: Int. Heavy Minerals Conf. AusIMM, Fremantle, pp. 81–85.