Mineral liberation, association and textural information about processed iron ore is vital for understanding and predicting iron-ore downstream processing performance. Such information is usually obtained by two major imaging techniques – Optical Image Analysis and Scanning Electron Microscopy. Both techniques use polished epoxy resin blocks with embedded sized particles suitable for imaging. Each particle section can be classified to a certain liberation or textural class. If the amount of analyzed particle sections is sufficient for statistical validity, the abundance of particles in each class gives an objective description of the whole sample.
Different particle sections in the block can touch each other and the processing software may consider such particles as just one particle. If the touching particles belong to different liberation or textural classes this will most probably result in distortion of the liberation, association and textural data of the sample.
CSIRO developed the new Mineral4/Recognition4 software for textural characterization of different ores, sinters and coke with a novel image analysis processing routine for separation of touching particle sections. This processing is based on two major methods: modified watershed separation and modified binary erosion separation. Depending on the ore type, particle size, imaging magnification and other factors, both methods can be adjusted to obtain a better final result. The article highlights the negative effect of particle touching, the importance of separation, and improvements made by distinct methods to incorrect separation by comparing different cases, including those when touching particles are not separated at all, separated manually and when they are separated automatically by each method, or their combination, with different options. The article demonstrates the effects that touching particle sections have on liberation and textural iron-ore characterization and shows that the proper use of automated separation developed within the Mineral4/Recognition4 system significantly reduces errors introduced by touching particle sections.
Increases in iron and steel production in recent decades and the subsequent need for vastly increased production of the necessary raw materials (e.g. iron ore) have led to the exploitation of ore bodies with new, complex textures and mineral composition. The characterization of such ores is vital to understand the routes for successful beneficiation, processing and agglomeration, as well as optimization of those processes.
Characterization of iron ores requires a significant amount of information about the ore, including: porosity, hardness, grain size, chemistry, mineralogy, texture, liberation and mineral associations. Two major methods for obtaining textural, liberation and mineral association information are Optical Image Analysis (OIA; Pirard et al., 2007; Gomes & Paciornik, 2008a and b; Donskoi et al., 2010, 2015, 2016b, 2017) and Scanning Electron Microscopy (SEM; Gottlieb et al., 2000; Gu & Guerney, 2000; Maddren et al., 2007). Both techniques use polished blocks with iron-ore particles embedded in epoxy resin for ore characterization and both have certain advantages and drawbacks (Gomes & Paciornik, 2008b; Donskoi et al., 2014). For example, different particle sections in the epoxy blocks can touch each other and this may result in problems with correct ore characterization.
This article highlights the negative effects caused by particles touching and discusses the various methods for improving particle separation by comparing different analysis cases using a complex iron-ore reference sample. The methods are compared and contrasted using the Mineral4/Recognition4 OIA software suite developed by CSIRO which provides a choice of different particle separation methods or their combination with a wide range of adjustable parameters and novel modifications to existing approaches.
Statement of the problem
Before being mounted in an epoxy resin block for analysis, either by SEM or OIA methods, the iron-ore particles are usually sized, so that each block represents a particular size fraction. After polishing, the block surface is imaged under a microscope using a certain magnification and then image analysis software identifies each particle section separately. In OIA a single polished block, typically 25 or 30 mm in diameter, is usually characterized by analyzing dozens or even hundreds of individual images. The acquisition of such a large number of images is usually performed in an automated mode using a microscope fitted with a motorized moving stage. Depending on the requirements each such image can be an original digital camera shot, possibly cropped to reduce aberrations and non-uniformity of illumination, or a combination of multiple shots digitally stitched together. To be identified as a single standalone particle, the object in the digital image should have no pixel connections with other objects; in other words, 100% of its boundary should connect either with epoxy or with the image frame. If objects touch each other, that is, if pixels belonging to different objects are adjacent to each other, they, in the absence of special processing, are considered as one particle. Such touching particle sections might randomly come into physical contact during the polished block preparation, or the epoxy area between them might be too small or not too clear to be optically or digitally resolved at the current magnification. Figure 1 shows an example of such touching objects, where three distinct particles are in close proximity to each other.
Figure 1a and b clearly shows that these touching particles exhibit very different iron ore textures; two particles are nearly pure hematite, although showing different porosities, and the third (center) is a goethitic particle. In both textural and liberation classifications the goethitic particle will report to a different class than the two hematitic particles. However, because they touch or nearly touch each other, from a simple image analysis point of view they are “connected” and so considered as one particle. The analysis outcome will therefore be quite different. In this example, using the textural classification described in more detail later in the article, the pseudo-particle 33 in Fig. 1c will report as a non-liberated hematitic/goethitic particle belonging to the Hematite-Goethite Hard textural class. If, however, particles are analysed individually, e.g. with a proper separation routine in place, as in Fig. 1d, particle 29 will be identified as Dense Hematite Very Hard, particle 32 as Vitreous Goethite Dense, and particle 35 as Hematite Medium Hard (see more details of textural classification later in the paper), all three being well liberated particles of different hardness. This situation could be further complicated if one of the touching particles was pure gangue (e.g. quartz) as that may result in the whole pseudo-particle being classified as belonging to a gangue class because of its iron content being too low.
The above example clearly shows that if touching particles are not separated during image analysis, the dimensional, liberation, association and textural characterization of the ore will be distorted. It should be noted though that touching particles do not directly affect the estimation of mineral abundances as the measured mineral areas are the same; however converting areas to volumes and therefore abundances normally requires application of some stereological correction mechanism and its outcomes may depend on separation.
CSIRO Mineral4/Recognition4 OIA suite in the context of previous work
The problem of touching objects is not new to image analysis and attempts to solve it have been undertaken in different OIA applications. Two major methods used by researchers were the watershed segmentation and erosion-dilation cycles. The work described in this paper was performed using the CSIRO Mineral4/Recognition4 OIA suite which employs both techniques in modified forms specifically designed to meet the needs of mineral particle separation. The two techniques can be used individually or in combination. The term “separation” is used within this article for image analysis techniques allowing individual particles that touch each other in the cross-section of the polished epoxy block not to be falsely recognized as one particle. It should not be confused with physical separation of ore particles during downstream processing, as particles that happen to touch each other in the epoxy block are not physically attached to each other in the original sample.
In addition to watershed segmentation and erosion-dilation cycles, other advanced OIA methods for object separation have been previously developed. Van den Berg et al. (2002) estimated the angle of the contact wedges in the outline of each pair of suspected touching grain sections and if this angle was smaller than a user-defined threshold value, intersection between grain sections was determined. Faessel & Courtois (2009) used a gap-filling method applied to the skeleton of the image background, while Filho et al. (2013) used optimal path forest operations. The applicability of such methods to iron ore is limited though, mainly because they are often based on some assumed knowledge about the sizes and shapes of objects being separated after imaging, e.g. rocks of a certain size or crop grains. Unfortunately the cross-sections of iron ore particles often have very complex shapes. In addition, the distribution of observed grain sizes can also be very wide even for well-sized fractions. This is because the cut during grinding and polishing of the epoxy resin mount for subsequent optical examination can not only go through the bulk of a particle, but also through the very top or bottom of it, thus producing much smaller sections. At the same time elongated particles may produce larger cross-sections than round ones obtained by the same sizing procedure.
Of the two methods used in Mineral4/Reconition4, watershed segmentation is probably the most popular and robust compared to all other object separation methods in OIA (Russ, 1998; Donskoi et al., 2015). The method receives its name from a topographic analogy: if a terrain area with a complex relief is filled with water that is poured onto it from the top, the water first appears in the deepest points (basins). As the water level rises, higher terrain features (watersheds) prevent the lowest points from connecting, leading to a series of isolated regions. A digital image, that is, a pixel map can be considered as a representation of such terrain, where the brightness of each pixel corresponds to height. A software algorithm can thus be applied to determine basins and watersheds, the latter considered as separation lines between basins (Russ, 1998). In order to apply this approach to iron ore characterization tasks, the image needs to be inverted, due to the fact that objects to be separated usually consist of minerals that are brighter than epoxy. The strong difference in reflectivity of different minerals makes direct application of watershed separation to grey scale ore images impractical; Mineral4/Recognition4 therefore applies the method to a specially prepared particle map, which will be discussed later.
The problem with a watershed segmentation approach, however, is that it tends to over-segment objects of complex shapes, in particular, those containing large concavities or protruding elements. Figure 2b shows two examples where such particles were erroneously separated by watershed segmentation. Figure 2c also shows how Mineral4/Recognition4 can identify and remove such incorrect separations.
To reduce the amount of erroneous separations many different modifications of watershed segmentation have been proposed. A fuzzy supervised classification procedure and a genetic algorithm in order to build the elevation map were used in the watershed algorithm proposed by Derivaux et al. (2010). Couprie et al. (2009) worked with seeded image segmentation and proposed the use of added graph cuts, random walker, and shortest path optimization algorithms to optimize watershed procedure. Qin et al. (2013) studied separation of touching corn kernels and applied an extended-maxima transform.
In all these methods the watershed algorithm was modified before the watershed was drawn. The majority of such modifications can be made if preliminary information about the shape and size distribution of objects to be separated is available in advance. As already mentioned, iron ore particle sections can vary quite significantly in size and shape and may show very irregular boundaries. For such separations a unique method is proposed in this article, where first an adjustable modified watershed algorithm is applied, and after that the validity of each separation is analyzed and, if certain criteria are not met, the separation is removed (see Fig. 2c). The details of this method are provided later when discussing Watershed separation in Mineral4.
For some cases an erosion–dilation cycle method can perform better particle separation than watershed segmentation. In this method mathematical morphology is used to disconnect the touching regions (Shatadal et al., 1995). Namely, in the first stage a binary map of the objects is created, following which this map is eroded to the extent when touching objects become separated. Later dilation is performed using an algorithm preventing connections between growing regions, and the newly established boundary is used as the mask for the separation of original objects. The method is rather simple and provides good results for convex grains with sharp edges, but may fail when particle shapes are fitting into each other in a dovetail manner. It also tends to produce false separations when objects contain relatively narrow necks between their parts. This method was also found to be very useful in Mineral4/Recognition4, especially as its implementation allowed a much wider choice of options, that is, watershed segmentation only, or erosion-based segmentation only, or a combination of both, to be used in certain complex scenarios. One example of such scenario is when a thin protruding part of one particle touches another particle, as shown by the arrow in Fig. 2c. Due to the specifics of the algorithm, which predominantly focuses on the bulk characteristics of objects, watershed segmentation is likely to ignore the finer details and therefore cut through the protruding part; such a cut-through may then be removed by software as erroneous. Erosion-based separation though will create a valid separation in the thinnest point of contact (Fig. 2d). For the sample considered in this article the erosion-dilation cycle method actually provided slightly better separation compared to watersheds.
Particle separation methodology in Mineral4
The work described in this article was performed using the CSIRO Mineral4/Recognition4 OIA package, which is primarily aimed at optical image analysis of iron ore fines (Donskoi et al., 2010), although it has been successfully applied to a wide range of other OIA tasks, including characterization of ores other than iron ore, iron ore sinters and coke (Donskoi et al., 2017). In this section, a brief overview of how particle separation is implemented within the package is provided.
Mineral4 is the component of the package directly responsible for optical image analysis, as well as other important tasks such as automated image acquisition. Recognition4 is the data processing and reporting component of the package used for post-processing of data measured by Mineral4. All the image analysis settings that affect particle separation are adjusted in Mineral4. It is therefore Mineral4 that is being referred to in this paper most of the time. This component is an extension of the Zeiss AxioVision (2006–2010) microscopy software suite and the basic image analysis functions used by Mineral4 belong to Zeiss software libraries. These functions implement standard binary and grey scale image analysis operations described in detail in the literature (Seul et al., 2000).
Initially, separation processes within the Mineral4 analysis workflow are discussed, followed by other operations that may affect particle separation or be affected by its results. An overview of the Particle Separation/Join module is then provided, followed by a more detailed explanation of the two automated separation routines, Separation by Erosion and Watershed Separation.
Separation within the Mineral4 workflow
Mineral4 has a well-defined analysis workflow in which each general operation, or rather set of closely related operations, has a specific place (Donskoi et al., 2010, 2015, 2016b, 2017; Poliakov & Donskoi, 2014). Such sets of operations are grouped into modules with every module having a dedicated form within the software allowing the operator to set individual parameters related to each specific operation and to observe the effect of those settings on the image processing. In fully automated mode the forms are not displayed so corresponding calculations occur according to analysis profile settings without the operator’s intervention. The analysis profile is a pre-recorded set of choices for different options and corresponding parameter values in various processing modules. The output of every module is a binary map, or in many cases a set of maps that represent each processing stage of the image being analyzed (Fig. S1, freely available as Supplementary Material linked to this article at https://pubs.geoscienceworld.org/eurjmin/). For example, the Particle Identification module produces a map of all known particles (as opposed to the epoxy), and following it in the workflow the Mineral Identification module executed for each individual mineral produces a map of that mineral. Later in the workflow what can be collectively described as “correction maps”, that is, maps used to re-assign ambiguously or falsely identified areas to their correct identification, are generated. All through the workflow the maps are being combined and refined until in the end the final set of maps, identifying all known particles, minerals and porosity, is obtained and measured.
In this workflow particle separation occurs between the initial Particle/Mineral Identification and correction procedures. By the time the “correction” group of procedures begins all particle matter has to be fully identified. Particle matter, including porosity, will not be added or removed during correction but only identified or re-identified as certain mineral. The Particle Identification by itself, while providing an important basis for further operations, cannot guarantee full identification of all particle matter, mostly because the wide reflectivity range covering all minerals does not allow for fine level adjustment, especially if it partially overlaps with the reflectivity range of epoxy. Such adjustments, possible on an individual mineral level, may further improve the particle map and help finalize it. It is therefore the moment after the initial particle map and mineral maps are combined when the resulting particle map can be considered fully defined, and therefore be subject to separation.
An operation that also requires a fully defined particle map is Porosity Identification. This is performed after Separation, as the porosity map, once defined, is then combined with the particle map for the subsequent processing. In a number of cases, in particular where two adjacent particles touch more than once, this may “solidify” the combined pseudo-particle making it impossible to be properly separated. It is therefore mandatory that Separation is performed before Porosity identification. At the same time it must be noted that identified porosity often closes substantial gaps in highly porous, friable particles that separation routines may attempt to further break down. To prevent this, particles can be “filled” prior to separation, and that will be discussed in more detail later.
It should be clear that the quality of particle separation may be affected by the quality of image analysis operations performed prior to separation. The very first important factor here is the preparation of polished blocks for imaging (Donskoi et al., 2015, 2018). Without going into detail regarding sample preparation methodology, it should be noted that dry-sized fractions are in many cases much harder to analyze properly compared to wet-sized ones because of the presence of adhering and non-adhering ultrafines. If this is the case, separation is also likely to be affected. In the same manner, properly sized fractions (captured between sieves where the coarser sieve size is about the square root of two times the finer sieve size) can be analyzed better than samples with much wider size ranges. This is because particles in samples with a wider size range are often packed much more tightly as the smaller particles tend to fill the gaps between the larger ones. Also, separation settings that work well on larger objects may ignore the small ones, and applying “stronger” separation, enough to separate the small particles, will probably break down large friable objects and detach parts of others.
Certain actual image analysis routines that occur prior to particle separation may have effect on it which should be taken into consideration. First of all, it can be suggested that during Image Improvement a delineation (usually to a value of 2–3 pixels) of the image is performed. The main purpose of delineation is the elimination of edge effects which occur on the boundaries between two areas of significantly different reflectivity, e.g. brighter minerals such as hematite and epoxy. Because of both optical and digital interference such boundary areas in digital photomicrographs may be represented by pixels of intermediate reflectivity corresponding to darker minerals that are not actually present in this area of the sample. Performing delineation increases the sharpness of the digital image and therefore reduces the amount of falsely identified darker matter, by changing the individual pixel’s brightness to either that of a brighter mineral or to that of less bright epoxy. Mineral4 has correction routines to further improve individual mineral identification affected by border effects, but from a separation point of view borders that are less fuzzy generally mean better distinction between nearly touching individual particles and therefore better separation.
Secondly, particle edges themselves often have visible relief making particle borders appear dark. The epoxy immediately adjacent to edges is also often darker or brighter than usual because of the unpolished surface of particles directly underneath the epoxy. Thus, the border epoxy area can mistakenly be identified as non-epoxy particle area during Particle or Mineral Identification, especially if there is some mineral with reflectivity close to that of epoxy present in the sample. The multi-threshold capabilities of Mineral4 (Donskoi et al., 2015) allowing more precise identification of individual minerals make such misidentifications far less likely, but certainly it should be kept in mind that particle or mineral identification in the reflectivity range close to that of epoxy should be done very thoroughly as errors there may affect particle separation.
During the final step of the workflow, when performing measurements of the image analysis results, Mineral4 provides various means for size-based filtering of the measured objects. In the most typical scenario, particles smaller than a certain size (e.g. half of the minimal size of the analyzed size fraction) are excluded from consideration to enable the Large Section stereological correction (Lin et al., 1995). Such a stereological correction is particularly important for characterization studies because smaller objects normally correspond to particles that have been sectioned closer to the side/end of the particle than to the middle. Statistically such sections have simpler morphologies and can therefore introduce significant bias into the textural and liberation measurements. It is therefore a standard practice to exclude smaller objects from measurement. In this paper this process is referred to as Small Section removal.
If Small Section removal takes place during analysis, subdivision of objects by separation will necessarily result in more objects being rejected by this filtering. At a first glance, assuming objects being removed do not have any preferential mineralogy, there is going to be no effect apart from losing some portion of the measured sample section area and therefore reduction of representativeness, which possibly can be compensated by analyzing more images. The reality is more sophisticated. First of all, if one mineral preferentially tends to appear in smaller grains, e.g. protruding from particles or adhering to their surface, over-separation will result in preferential exclusion of this mineral, so the overall statistics may be slightly biased. Even more importantly, certain minerals, in particular if they are highly porous or close in color and reflectivity to epoxy, may be more prone to over-separation and therefore to preferential exclusion.
Mineral4 allows the operator to limit not only the minimal but also the maximal particle section size to be measured so that all particles above a certain size limit can also be discarded. While the primary goal of this mechanism is to allow measurements of “virtual” size fractions within a wide size distribution, it can also be used to filter out larger than expected objects that should not appear in the size fraction being measured. Such objects are most likely unseparated touching particles, and so to a certain extent such filtering can improve the results of poor separation, at the potential cost of a reduction in the measured sample area.
Mineral4 separation module overview
Figures S2 and S3 (in Supplementary Material) show the Particle Separation/Join module form of Mineral4. The controls in the top-right corner of the form (i.e. “View”) are used to adjust the size of the image on the display. The module also allows changing between the two different views. The “Original” view displays the image itself overlaid with separation lines and screenshots of this view have been used widely throughout the paper. The alternative, “Black/color” view, shows the same separation lines over a particle map, with each individual particle painted with a random false color. This view allows easily distinguishing between separated and unseparated particles.
The rest of the form controls are combined into three groups, responsible for two automatic and one manual mode of separation. The two different automatic routines, “Separation by Erosion” and “Watersheds Separation”, are completely independent. According to the operator’s choice, recorded in the analysis profile, either one of those routines or a combination of both can be used. Each of these routines will be described in more detail in subsequent subsections.
A Manual Separation routine is also provided with this module, similar to most other Mineral4 modules that also include manual routines. While those routines are not used in high throughput automated image analysis, they can be useful when performing specific analysis of individual images or for research purposes. In particular, while working on this paper a benchmark analysis of an image set was performed using manual separation for the purpose of generating objective sample measurements. All separation options were then compared to this benchmark when discussing textural and liberation analysis.
Figure 3 shows how color-coding of individual particles in the “Black/color” view can assist the operator estimate the quality of separation. A number of objects of interest have been outlined in Fig. 3a. Of those objects, A and B are porous particles that have small areas detached even before separation, as shown in Fig. 3b. Objects C and F are also porous particles, whereas D and G are touching particles (Fig. 3b shows them with the same false color) of somewhat different textures. E and H represent a boundary case and can correspond to either two touching particles of the same texture or a single particle with a crack going through the section.
Figure 3c shows the effect of watershed separation applied to the image. Particles D, G and E are now correctly separated however the highly porous particles C and F are now also broken by separation lines. Note that particle B has several separation lines attempting to cut through it, but all those incorrect separations have been automatically removed by Mineral4. Finally, the measure map in Fig. 3d shows the overall effect of separation on what is being measured. Individual particles in this map are color-coded and numbered, while fragments colored white represent small detached objects that have been discarded due to the small section removal. It can be seen that unattached parts of particles A and B as well as some incorrectly separated parts of particles C and F are excluded from the measurement while the remainder of particle F has been subdivided into two separate objects. Note though that the overall area affected by incorrect separation in this example is insignificant compared to the correctly separated areas (D, E and G). Due to the great variety of sizes, shapes and porosity of iron ore particles it is impossible to reach a perfect automated separation, especially keeping in mind that individual images may have their own peculiarities. Therefore what is considered good separation means maximum gain with minimum loss.
Of the two automated separation routines used by Mineral4, Watersheds Separation is the most well established and highly refined. It is important to notice that separation of ore particles is not the only application of this routine within the system. Others applications include, for example, segmentation of continuous highly porous structures into structural elements for subsequent analysis. Some of the settings, such as the various smoothness parameters (see “Watersheds Separation” frame in Fig. S3), are more applicable to those studies. For the work described in this paper they have been set to reasonable values typical for particle separation, and not subsequently changed between analyses.
Figure 4a shows an example of four iron-ore particles that touch each other in the polished block section. In the absence of separation the particles form a single super-particle as shown in Fig. 4b, which will result in incorrect dimensional, textural, liberation and association measurements. As has already been discussed, Mineral4 uses the combined particle map (the original particle map superimposed with the individual mineral maps) as the input for Watersheds Separation. Prior to separation the map has to be prepared. First of all, there is an option of filling all pores in particles (“Fill pores” checkbox). Using this option helps prevent over-separation of highly porous objects. There is, however, risk of establishing solid connections between particles that touch more than once, as the space between such touching points will also be filled. Alternatively, or in combination with filling, the binary Close function can also be applied to the map a number of times (according to the “Closing Depth” setting), thereby making particles less prone to over-separation. Finally, the map is subjected to a number (according to the “Separation Coefficient” setting) of binary Erosion operations and then inverted. In the example shown in Fig. 4d (fill pores, no Close, triple Erosion) the objects are already separated after these operations and the application of Watersheds function (Fig. 4e) directly to the inverted map provides a set of separation lines that, after removing areas covered by them from the particle map, results in all touching particles being properly separated (Fig. 4f).
Although this example uses Watersheds to draw separation lines, essentially the particles have already been separated by Erosion. This is not necessarily the case for all touching particles in the sample. Whether adjacent particles are separated or not depends primarily on the combination of Close and Erosion parameters mentioned above. Also, for some samples strong Erosion may be too destructive, so weaker Erosion with subsequent extra separation may be needed. The “Fine Separation” checkbox allows the inverted map (Fig. 4d) to be subjected to Euclidean Distance function prior to Watersheds, which assigns to each black pixel a grayscale value based on their proximity to the white pixels. That allows a distinction to be made between bulk objects and weaker connections between them, making the latter ready for watershed separation. The result of the application of this function is shown in Fig. 5a.
Figure 5b shows the result of Watersheds transformation applied to Fig. 5a; it is interesting to compare it to Fig. 4e as all the watershed lines from the latter are present but some extra lines going through the weak points also appear. Figure 5c shows the effect of those extra lines – the most complexly shaped particles are now divided into sub-particles.
The example in Fig. 5c is, of course, an exaggerated one, as we have already demonstrated that, in this particular case, separation was already good before turning “Fine Separation” on. In many cases, however, using Fine Separation is fully justified as the particles are either too densely packed or too friable and therefore require much stronger closing or weaker erosion at the particle map preparation stage. The example shows though, that the separation lines present after fine separation tend to cut between particles and also through “weak” (e.g., thin or associated with large cavities) areas of particles themselves. This is the main weakness of Watersheds Separation in general, but Mineral4 is capable of overcoming this problem by applying the “Removal of Wrong Separations” routine. Figure 5d shows the final result of the same fine separation scenario with the removal of Wrong Separations turned on. In this particular example all separation lines going through particles have been removed by the algorithm while all separations between particles are still in place. While in general the routine cannot guarantee the correct removal of 100% of any wrong separations present over multiple images, it still can provide a significant improvement of the separation procedure.
Figure 6 provides an explanation of how the removal of wrong separations algorithm works. It is assumed that a correct separation line goes through an area where two particles are very close to each other to the point where they touch, while the immediate vicinity of the touching point forms two clefts, or channels. A separation line going through this connection will cover the touching point but also most of the clefts. On the other hand, an incorrect separation line cutting through the bulk of the particle may also go through some pores and dents, but will mainly cover the particle body.
Figure 6a gives a simplified view of two such separation line intersections with the particle map. The intersection outlined with green is a valid separation, the one outlined with red is an incorrect one, and only the parts of separation lines that have some particle matter covered are being considered. If a binary Close operation is applied to the particle map, it will mostly affect the narrow clefts. Fine pores will be filled too but, area-wise, the effect typically would not be significant. On the contrary, wide dents and large pores would not be significantly affected unless the Close operation is extremely strong. If particle areas covered by the separation line sections in both Fig. 6a (before Close) and Fig. 6b (after Close) are considered, it is obvious that the area covered by the correct separation line has grown considerably after closing, while for the incorrect line the change is fairly small. By introducing the “Limiting Ratio”, that is, the ratio of the particle areas covered by the separation line after and before the close operation, the two types of separations can be distinguished and the incorrect ones removed. A separation line cutting directly through a particle (e.g. the red line in Fig. 6) will have the ratio close to one, but in the presence of a strong cleft pattern (e.g. the green line in Fig. 6) the ratio will likely be at least two or above, so that choosing a Limiting Ratio of 1.5 will easily distinguish between the two cases. In reality, due to the presence of less clear cases the parameter may require fine tuning to achieve optimal results. Mineral4 also allows the operator to choose the number of Close operations to suit a particular case. It should be noted that separation line thickness, the optimal value of which depends on particle size, magnification, particle boundary smoothness, etc., will also affect the areas being compared and therefore the quality of removal.
Watersheds Separation with “Removal of Wrong Separations” routine subjects the particle map to a relatively strong separation followed by filtering out of the incorrect attempts. Although this approach has been shown to provide good results in many cases, one possible drawback is that in certain cases valid separations between particles may also be removed. In a typical scenario such a separation will go through a thin but relatively long particle segment that is touching another particle. The separation line may cut through this protruding part without forming the cleft pattern, and will be subsequently removed. An example of such an incorrect removal was shown earlier in Fig. 2c. To overcome this issue, a simpler erosion-based separation option has been also added to Mineral4 (see “Separation by Erosion” frame in Fig. S3).
Applying binary Erosion cycles according to an operator set “Erosion depth” to the particle map (Fig. 4c) breaks it into individual objects (Fig. 7a), which is similar to particle map preparation within Watersheds Separation routine. It is assumed that all particles have already been separated after this stage. Further, the map is subjected to binary Scrap, when all objects smaller than a set “Scrap size” are removed, which helps avoid excessive separation of protruding parts of particles. The Erosion is then reversed by applying a function opposite to it. In this particular case binary Thinning of the epoxy map is used, which can be considered an equivalent of Dilation which prevents dilated objects from establishing connections. As the Thinning applied is stronger than the original Erosion the epoxy between the dilated objects forms distinct thin lines in the areas where the original objects were in close proximity to, or touching, each other (Fig. 7b). These lines can be identified (Fig. 7c) and then used as separation lines (Fig. 7d). There also is an option to fill the particle map before separation which reduces breakout of particles during the erosion stage. It is similar to the option provided for Watersheds Separation but is applied independently, so it is possible, for example, to use a combination of watershed separation with fill and erosion without fill.
The checkboxes provided within both “Separation by Erosion” and “Watersheds Separation” frames allow the operator to observe the effect of the two different separation methods separately or in combination. Most of the cases presented further in this paper will use only one of the methods to make the effect of individual settings more clear. It can be expected, however, that well-defined analysis profiles will make use of both methods combined. For example, in the case presented in Fig. 2 the combination of Watersheds Separation with Removal of Wrong Separations (Fig. 2c) and Separation by Erosion (Fig. 2d) gives the best result.
The effect of different separation methods on ore characterization
To study the effect of particle separation on the measured properties of iron-ore fines, a sample of Australian hematitic/goethitic iron ore (−250+125 μm size fraction) was examined. The sample was mounted in a polished epoxy block and then imaged at 0.53 μm/pixel resolution with ×200 optical magnification using 2 × 2 mosaic stitching. Zeiss AxioPlan 2 Imaging microscope equipped with AxioCam HRc colour digital camera and a moving stage was used for this work. The resulting set of 189 images of approximately 1750 × 1750 pixels each (the actual number of pixels varied slightly between images because of stitching) was subjected to image analysis a number of times using the CSIRO Mineral4/Recognition4 package (Donskoi et al., 2010, 2017). All analysis and reporting parameters were the same except for certain parameters within the Mineral4 Particle Separation/Join module, which varied between analysis sessions, thus resulting in particle separation being performed differently. In particular, particle sections with an area below 1000 square micrometers were excluded from the measurement in all analysis cases as a part of the large section stereological correction (it should be noted that, depending on the task, the large section stereological correction exclusion limit may be significantly higher). The results of each analysis were then compared. To provide a reliable basis for such a comparison one of the analysis sessions was performed in Manual Separation mode – that is, all touching particles were separated by the operator. This manual analysis session was considered a benchmark against which all other analysis sessions were assessed.
It is important to understand that, while the manual analysis session gives the best possible representation of individual particles that an individual operator can provide, it is not necessarily perfect or free of bias. Figure 3a can provide some useful examples. It is often reasonably easy to distinguish two touching particles from an individual particle of a complex shape (G), especially if they have different mineralogy and texture. However there also are cases when the difference is not so obvious (D and E), so it is not clear whether the objects in question are two sections of the same particle or two particles with similar texture touching each other. In many cases there appears to be some “other than epoxy” material in the image connecting the two objects together. That material can in fact be some genuine dark (low reflectivity) mineral possibly providing a robust physical connection between the particles. At the same time it could be adhering ultra-fines connecting the particles in the section but not actually holding them together – such objects would behave as individual particles during downstream processing. Also, the perceivable connection may not exist and the “other than epoxy” color may correspond to under-surface reflections, that is, to particle material that is not in the section being studied but underneath (that material may or may not be part of the particle but is certainly not part of the section). Sometimes it is possible to distinguish between such cases by a very close inspection of the “connection”, possibly at higher magnification. However this operation is labor-intensive and can only be performed during a manual study of the sample.
To complicate things further, it is not unusual for an individual particle section to be nearly separated by a crack or large pore, while the overall shape and morphology may be suggesting that this is in fact a single particle (Fig. 3a, particles H and possibly E). These examples emphasize the fact that the operator may prefer to consciously avoid separating such objects if some connection is present. Even more importantly, touching objects displaying the same morphology and a continuous shape may not even be considered by the operator as candidates for separation in the first instance, as in the operator’s perception this is simply one particle. As the texture of the separated and unseparated objects is essentially the same, separating them will not have any noticeable effect on textural and liberation/association figures. Based on that, during Manual Separation special attention has been paid to touching objects displaying different morphologies, even if the difference was only slight. Such objects were studied very thoroughly before a decision was made on whether they should be separated.
Altogether, along with Manual Separation, a number of fully automated analysis sessions were performed to cover typical separation choices an operator might face. Table 1 shows the list of codes used for the eight analysis cases. These cases cover the typical scenarios of reasonably good separation and over-separation that may occur during analysis, a case with no separation at all was also included. Other analysis cases from the full study, not included in this article because of space limitations, also confirm the findings presented.
It should be noted that three of the cases can be directly compared to WS_BASE as “using stronger separation”, or “being more separated”. In particular, WS_NO_FILL used otherwise the same separation procedure over a non-filled particle map, WS_NO_REM didn’t remove wrong separations, and COMBO1 combines WS_BASE with erosion-based separation. This observation provides a good basis for comparing analysis outcomes for those cases.
The effect of separation on textural classification
Table 2 shows the detailed results of textural classification (Donskoi et al., 2007) of the sample for analysis in the absence of separation and their comparison to the manual separation results; more results showing also two automated analysis cases with separation, WS_BASE and EROSION, are presented in Table S1 in Supplementary Material. Rows in the first column of the top part of the table list the generic textural classes corresponding to those in a typical iron ore classification used by Recognition4; their estimated hardness (out of three possibilities – hard, medium and soft) is also provided. For non-shale textures, information is also provided on whether they are considered hematitic, goethitic or combined, that is, hematitic-goethitic (in the H/G column). For all cases the number of particles in each textural class and the total area percentage of all particles in that class in the sample are provided. The overall measured numbers of particles and sample areas are also given. The bottom part of the table shows the same statistics calculated for certain meaningful groups of textures: shales (including all “Shale” groups and Quartz) and non-shales (the latter also subdivided into Hematitic, Goethitic or Hematitic/Goethitic); and Hard, Medium and Soft textures separately. For all cases except for manual separation, the difference from the manual analysis is also presented; those differences, or “errors”, form the basis for comparison of the different separation methods. For the manual case only, the data in the bottom part of the table are the actual figures, not “errors”.
While Table 2 provides detailed information both on the textural classification system used in this research and on how individual data for each textural class, as well as group of textures, are calculated, Table 3 lists the significant findings (particle number and area values as well as calculated errors for different textural groups) for all considered analysis cases. Essentially it is the pivoted copy of the bottom section (starting from the “All particles” row) of Table 2 with all analysis cases present. Explanations are provided further for Table 2 and these also apply to Table 3, keeping in mind that the rows and columns are inverted.
Only textural classes present in the different analysis reports for the sample are shown in Table 2, a comprehensive classification scheme may include classes not observed in this exercise. “Zero Particle” is a special class for pseudo-particles showing no mineralogy. In most cases they represent epoxy defects and under-surface reflection. While Mineral4 can discriminate most of these based on their low average reflectivity, it is not always possible to eliminate all without losing some dark, low reflectivity porous gangue particles or particle regions. The remaining zero particles are normally discarded during Recognition4 reporting. They are listed here for consistency as they contribute to the overall measured particle area, but their effect on the estimated quality of separation was negligible.
The difference from the manual analysis is shown in both relative and absolute units, and this requires some explanation. Firstly, while the percentage change in the number of particles and, particularly, in the abundance (area percentage) for highly represented textural classes is of importance, textural classes with relatively lower abundance will exhibit very high relative changes after only a small change in particle numbers or areas. Secondly, the overall error of each analysis method can only be estimated by adding up comparable absolute values. In fact, modules of those absolute values are probably best used.
To explain the latter statement, let us consider a hypothetical situation of a generally well-separated sample that still has small grains of quartz touching 100 pure hematitic particles. The presence of quartz moves them, say, from Dense Hematite Very Hard to Hematite Siliceous. In a properly separated sample the quartz grains will be detached from hematite and then discarded because of their size, so the only significant difference, apart from a slight area loss, between the two analysis sessions will be an increase in Dense Hematite by 100 particles and a corresponding decrease in Hematite Siliceous. However by adding up the signed differences we will end up with zero, which apparently doesn’t represent the significantly changed texture of the sample. Therefore, the “error” of each analysis session is determined as the sum of absolute values of changes for each textural class or group of classes expressed as a percentage of the overall sum of values for all classes for manual analysis. Note that areas are already expressed as a percentage, while for particle numbers the percentage is calculated.
The proposed metrics allow for a comparison of the errors for different separation cases and also explore the behavior of different groups of textures depending on the separation method chosen. However the metrics also have their drawbacks. In particular, in the example described above the overall error for the unseparated sample in absolute values was 200, that is, double the number of re-classified particles. While a better estimate of the error than zero, it is obviously an over-estimate. Generally speaking, areas that are completely lost due to separation and subsequent small section removal are only counted once (but, as only small areas are removed, their contribution is also small). However objects that change classification are counted twice by area as shown in the example above. They are also counted at least twice by the number of particles, once again as in the example above, and possibly even more, e.g. a pseudo-particle of a complex texture that becomes two particles of simple textures after separation gives an overall change of three. Similarly, three touching particles are likely to give a change of four and so on. This effect is somewhat mitigated by the fact that individual objects are not being tracked and, while some classes may lose certain particles and gain others, only the net difference for the class is taken into account. Still, based on the explanations above, the authors suggest that to understand the actual error the overall error calculated by area percentage should be divided by two. The coefficient to be used with the error in the number of particles is, as shown, less obvious so we keep the original figure. The comparisons between analysis cases are still valid, however it should be kept in mind that the actual error, understood as the number of incorrectly classified particles (or pseudo-particles formed by touching particles), is less than half of that figure.
Also importantly, the particle-based metric does not distinguish between particle matter migrating from one textural class to another and also that being lost due to small section removal. Therefore, in a purely hypothetical scenario where the automatic separation routine correctly separates all touching sections but randomly “destroys” 5% of all particles, we will see a 5% error in each textural class (and overall). The measured textural characteristics of the ore will however be fully correct, as a 5% decrease in the area of each class will be matched by 5% decrease in the overall area. The area-based metric therefore is not affected by random loss of matter. In reality the loss of matter is not random as can be seen from the discussion below, and so it does affect characterization, but it is important to keep in mind that the aim of separation is to correctly characterize the sample, not to avoid any area loss (which can be easily compensated for by analyzing more of the sample). Still, both metrics appear to provide valuable information on the effect of separation. It is reasonable to say that the particle-based metric allows us to see “what” happens to the sample after separation and the area-based metric tells “how” the changes affect overall textural distribution.
The unseparated sample had significantly less particles (11,364, or 5.83% less) than the manually separated one (12,067). Its overall area was slightly (0.32%) higher though as there was some loss of measured material in the separated sample because of small section removal.
For the most abundant texture (39.18% by area), Dense Hematite Very Hard, the unseparated sample had 7.53% less particles (4510 against 4877), or 3.04% less area. It also had 39.13% more Shale-Hematitic particles however based on the number of particles (32) or area percentage (0.37%) involved, this difference was less significant. The Shale-Hematitic particles in the unseparated sample were also much larger on average than in the separated one (39.13% more particles correspond to 146.67% more area). It can be easily assumed that many of the Shale-Hematitic particles in the unseparated sample were actually cases where hematitic particles were touching gangue. This was also likely the case for other complex textures containing hematite.
The number of particles (or pseudo-particles) in the unseparated sample that changed their classification after separation was 803. This was the sum of the absolute values of differences in the number of particles between UNSEP and MAN cases in each textural class. That accounts for 6.65% of all particles and also corresponds to a 3.75% error by sample area (the sum of absolute values of differences in each class divided by two).
21.64% (or 16.06% by area) of all shales were classified incorrectly without separation, whereas only 6.44% (or 3.58% by area) of non-shale particles were classified incorrectly without separation. Keeping in mind, however, that non-shale particles were much more highly represented in the sample, their contribution to the overall 3.75% error by area was 3.53% with only 0.22% contributed by shales.
The following results from Table 3 appear to be of particular interest (unless specifically stated, all results are being compared to manual separation, which is considered as a benchmark representing the actual condition of the sample):
The unseparated sample (UNSEP) had significantly (−5.83%) less particles, so proper separation changes the particle count quite significantly. At the same time, the effect of separation on the overall measured sample area was much lower (0.32%).
In the “Measured value” section of Table 3, most separation cases are within or slightly outside the range of 1% error by the number of particles. The only analysis case significantly outside that range, WS_NO_FILL, gives a rather high over-estimate of 5.25%, which is still less than that for UNSEP by absolute value.
Also in the “Measured value” section, all cases considered, except for UNSEP, resulted in the area being under-estimated with the absolute value of the error greater than that for UNSEP. The highest error is once again for WS_NO_FILL, followed by WS_NO_REM.
Referring to the “Overall” column of the “Value by group / difference” section of Table 3:
From a textural characterization aspect, the lowest error by area was observed for EROSION (1.41%), which also gave the lowest error by number (1.99%). The WS_BASE case did not perform that well by comparison (1.77% by area, 2.64% by number) but even that result was much better (i.e. less than half the error) than the case in the absence of separation.
Other watershed-based cases with separation comparatively stronger than that for WS_BASE case provided even worse results, however only WS_NO_FILL was worse than UNSEP.
Judging by the fact that the error for the COMBO1 case, which was a combination of WS_BASE and EROSION cases WS_5_145 and E_3_1000, was relatively high (3.48% by number, 2.24% by area) and noticeably higher than for each of the individual methods participating, it can be assumed that the two methods separate particles differently, otherwise the separations would overlap and the combined effect would have been closer to the individual ones.
To test this hypothesis, COMBO2 case was introduced, which combined “weaker” versions of WS_BASE (using a higher limiting ratio, so more separations are removed) and EROSION (using a lower erosion coefficient), see Table 1 for details. This case gave a much better result (2.04% by number and 1.60% by area), actually better than WS_BASE case (it should be noted that the weaker watershed separation applied individually gave higher error compared to WS_BASE). This result suggests that combining two methods with weak (i.e. under-separated) individual settings might be the easiest way for the operator to achieve a reasonable separation.
Based on the observed results, it is possible to roughly describe the EROSION, COMBO2 and WS_BASE cases as reasonably well-separated, since the error of separation for all those cases, depending on the choice of metrics, was 2–3 times less than that for the unseparated case. Other automated separation cases can be directly compared to WS_BASE as more separated. Most of those cases, except for WS_NO_FILL, still show errors noticeably lower than that exhibited by analysis performed without separation.
It must be noted that the estimated quality of separation for each set of parameters in this experiment certainly cannot be used to suggest that one set of parameters is “good” for all possible separation tasks, while another is “bad” and should not be used for analysis in general. First of all, all analyses have been performed for a particular set of images, that is, a certain size fraction of a certain ore imaged at certain conditions. According to practice, changing just one of these conditions can significantly affect the quality of separation; in particular, experienced Mineral4 operators usually choose the analysis profile according to magnification and ore type; the magnification, in turn, is largely determined by the size fraction being processed. Secondly, the effect of changing one major parameter can be quite significant and is normally accompanied by fine-tuning of minor parameters in order to achieve comparable processing quality. This particular article, however, was aimed at comparing exactly the effects of major parameter changes.
It is also useful to compare the effect of separation on different textural groups. Looking at the relative errors for different groups of minerals, it was clear that with no separation (UNSEP) there was an obvious trend for the more friable textural classes to show much higher errors than the denser textural classes (e.g. shale vs. non-shale – 16.06% and 3.58% by area correspondingly, goethitic vs. hematitic – 4.54% and 2.13%, soft vs. medium vs. hard – 9.11%, 3.67% and 2.87%). For the better separation cases the errors for all groups were generally lower and the error for friable material often decreased quite significantly (e.g. 3.60% for EROSION). In particular, the fill procedure seems to be beneficial for the friable textures (note that the only case where the soft material error was higher compared to UNSEP, 14.35%, was the WS_NO_FILL), which means that in the significant presence of such textures using fill should be considered a priority. Significant over-separation appears to have had a detrimental effect on both soft and hard textures (referring again to WS_NO_FILL, the error for hard material was 4.36%, compared to 1.20% for EROSION).
The error for goethitic textures was also consistently higher than that for hematitic. Interestingly, the error for hematitic-goethitic textures was typically higher than that for goethitic, which can probably be explained by the presence of Vitreous Goethite Dense in the goethitic group – this low porosity mineral tends to behave reasonably well during separation as it is less likely to be wrongly separated.
A summary of the effect of separation on different groups of textures can be found in Fig. 8.
It must be noted that there are other possible explanations for different errors observed for different classes and groups of classes. One particular effect, discussed in more detail in the following section, is that in a simple case of two distinct classes the more abundant class is less affected by touching and therefore should have less separation error. This observation, however, cannot be easily extrapolated to scenarios with multiple classes and complex groups of classes. Indeed, Fig. 8b shows that hematitic-goethitic textures consistently report the highest error even though they are twice more abundant than goethitic.
The effect of separation on liberation analysis
While textural characterization is extremely important for the prediction of downstream processing performance (Donskoi et al., 2012), it has not been fully embraced by the whole iron-ore industry, so in many cases it relies on other measurable ore characteristics. One such characteristic is mineral liberation (Delbem et al., 2015; Ueda et al., 2016; Donskoi et al., 2016a; Tripathy et al., 2017), and it is natural to expect that liberation measurements too are going to be affected by particle separation. In order to test that, Recognition4 was used to generate liberation reports for all analysis cases considered in this paper.
As conventional liberation studies are based on mineral abundances, Table 4, showing those abundances calculated for the manually separated case, is provided to give a better understanding of the sample. It can be seen that the most abundant mineral in the sample is hematite, followed by vitreous and ochreous goethite.
From the earlier discussion and observations made in the previous section, it is natural to expect that an unseparated sample contains more complex textures, and therefore less fully liberated particles. It can be expected therefore that proper separation will increase the amount of fully liberated material, and that excessive separation will display the same, but even stronger, effect. In order to verify that statement, the Liberation Analysis module of Recognition4 was used. The module allows for the calculation of statistical information for 11 or 12 different liberation classes of individual minerals or “phases” (that is, combinations of minerals) within the sample. Table 5 shows an example of liberation data for “iron-bearing minerals with low impurities” phase (A), comprised of “kenomagnetite”, hematite, “hydrohematite” and vitreous goethite, calculated for the MAN analysis case. The other phase (B) includes the rest of the minerals which for this exercise can be considered “gangue or high impurity” (it should be understood that low impurity ochreous goethite deposits are also common). Every particle is assigned to a particular liberation class, e.g. 0–5%, 5–15%, 15–25% and so on, based on the combined weight percentage of the minerals belonging to phase A within the particle, note that 0–5% phase A class is also 95–100% phase B class and so on. For each liberation class a significant amount of other information is calculated. This includes calculated values such as area percentage (Ar%), average phase A grade (AvG%), weight percentage (Wht%), specific gravity (SG), distribution by total iron (FeD), average iron grade (FeG), phase A recovery (Rec), cumulative recovery (C Rec) and cumulative yield (C Yld), mineral (KMg – kenomagnetite, Hem – hematite, HHm – hydrohematite, VGt – vitreous goethite, OGt – ochreous goethite, Kln – kaolinite, Qtz – quartz) and porosity (Por) contents by weight and volume. Finally, the number of particles (N) and their percentage of the whole sample (N%) are also calculated.
The hypothesis of the effect of separation on liberation analysis is, as mentioned, that higher particle separation by OIA should mean higher liberation of phases. As a general rule, for a non-separated sample the non-liberated particles reported in the middle of the distribution are either touching particles, possibly having quite different mineralogy, or genuine particles of complex mineralogy. Correct separation of touching particles will reduce the amount of “false” complex mineralogy matter. Erroneous separation of genuine particles will, statistically, have a similar effect by possibly breaking particles of complex mineralogy into parts, at least some of which may have simpler mineralogy. An example of the latter effect will be a generally homogenous particle containing a small grain of another mineral. If separation leaves that grain in one half of a particle the other becomes fully homogenous and liberated.
Typically classes at both ends of the liberation classes distribution spectrum, i.e. that class showing fully liberated phase A (95–100% class) and that class showing fully liberated phase B (0–5% class), are of most importance for liberation studies. Unfortunately phase A in the above example is highly abundant and present in most particles. As a result, the fully liberated phase B is very rare (0.35% by weight) even though ochreous goethite has been included in it, making statistically valid observations on the behavior of this phase under different separation conditions difficult. Therefore in order to study the effect of separation on mineral liberation, a different phase A was chosen, consisting only of hematite (representing over 75% by weight of the sample).
Table 6 shows the overall weight percentage figures for 11 liberation classes for hematite in all eight analysis cases considered in this paper. In order to observe some trends the cases have been ordered by “separation strength”, low to high, which has been expressed as the total measured particle section area after small section removal (see Table 3), high to low. While it appears to be a reasonable indicator of the level of separation (that is, higher separation meant more fine objects excluded from consideration because of small section removal), it is important to emphasize that it does not directly translate into the “quality of separation”. The previous section showed that the relationship was more complex, although the “good” separation cases from a textural viewpoint were still reasonably close to the benchmark manual separation by this parameter. Note that the benchmark MAN case should be excluded from consideration when observing the trends as it was obtained differently and is supposed to contain no incorrect separations at all.
A number of observations can be made from Table 6. In particular, there was a clear trend for the amount of fully liberated hematite to grow with the increase of separation. In particular, the three rows in the bottom of the table, corresponding to cases with higher area loss, all show noticeable overestimation of fully liberated hematite, while for all other cases the amount of hematite in this liberation class was underestimated, with the unseparated case being the most significant. Each of the three cases at the bottom section are also directly comparable to WS_BASE as being more separated, and the comparison of numbers also confirms that extra separation means extra estimated liberation of hematite. It is also interesting to note that the best case from a textural viewpoint, EROSION, did not show the result closest to MAN for hematite liberation.
Before considering the effect of separation on phase B, that is, minerals other than hematite, it is worth comparing the behavior of fully liberated phases A and B for the unseparated and manually separated cases. The absolute difference in weight percentages for the 0–5% class was 1.39% compared to 2.27% for the 95–100% class. If, however, relative differences were considered then the differences were 12.35% of the class weight for the 0–5% class and only 4.81% for the 95–100% class, that is, proper separation appears to have had a stronger effect on the less abundant phase (i.e. phase B), roughly three times greater for the phase that was three times less abundant (see Table 4; note that although weight percentages have been used to compare phase abundances, the same ratio expressed in area % would be closer to two). This can be explained relatively easily. If we consider a sample with the same ratio of phases consisting only of liberated phase A and phase B particles that touch each other randomly, phase B particle is three times more likely to touch phase A particle and form a pseudo-particle belonging to a non-liberated class, than to touch another phase B particle and stay liberated. Phase A particle, in turn, is three times more likely to stay liberated than not after touching. Of course, not all particles touch each other and not all genuine particles are fully liberated, but proper separation, which reverses the effect of touching just described, should still be expected to affect the less frequent phase B more.
It is worth mentioning that of the analysis cases considered manual separation had the highest liberation of phase B. It is hard to trace a clear trend in the behavior of phase B liberation against “separation strength”, although it should be admitted that the figures are slightly higher in the bottom, more highly separated part of Table 6. The highest, and the closest to the MAN case percentage was recorded by the watersheds without removal WS_NO_REM case, which did not perform very well texturally and also gave the highest overestimate of phase A liberation. The graph in Fig. 9 illustrates these observations. For the five cases displayed: unseparated (UNSEP), manually separated (MAN), one of the best cases (EROSION) and a reasonably good case (COMBO2) from a textural point of view, and an over-separated (WS_NO_REM) case, the behavior of all classes conforms to expectations – that is, with more separation the weight percentages of non-liberated classes decrease and the weight percentage of fully liberated phase A noticeably increases. The exception was the fully liberated phase B, which was highest for manual separation. Also, it was the only class where there was a noticeable relative difference between the manual case and other texturally sound cases.
While this effect does not fully coincide with the expectations in general, the exception is still not difficult to explain based on the observations made earlier during textural analysis. Soft and friable material was significantly represented in this liberation class and, as shown earlier in this article, separation introduces the highest error for this sort of material. A significant amount of this material can be expected to be lost due to separation and subsequent small sections removal, accounting for most of the area loss, so the liberation figures for phase B, consisting mostly of such material, will be affected. Modal analysis can confirm this observation – for example, for the case where the highest area loss was recorded, the non-filled WS_NO_FILL, the weight percentages for the most typically friable mineral types, ochreous goethite and kaolinite, were 4.01% and 0.50% respectively, compared to 4.19% and 0.54% after manual separation (Table 4), which means a 5–10% loss for each. This observation again emphasizes the importance of using “non-destructive” (e.g. using fill) separation in the presence of such porous material.
Except for this minor discrepancy, it should also be emphasized that the proper separation cases were in very good agreement with the benchmark in all liberation classes, which was significantly better than in the absence of separation and also noticeably better than in over-separated cases (Fig. 9).
Correct characterization of iron ores is vitally important for the iron-ore industry, especially as old high-grade deposits are becoming depleted and ore bodies of new, complex mineralogy are being mined. Optical image analysis is one of the major tools used for iron-ore characterization. It is important therefore to ensure that the image analysis is performed to the highest possible standard, as any errors and biases may affect the quality of characterization and, as a result, the quality of resource estimation and of the ability to predict downstream processing performance.
One of the problems to be resolved in automated optical image analysis applications for the characterization of discrete objects is automated separation of touching objects. Failure to separate individual particles of iron ore may result in distortion of dimensional, liberation, association and textural characterization of the sample. A number of particle separation algorithms have been suggested over time, the most popular of which belong to two major groups of algorithms, watershed-based and erosion-based. Although those algorithms have reached a significant level of sophistication, their applicability to separation of iron ore particles remains limited, mostly due to the wide size, shape and textural range of a typical iron-ore fines sample.
The CSIRO Mineral4/Recognition4 optical image analysis package successfully uses both approaches to provide the best possible particle separation during automated analysis. Its novel approach to watershed-based separation uses a modified procedure that, unlike many other methods, tests each attempted separation in order to verify that it satisfies validity criteria. The package also supports erosion-based separation, and the two methods can be used individually or in combination depending on the particular task. Mineral4/Recognition4 has a well-defined analysis workflow where particle separation takes an optimal position to ensure the quality of other operations. A number of analysis parameters can be adjusted by the operator to ensure good separation; once the parameters are defined and recorded the analysis can be performed without the operator’s intervention.
To determine the effect of the major separation parameters, such as the choice between watersheds and erosion, filling the pores in particle sections, removal of wrong separations, a hematitic/goethitic iron ore sample was analyzed using a number of different analysis profiles and the results were then compared to those from manual separation. The analysis cases covered a wide range, from no separation at all to significant over-separation. Apart from separation itself, an important factor affecting the measured results was the large section stereological correction that demanded objects smaller than a defined size limit to be excluded from consideration. The study attempted to evaluate the effects of major parameter changes, therefore no fine tuning of analysis parameters was performed.
It was determined that, using the proposed metrics, in the absence of separation the textural characterization error was at 3.75% using an area-based metric and at 6.65% using a particle-based metric (the latter was likely to include a significant over-estimate). Most automatically separated samples provided noticeably better textural characterization. The best analysis case showed an error of 1.41% by area and 1.99% by particle number, about three times better than characterization without particle separation, and a number of other cases also showed very significant improvement. Even the cases with important separation parameters significantly different from the optimal combinations typically resulted in better textural characterization compared to the non-separated case. Strong over-separation of the sample, however, resulted in a worse outcome.
Of the iron ore textural groups studied, it was found out that excessive separation (as well as the absence of separation) mostly affects the softer, more friable textures with relatively more abundant gangue minerals (e.g. kaolinite) and ochreous goethite. In particular, the effect of separation without fill on the characterization of those textural groups for the sample studied was negative. Based upon this finding, one of our suggestions is that in the presence of such porous textures, particles should be pre-filled, especially if watershed-based separation is used.
Liberation characterization for hematite was also performed for all the analysis cases selected for the study. This analysis confirmed that both the absence of separation and over-separation have a significant effect on the measured abundances of highly liberated classes. It was shown that moderate separation significantly improved measured abundance figures for all liberation classes, including fully liberated hematite.
It is therefore recommended that automated or manual particle separation is applied during textural and liberation studies of iron ore using optical image analysis, as it can significantly improve the quality of characterization. Particular attention should be paid to avoiding over-separation of the sample or preferential separation of soft textures though. Therefore stronger particle preservation mechanisms, including pre-fill during separation, and less destructive separation methods are suggested for the studies where textural information is of importance.
The authors wish to thank the staff of the CSIRO Carbon Steel Futures group for valuable discussions and support during this work, and the internal reviewers of the article Mark Pownceby and Michael Peterson for their very useful suggestions. The authors would also like to express their gratitude to Carlos Rodriguez-Navarro and the two anonymous reviewers for their thorough review of the manuscript and helpful editorial recommendations.