**1. Introduction**

More than 150 years ago Henry Clifton Sorby [1] used optical microscopy for the characterization of rocks and minerals. He developed the basic techniques of petrography, using the polarizing microscope to study the structure of rock thin sections. In the early twentieth century Murdoch [2] started to use measurements of ore mineral reflectance combined with microchemical techniques for mineral identification.

Presently, the mineralogy of commercial raw materials, such as iron ore, can be determined from X-ray diffraction (XRD) analysis, but for subsequent processing purposes it is also important to understand the actual abundance of each mineral, association/liberation characteristics, dimensional characteristics of particles and mineral grains, surface roughness, porosity and density, the presence of different textures, the reciprocal position of mineral grains, and other morphological and morphometric characteristics. For these purposes, imaging techniques such as scanning electron microscopy [3–6], Raman spectroscopy [7] and optical image analysis (OIA) [8–12] are used. Generally, these methods can segment different minerals, but identification of different morphologies of the same mineral requires the further application of textural/structural segmentation.

The standard method for mineral segmentation in OIA is thresholding [13], where minerals are segmented by their color and brightness. An example of hematite thresholding in crushed iron ore sinter is given in Figure 1. Segmentation of hematite, which is the brightest mineral in the digital image, is actually performed using three color channels, but for simplicity, only the red channel reflectivity histogram used to determine the selected phase is shown. The reflectivity range of hematite corresponds to the last peak in the histogram. Figure 1b shows a partially successful attempt to automatically identify hematite. In this example, the hematite areas with relatively lower reflectance are not identified, and from the reflectivity histogram it is evident that the hematite peak is only partially covered by the range between the lower and the upper limits, or thresholds. Only the image pixels with red channel reflectivity within those thresholds are identified as hematite in this example. However when the whole of the last peak in the reflectivity histogram is thresholded, the hematite becomes fully segmented (Figure 1c). The use of multispectral image acquisition systems based on narrow bandwidth (e.g., 10 nm) interference filters show more efficient segmentation of minerals compared to colour imaging using tristimulus (red, green, blue) filters [14–17].

For sinter characterization it is very important to segment the primary, or unreacted, hematite remaining after the sintering process, from the secondary hematite which precipitated from the sinter melt during cooling. The sinter particle at the left hand side of Figure 1a has only secondary hematite present, whereas the particle to the right contains both phases. The large hematite grain indicated by an arrow in the bottom-right corner of the image is a good example of primary hematite. Figure 1b clearly shows that, after partial thresholding, the amounts of both types of hematite were underestimated, which means that thresholding alone is unable to segment one type of hematite from another. The size of hematite grains also cannot be reliably used for segmentation. While primary hematite grains are generally large, it is clear that some of the secondary hematite grains in the particle to the left are larger than some of the primary hematite grains in the particle to the right.

For coke characterization it is important to segment Inert Maceral Derived Components (IMDC) and Reacted Maceral Derived components (RMDC) [18,19]. However, they also cannot be segmented by simple thresholding as discussed in the section on coke characterization.

In order to quantify coke phases that are difficult to segment automatically, as well as certain sinter phases such as primary and secondary hematite discussed above, the traditional approach employs manual point counting by a trained petrographer. The problem with this approach, apart from it being labor intensive and thus expensive, is that it can be very subjective. It is even possible for an individual petrographer to report different results for the same sample if re-analyzed after a significant time.

Automated optical image analysis reduces the subjectivity and makes the characterization more consistent. The approach adopted duringOIA to segment phaseswith similar reflectivitywould be analogous to what petrographers employ during manual point counting—i.e., segmentation by structure/texture.

CSIRO's Carbon Steel Futures group developed the optical image analysis software Mineral4/Recognition4 for optical image analysis of major ironmaking feedstock materials such as ores (iron ore in particular) including lump and fine ores, sinters, pellets and briquettes, coal, coke etc. [9,10]. It can comprehensively characterize phase abundances, porosity, liberation/association, texture, and other sample characteristics. The first and the most important step during characterization is the correct identification of phases (see [14]). Even if a multi-thresholding [20] approach is used, it will not necessarily allow for the acceptable segmentation of phases when their reflectivities overlap. To achieve this, a textural identification module was developed for the software, allowing the segmentation of phases which have the same or significantly overlapping reflectivity, but different morphology. To characterize a particular material, an "analysis profile" is developed, which records individual parameter settings and adjustments made during different stages of image analysis. The textural identification unit is a subset of the mineral/phase identification stage. It is engaged when necessary and can perform differently for different materials/phases according to the profile settings.

(**a**)

**Figure 1.** Thresholding of hematite in sinter during image analysis: (**a**) original image; (**b**) partial identification of hematite (segmentation of the brightest part); and (**c**) full identification. Reflectivity histograms show the different red channel low and high threshold values and represent screenshots from Zeiss AxioVision software.

This article demonstrates algorithms developed for textural segmentation of different ironmaking feedstock materials performed by the textural identification module in Mineral4/Recognition4. These algorithms are based on well-established image analysis procedures such as binary Erosion, Dilation, size-based noise reduction (Scrapping) etc. [21]. Similar approaches can also be used for image analysis of any other materials within a very wide range of possible OIA applications.
