Abstract
D.Ing. (Electrical Engineering)
Process Mineralogy is the application of mineralogical techniques to the exploration
of ore deposits and the design and optimisation of mineral processing flowsheets. Samples
can be drill cores, rocks and milled particles, to give a few examples. X-ray microtomography
has emerged as a complementary technique to the existing two-dimensional
imaging modalities and bulk mineralogical methods. The applications of analysing X-ray
micro-tomography scans include analysing packed particle beds to determine particlesize
distributions, mineral exposure and liberation, as well as analysing the pore network
within ores targeted by the oil and gas industry. X-ray micro-tomography suffers from
several artefacts, including beam hardening, blurring and streaks, of which beam hardening
and streaks are particularly problematic and common when scanning metal-bearing
ores.
A fundamental step in analysing a tomogram is to segment the different groups of
minerals that are present within the sample. This is necessary to measure mineral grain
properties and as a precursor to segmenting and analysing particles in a crushed or milled
sample. In order for X-ray micro-tomography to provide accurate measurements, this first
step of segmenting minerals must be performed accurately.
Machine learning has been used in image processing for a variety of applications,
including the analysis of optical microscopy images for medical purposes, and recently
the analysis of tomograms. The primary focus of this work is the application of machine
learning algorithms to the segmentation of minerals, as well as a means for measuring the
accuracy of those algorithms.
Four main problem areas were identified in this work. The first is the need for a
suitable algorithm for filtering tomograms to reduce the quantity of noise that is present
while minimising the additional blurring of the edges of mineral grains. The second problem
statement focuses specifically on machine learning, while the third problem statement
is directed at the description of voxels by means of several features. The fourth
problem area is measuring the accuracy of any measurements made on the segmented
tomograms. Without an analysis of the measurement accuracy, X-ray micro-tomography
will not be accepted by the industry at large. This work demonstrates a method by which
back-scattered electron images from a scanning electron microscope may be aligned to a
tomogram, and used to quantify the accuracy of mineral segmentation algorithms...