XXXI International Mineral Processing Congress 2024 Proceedings/Washington, DC/Sep 29–Oct 3 905
Modal Mineralogy Analysis
Two different methodologies were applied to analyse the fine
and coarse fractions. The fine fraction images were labelled
with the superpixel +classifier algorithm, while the coarse
fraction images were labelled with the instance segmenta-
tion algorithm. Figure 5 shows the software user interface
and the information extracted from the sample during the
optical microscopy scanning process. The MinDet set up
can detect particles in-pulp using optical reflective prop-
erties, identify the boundaries of particles (for sizing) and
classify the minerals (for surface grade calculation) within
pulp images, estimate the number of images required for a
representative sample to obtain an acceptable accuracy for
particle size distribution and mineralogy.
Deep Learning Instance Segmentation
Deep Learning (DL) is a subset of Machine Learning (ML)
that utilises deep neural networks to accomplish tasks like
machine vision, autonomous driving, and natural lan-
guage processing. DL can learn complex features to distin-
guish minerals and boundaries but typically requires large
and clean labelled dataset known as supervised learning.
However, it is possible overcome this requirement by utilis-
ing pre-trained models on other datasets and tweaking the
model parameters for a study with a smaller dataset (also
known as transfer learning). A feasibility study was done
previously to compare the transfer learning performances of
two DL instance segmentation algorithms on thin section
samples (Koh et al., 2021). The sample preparation tech-
niques to prepare samples and develop the ML algorithm
was demonstrated in detail by the authors elsewhere (Koh
et al., 2024).
Superpixel +Classifier Instance Segmentation
When collection of a labelled dataset of suitable size is
not feasible, it is still possible to segment images through
unsupervised learning. Such an example would be super-
pixel techniques which groups pixels according to common
characteristics. One of the most popular superpixel method
is the Simple Linear Iterative Clustering (SLIC) algorithm
(Achanta et al., 2012). The trade-off is that the superpixels
could only be subsequently classified through low-level fea-
tures like pixel colour and brightness.
RESULTS AND DISCUSSION
To reduce the effects of slurry sampling on the results in
this study, the validation of PSDs and modal mineralogy
was done with the MinDet batch mode as the sample qual-
ity can be controlled while being prepared. This focuses
the accuracy of the methodology instead of the variance
from slurry sampling which can be estimated based on
the sampled mass or area for a given slurry stream. As the
result, a more adequate comparison and error estimation is
obtained when MinDet reading is compared to the ground
Figure 4. Sampling and scanning unit, the compartments are
numbered as 1–Control Unit, 2–Sampling Unit, 3-Digital Microscope,
4–3D Traverse, 5–Automatic Valves, 6–Input/Output streams
Modal Mineralogy Analysis
Two different methodologies were applied to analyse the fine
and coarse fractions. The fine fraction images were labelled
with the superpixel +classifier algorithm, while the coarse
fraction images were labelled with the instance segmenta-
tion algorithm. Figure 5 shows the software user interface
and the information extracted from the sample during the
optical microscopy scanning process. The MinDet set up
can detect particles in-pulp using optical reflective prop-
erties, identify the boundaries of particles (for sizing) and
classify the minerals (for surface grade calculation) within
pulp images, estimate the number of images required for a
representative sample to obtain an acceptable accuracy for
particle size distribution and mineralogy.
Deep Learning Instance Segmentation
Deep Learning (DL) is a subset of Machine Learning (ML)
that utilises deep neural networks to accomplish tasks like
machine vision, autonomous driving, and natural lan-
guage processing. DL can learn complex features to distin-
guish minerals and boundaries but typically requires large
and clean labelled dataset known as supervised learning.
However, it is possible overcome this requirement by utilis-
ing pre-trained models on other datasets and tweaking the
model parameters for a study with a smaller dataset (also
known as transfer learning). A feasibility study was done
previously to compare the transfer learning performances of
two DL instance segmentation algorithms on thin section
samples (Koh et al., 2021). The sample preparation tech-
niques to prepare samples and develop the ML algorithm
was demonstrated in detail by the authors elsewhere (Koh
et al., 2024).
Superpixel +Classifier Instance Segmentation
When collection of a labelled dataset of suitable size is
not feasible, it is still possible to segment images through
unsupervised learning. Such an example would be super-
pixel techniques which groups pixels according to common
characteristics. One of the most popular superpixel method
is the Simple Linear Iterative Clustering (SLIC) algorithm
(Achanta et al., 2012). The trade-off is that the superpixels
could only be subsequently classified through low-level fea-
tures like pixel colour and brightness.
RESULTS AND DISCUSSION
To reduce the effects of slurry sampling on the results in
this study, the validation of PSDs and modal mineralogy
was done with the MinDet batch mode as the sample qual-
ity can be controlled while being prepared. This focuses
the accuracy of the methodology instead of the variance
from slurry sampling which can be estimated based on
the sampled mass or area for a given slurry stream. As the
result, a more adequate comparison and error estimation is
obtained when MinDet reading is compared to the ground
Figure 4. Sampling and scanning unit, the compartments are
numbered as 1–Control Unit, 2–Sampling Unit, 3-Digital Microscope,
4–3D Traverse, 5–Automatic Valves, 6–Input/Output streams